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.1003468 | Slowness and Sparseness Have Diverging Effects on Complex Cell Learning | Following earlier studies which showed that a sparse coding principle may explain the receptive field properties of complex cells in primary visual cortex, it has been concluded that the same properties may be equally derived from a slowness principle. In contrast to this claim, we here show that slowness and sparsity drive the representations towards substantially different receptive field properties. To do so, we present complete sets of basis functions learned with slow subspace analysis (SSA) in case of natural movies as well as translations, rotations, and scalings of natural images. SSA directly parallels independent subspace analysis (ISA) with the only difference that SSA maximizes slowness instead of sparsity. We find a large discrepancy between the filter shapes learned with SSA and ISA. We argue that SSA can be understood as a generalization of the Fourier transform where the power spectrum corresponds to the maximally slow subspace energies in SSA. Finally, we investigate the trade-off between slowness and sparseness when combined in one objective function.
| A key question in visual neuroscience is how neural representations achieve invariance against appearance changes of objects. In particular, the invariance of complex cell responses in primary visual cortex against small translations is commonly interpreted as a signature of an invariant coding strategy possibly originating from an unsupervised learning principle. Various models have been proposed to explain the response properties of complex cells using a sparsity or a slowness criterion and it has been concluded that physiologically plausible receptive field properties can be derived from either criterion. Here, we show that the effect of the two objectives on the resulting receptive field properties is in fact very different. We conclude that slowness alone cannot explain the filter shapes of complex cells and discuss what kind of experimental measurements could help us to better asses the role of slowness and sparsity for complex cell representations.
| The appearance of objects in an image can change dramatically depending on their pose, distance, and illumination. Learning representations that are invariant against such appearance changes can be viewed as an important preprocessing step which removes distracting variance from a data set in order to improve performance of downstream classifiers or regression estimators [1]. Clearly, it is an inherent part of training a classifier to make its response invariant against all within-class variations. Rather than learning these invariances for each object class individually, however, we observe that many transformations such as translation, rotation and scaling apply to any object independent of its specific shape. This suggests that signatures of such transformations exist in the spatio-temporal statistics of natural images which allow one to learn invariant representations in an unsupervised way.
Complex cells in primary visual cortex are commonly seen as building blocks for such invariant image representations (e.g. [2]). While complex cells, like simple cells, respond to edges of particular orientation they are less sensitive to the precise location of the edge [3]. A variety of neural algorithms have been proposed that aim at explaining the response properties of complex cells as components of an invariant representation that is optimized for the spatio-temporal statistics of the visual input [4]–[12].
The two main objectives used for the optimization of models of neural representations are sparseness and slowness. While in the context of unsupervised representation learning the two objectives have been proposed to similarly explain the receptive field properties of complex cells, there are important differences between them that may help to identify the algorithms used in biological vision. Intuitively, the slowness objective can be seen as a measure of approximate invariance or “tolerance”, whereas sparseness is better interpreted as a measure of selectivity. Tolerance and selectivity—or slowness and sparseness, respectively—can be understood as complementary goals which both play an important role for solving the task of object recognition [13]. A prominent view that goes back to Fukushima's proposal of the necognitron (1980) is that these goals are pursued in an alternating fashion by alternating layers of S and C cells where the S cells are optimized for selectivity and the C cells are optimized for tolerance. This idea has been inspired by the finding of simple and complex cells in primary visual cortex which also motivated the terminology of S and C cells.
Thus, based on the strong association between complex cells and invariance, one would expect that slowness rather than sparseness should play a critical role for complex cell representations. In this study, we investigate the differences between slowness and sparseness for shaping the receptive field properties of complex cells.
While for natural signals it may be impossible to find perfectly invariant representations, slowness seeks to find features that at least change as slowly as possible under the appearance transformations exhibited in the data [16], [9]–[12], [14]–[27]. In contrast to sparse representation learning which is tightly linked to generative modeling, many slow feature learning algorithms follow a discriminative or coarse-graining approach: they do not aim at modeling all variations in the sensory data but rather classify parts of it as noise (or some dimensions as being dominated by noise) and then discard this information. This is most obvious in the case of slow feature analysis (SFA) [21]. SFA can be seen as a special case of oriented principal component analysis which seeks to determine the most informative subspace under the assumption that fast changes are noise [28]. While it is very likely that some information is discarded along the visual pathway, throwing away information in modeling studies requires great caution. For example, if one discards all high spatial frequency information in natural images one would easily obtain a representation which changes more slowly in time. Yet, this improvement in slowness is not productive as high spatial frequency information in natural images cannot be equated with noise but often carries critical information. We therefore compare complete sets of filters learned with slow subspace analysis (SSA) [9] and independent subspace analysis (ISA) [4], respectively. The two algorithms are perfectly identical with the only difference that SSA maximizes slowness while ISA maximizes sparsity.
For sparseness it is common to show complete sets of filters, but this is not so in case of slowness. Based on the analysis of a small subset of filters, it has been argued that SSA may generally yield similar results to ISA [9]. In contrast, we here arrive at quite the opposite conclusion: by looking at the complete representation we find a large discrepancy between the filter shapes derived with SSA and those derived with ISA. Most notably, we find that SSA does not lead to localized receptive fields as has been claimed ([9], [29] —but see [28], [30]).
Complete representations optimizing slowness have previously been studied only for mixed objective functions that combined slowness with sparseness [8], [31]–[33] but never when optimizing exclusively for slowness alone. Here we systematically investigate how a complete set of filters changes when varying the objective function from a pure slowness objective to a pure sparsity objective by using a weighted mixture of the two and gradually increasing the ratio of their respective weights. From this analysis we will conclude that the receptive field shapes shown in [8], [31]–[33] are mostly determined by the sparsity objective rather than the slowness objective. That is the receptive fields would change relatively little if the slowness objective was dropped but it would change drastically if the sparsity objective was removed. These findings change our view of the effect of slowness and raise new questions that can guide us to a more profound understanding of unsupervised complex cell learning.
The central result of this paper is the observation that the effect of the slowness objective on complex cell learning is substantially different from that of sparseness. Most likely this has gone unnoticed to date because previous work either did not derive complete representations from slowness or combined the slowness objective with a sparsity constraint which masked the genuine effect of slowness. Therefore, we here put a large effort into characterizing the effect of slow subspace learning on the complete set of filter shapes under various conditions. We first study a number of analytically defined transformations such as translations, rotations, and scalings before we turn to natural movies and the comparison between slowness and sparseness.
The general design common to SSA and ISA is illustrated in Figure 1. We apply a set of filters to the input and square the filter responses. Two filters form a 2-dimensional subspace (gray box in Figure 1) and the sum of squared filter responses of these two filters yield the subspace energy response. This can be seen as the squared radial component of the projection of the signal into the 2D subspace formed by the two respective filters. For example, if the filters are taken from the Fourier basis and grouped such that the two filters within each subspace have the same spatial frequency and orientation and phase difference, the output at a fixed time instant is the power spectrum of the image . As input we used image patches sampled from the van Hateren image database [34] and from the video database [35], vectorized to -dimensions, and applied SSA to all remaining 120 AC components after projecting out the DC component.
In the first part of our study, the input sequence consisted of translations. As time-varying process for the translations, we implemented a two-dimensional random walk of an window over the full image. The shift amplitudes were drawn from a continuous uniform distribution between 0 and 2 pixels, allowing for subpixel shifts. The filters obtained from SSA are shown in Figure 2A. Each row contains the filter pairs of 6 subspaces, sorted by descending slowness from left to right and top to bottom. The filters clearly resemble global sine wave functions. The wave functions differ in spatial frequency and orientation between the different subspaces. Within each subspace, orientation and spatial frequency are almost identical, but phases differ significantly. In fact, the phase difference is close to (), resembling quadrature pairs of sine and cosine functions as it is the case for the two-dimensional Fourier basis. Accordingly, the subspace energy output of the resulting SSA representation is very similar to the power spectrum of the image .
In fact, one can think of SSA as learning a generalized power spectrum based on a slowness criterion. While the power spectrum is known to be invariant against translations with periodic boundary conditions, perfect invariance—or infinite slowness—is not achieved for the translations with open boundary conditions studied here (see Figure 2 B). The slowness criterion is best understood as a penalty of fast changes since it decomposes into an average over penalties of fast changes for each individual component (see methods). Therefore, we will always show the inverse slowness for each component such that the smaller the area under the curve the better the average slowness.
Compared to random subspaces, the decrease in , i.e. the increase in slowness, is substantial: the average inverse slowness decreases approximately by a factor of three. The low frequency subspaces are clearly the slowest subspaces, and slowness decreases with increasing spatial frequency. At the same time, however, the inverse slowness of all learned subspaces is still larger than 0, i.e. even for the slowest components, perfect invariance is not achieved. This is not surprising, as perfect invariance is impossible whenever unpredictable variations exist as it is the case for open boundary conditions.
In Figure 2 C, we show that SSA can indeed find perfectly invariant filters starting from a random initial filter set if one imposes periodic boundary conditions. To this end, we created pink noise patches with circulant covariance structure, i.e. the pixels on the left border of the image are correlated with pixels on the right border as if they were direct neighbors. As time-varying process, we implemented a random walk with cyclic shifts where the patches were translated randomly with periodic boundary conditions. As in the previous study, the shift amplitudes were drawn from a continuous uniform distribution between 0 and 2 pixels. Since the Fourier basis is the eigenbasis of the cyclic shift operator it should yield infinite slowness for the cyclic boundary conditions. Indeed, the filters learned from these data recover the Fourier basis with arbitrary precision. Perfect invariance is equivalent with the objective function converging to 0. This means that the response of each subspace is identical for all shifts. Figure 2D shows the inverse slowness of the individual components. For all filters, is very small (), close to perfect invariance and infinite slowness.
Given that the SSA representation learned for translations is very similar to the Fourier basis and since the Fourier basis achieves perfect invariance for cyclic shifts we proceeded to investigate whether the Fourier basis is optimal even for non-cyclic translations as well. We created three different data sets, with random translations as in the first study, but the maximal shift amplitude of the 2D random walk was 1, 2, and 3 pixels, respectively. As initial condition, we used the Fourier basis (Figure 3, ‘’) instead of a random matrix. The optimized bases are denoted as where indicates the maximal shift amplitude. We show the 2D-Fourier amplitude spectrum of the filters rather than the filters in pixel space because it is easier to assess the differences between the different bases. The DC component is located at the center of the spectrum.
During optimization, the basis slightly departs from the initial condition but remains very localized in the Fourier domain (Figure 3, ‘’). The low frequency filters become sensitive to higher frequencies while the high frequency filters become also sensitive to lower frequencies as the initial filters blur out towards the border or center, respectively. The objective function is improved for the optimized filters not only on the training but also on the test set (cf. Table 1). The slowness of the 60 individual components evaluated on identically created test sets (, , and , respectively) is shown in Figure 3. The Fourier filters are slower than the optimized filters for the first 20–30 components, then about equal for 10 components, and significantly faster for the remaining components. Apparently, the SSA objective sacrifices a little bit of the slowness of the low frequency components to get a comparatively larger gain in slowness from modifying the high frequency components. The optimization of average inverse slowness in contrast to searching for a single maximally slow component is a characteristic feature of SSA.
Even though we expect changes in natural movies to be dominated by local translations, it is instructive to study other global affine transforms as well. Therefore, we applied SSA to 3 additional data sets: The first data set contains patches from the van Hateren image set which were rotated around the center pixel. The second data set consists of patches from the van Hateren image set which were also rotated around the center pixel but where we kept only the pixels within a predefined circle. Specifically, we reduced the number of dimensions again to 121 pixels by cutting out the corners which left an circular image patch. The patches in the third data set were sampled with sizes ranging from to pixels and then rescaled to pixels, in order to obtain a patch-centered anisotropic scaling transformation. The preprocessing was identical to the previous studies and the initial filter matrix was a random orthonormal matrix. The filters and the objective of the individual subspaces of the rotation data are shown in Figure 4A. The filters resemble the rotation filters found with steerable filter theory [28]. The slowness of all components is significantly larger than for random filters, but with clearly decreasing slowness for the last subspaces. Notably, the last subspaces have no systematic structure. This can be explained by the fact that when rotating a square patch, the pixels in the 4 corners are not predictable unless for multiples of rotations. Therefore the algorithm cannot find meaningful subspaces that would preserve the energy for the pixels in the corners. The filters in Figure 4B from the disc shaped patches do not show these artifacts. Here, all filters nicely resemble angular wave functions as expected from steerable filter theory and also exhibit better slowness. Finally, the scaling filters are shown in Figure 4C. All filters resemble windowed wave functions that are localized towards the boundaries of the patch. This indicates that a scaling can be seen as a combination of local translations which go inward for downscaling and outward for upscaling. All subspaces defined by the learned filters are significantly slower than the random subspaces.
After characterizing the result of slow subspace learning for analytically defined transformations we now turn to natural movies and the comparison between slowness and sparseness. Specifically, we compare slow subspace analysis (SSA) to independent subspace analysis (ISA) in order to show how the slowness and the sparsity objective have different effects on the receptive field shapes learned. To this end, we combine the two objectives to obtain a weighted mixture of them for which we can gradually tune the trade-off between the slowness and the sparseness objective. In this way, we obtain a 1-parametric family of objective functions(1)for which the parameter determines the trade-off between slowness and sparseness. Specifically, we obtain SSA in case of and ISA for . As one can see in Figures 5 the filters learned with SSA () look very different from those learned with ISA (). This finding contradicts earlier claims that the filters learned with SSA are comparable to those learned with ISA. The most obvious difference is that the slowness objective works against the localization of filters that is brought forward by the sparsity objective.
For we will refer to the resulting algorithm as independent slow subspace analysis (ISSA). If a representation is optimized for its performance with respect to the slowness objective decreases monotonically with . At the same time, its performance with respect to increases with . The percentages shown indicate the increase in slowness and sparseness relative to the maximal gain that can be achieved if one optimizes solely for one of the two objectives. Note that the shapes of these curves depend on the objective functions used and are not invariant under pointwise nonlinear transformations. The values shown here are determined directly by the objective functions without any additional transformation (see Eqs. 3,11). Remarkably, it is possible to derive a representation which performs reasonably well with respect to both sparseness and slowness simultaneously. At an intermediate point where both objectives, and , are reduced by the same factor in our units, the performance is still larger than 80% for each. Interestingly, for this trade-off the receptive fields look quite similar to those obtained with ISA. This may explain why previous work on unsupervised learning with combinations of sparseness and slowness did not reveal that the two objectives drive the receptive fields towards very different shapes.
The trade-off in performance with respect to slowness and sparsity for natural movies, translation, rotation, and scaling is summarized in Figure 6. It shows the ISA filters (A), the ISSA filters at the intermediate point of slowness and sparsity for natural movies (B), translation (C), rotation (D), and scaling (E) and in the same order the SSA filters in (F,G,H,I). The concave shape of the curves (upper left) indicates that the trade-off between the two objectives is rather graceful such that it is possible to achieve a reasonably good performance for both objectives at the same time.
Unsupervised learning algorithms are a widespread approach to study candidate computational principles that may underly the formation of neural representations in sensory systems. Slowness and sparsity both have been suggested as objectives driving the formation of complex cell representations. More specifically, it has been claimed that the filter properties obtained from slow subspace analysis would resemble those obtained with independent subspace analysis [9] and that the optimal stimulus for SFA is localized [29]. Here, we showed that there is a striking difference between the sets of SSA and ISA filters: While the sparsity objective of ISA facilitates localized filter shapes, maximal slowness can be achieved only with global receptive fields as found by SSA.
The different implications of slowness and sparseness are most notable in filters containing high spatial frequencies. For low spatial frequency filters the number of cycles is small simply because it is constrained to be smaller than the product of spatial frequency and simulation window size. Since previous studies have inspected only low spatial frequency filters the different effect of sparseness and slowness has gone unnoticed or at least not been sufficiently appreciated [6], [9], [29]. A signature of the drive towards global filters generated by slowness can be found in the bandwidth statistics presented in [6]. Global filter shapes correspond to small bandwidth. While the authors mention that the fraction of small bandwidth filters exceeds that found for physiological receptive fields they rather suggested that this may be an artifact of their preprocessing, specifically referring to dimensionality reduction based on principal component analysis. However, the opposite is the case: the preprocessing rather leads to an underestimation of the fraction of small bandwidth filters. Principal component analysis will always select for low spatial frequency components and thus reduce the fraction of small bandwidth filters because it is the high spatial frequency components which have the smallest bandwidth.
While it is difficult to make rigorous statements that are model-independent, there are general arguments why the lack of localization is likely a generic consequence of slowness rather than a spurious property that was specific to SSA only: By definition a neuron cannot be driven by stimuli outside of its receptive field (RF). Therefore, whenever a stimulus is presented that drives the neuron inside its RF, the neuron must stop firing when the stimulus is shifted outside the RF. This suggests very generally, that in the presence of motion the objective of slowness or invariance necessarily requires large RFs. Sparsity, in contrast, encourages neurons to respond as selectively as possible. One obvious way to achieve this is to become selective for location which directly translates into small RF sizes.
In addition, analytical considerations suggest that slowness is likely to generate global filters with small bandwidth. For small image patches it is reasonable to assume that the spatio-temporal statistics are dominated by translational motion. Thus, it is not surprising that the filter properties of SSA found for natural movies resemble those for translations. In computer vision, there is a large number of studies which derive features that are invariant under specific types of transformations such as translations, scalings and rotations. An analytical approach to invariance is provided by steerable filter theory [36], [37] which allows one to design perfectly invariant filters for any compact Lie group transformation [38]. The best known example is the power spectrum which is perfectly invariant under translations with periodic boundary conditions [28]. For the other Lie group transformations studied in this paper, the symmetry was broken due to discretization and boundary effects. In these cases the representations found with SSA can be seen as a generalization of the Fourier transform whose subspace energies are not perfectly invariant anymore but at least maximally stable under the given spatio-temporal statistics. A very similar argument has also been made for SFA [30].
The receptive fields of complex cells determined from physiological experiments rarely exhibit multiple cycles as predicted by SSA. This indicates that complex cells in the brain are not fully optimized for slowness. It may still be possible though that slowness plays some role in the formation of complex cells. The trade-off analysis with the mixed objective has shown that giving up some sparsity allows one to achieve both relatively large sparsity and slowness at the same time with localized receptive fields.
Having established how exactly sparseness and slowness differ in their implied receptive fields also helps to address the roles of sparseness and slowness experimentally. Li & DiCarlo [39], [40] found neural correlates of the learning of invariances by manipulating the statistics of the presented stimuli. Since their recordings were from area IT where receptive fields are known to be very large, it would be very interesting to see the effect of similar experiments, made during the critical period, on complex cells in primary visual cortex. To distinguish between slowness and sparseness it might also be instructive to vary the temporal continuity of the training stimuli, e.g. by comparing the effect of smooth translations with discrete jumps on the learnt receptive fields. Another, possibly more direct approach to distinguish between sparseness and slowness might be to compute the respective objective functions directly on the sensory responses over development. While such an experiment has already been done for sparseness by [8] who interestingly found that sparseness decreases throughout development, we are not aware of the equivalent evaluation of any change in neuronal slowness.
Independent of what happens during development, the comparison of slowness and sparseness raises questions about how we should view the role of complex cells with respect to the tolerance-selectivity trade-off. Given that large receptive fields are advantageous for invariance or slowness, the small receptive field size of complex cells suggests that complex cells do not aim at achieving maximal tolerance but rather lean towards preserving a high degree of selectivity. For both ISA and SSA some degree of invariance is already built into the architecture which resembles the energy model of complex cells and will always find two-dimensional invariant subspaces. Instead of prescribing the invariant subspace dimensionality we wanted to know what happens if the subspace dimensionality is learned as well. This can be done by learning complex cells with SFA on the full quadratic feature space and then investigating the spectrum of the resulting quadratic forms. Comparing the number of subspaces employed by SFA to maximize slowness to empirical measurements in V1 [41], [42] it turns out that the number of subspaces employed by real neurons, and therefore the degree of invariance is smaller than predicted by slowness (see Figure S1).
The deeper principle underlying both sparsity and slowness is the idea of generative modeling [25]. From a generative modeling perspective, one is most concerned about modeling the precise shape of all variations in the data rather than just optimizing some fixed architecture or feature space to be as invariant or sparse as possible. More specifically, in a generative modeling framework all ingredients of the model are formalized by a density model and thus the likelihood becomes the natural objective function. This holds also true for the studies which combined the slowness objective with a sparsity objective in the past [8], [31]–[33]. The generative power of these models, however, still needs to be significantly improved in order to be able to explain object recognition performance of humans and animals. A better understanding of the partially opposing demands of slowness and sparseness on the response properties of visual neurons will help us understand the computational strategy employed by the visual system in reaching that performance.
The algorithm of slow subspace analysis (SSA) has previously been described by Kayser et al [9]. Just like in independent subspace analysis [4] also in SSA the -dimensional input space is separated into independent subspaces of dimensionality and the (squared) norm of each subspace should vary as slowly as possible. The output function of the -th subspace is then defined as(2)where K is the dimensionality of the subspace, the number of the subspace, and is the orthonormal filter matrix. It is important to notice that, for an input signal with zero mean and unit variance, has mean . For , the set of squared subspace norms corresponds to the power spectrum of the Fourier transform if the set of filters are the discrete Fourier transform.
The objective function of SSA has been called “temporal smoothness” objective by Kayser et al. [9] and is given by(3)Note, however, that increases with the amount of rapid changes and is minimized subject to . To find the optimal set of filters under the given constraints we use a variant of the gradient projection method of Rosen [43] which was successfully used for simple cell learning before [22].
In order to compute the gradient of the objective function we have to compute the temporal derivative of the output signal first, using the difference quotient as approximation:(4)As we use discrete time steps, we can set which leads to . This simplifies the objective function (3) as the temporal difference mean . The objective function can be further simplified by using the fact that for and having zero mean and unit variance, which leads to . The complete objective function is then(5)For every iteration, the gradient of the objective function is computed, scaled by the step length , and subtracted from the current filter set(6)The partial gradient with respect to is(7)with(8)The matrix containing the resulting filter set is then projected onto the orthogonal group using symmetric orthogonalization [44](9)yielding the closest orthonormal matrix with respect to the Frobenius norm [45]. Along this gradient a line search is performed where the initial step length is reduced until the objective function on is smaller than before the iteration proceeds.
The optimization is initialized with a random orthonormal matrix . As stopping criterion the optimization terminates when the change in the objective function is smaller than the threshold . In all our simulations we used a subspace dimension of . A python implementation of the algorithm can be found as part of the natter toolbox http://bethgelab.org/software/natter/.
Independent subspace analysis (ISA) has originally been proposed by Hyvärinen and Hoyer [4]. The only difference between SSA and ISA is the objective function. Generally speaking, ISA is characterized by a density model for which the density factorizes over a decomposition of linear subspaces. In most cases the subspaces all have the same dimension, and in case of natural images the marginal distributions over the individual subspaces are modeled as sparse spherically symmetric distributions. Like Hyvärinen and Hoyer [4] we chose the spherical exponential distribution(10)where is the subspace response as defined in Equation 2, is a scaling constant and the normalization constant. Correspondingly, the objective function reads(11)The scaling and normalization constants and can be omitted. This leads to the gradient(12)with as defined in Equation 8. The optimization is identical to SSA where only objective and gradient are replaced. For the numerical implementation of ISA we used a python translation of the code provided by the original authors at http://research.ics.aalto.fi/ica/imageica/.
The time-varying input signal was derived from the van Hateren image database [34] for translations, rotations and scalings and the van Hateren movie database [35] for movie sequences. The image database contains over 4000 calibrated monochrome images of pixels, where each pixel corresponds to of visual angle. We created a temporal sequence by sliding a window over the image. Step length and direction for translation, angle for rotation and anisotropic scaling factors were sampled from a uniform random process. If not stated otherwise, the translation was sampled independently for x- and y direction from a uniform distribution on , the rotation angle from a uniform distribution on and the scaling factors independently for x- and y-direction from a uniform distribution on . The movie database consists of 216 movies of pixels with a duration of 192 s and 25 frames per second. The images were taken in Holland and show the landscape consisting mostly of bushes, trees and lakes with the occasional streets and houses. The video clips were recorded from Dutch, German and British television with mostly wildlife scenes but also sports and movies. For each stimulus set we sampled patches.
The extracted image patches were treated as vectors by stacking up the columns of the image patches, resulting in a 121-dimensional input vector . We projected out the DC component, i.e. removed the mean from the patches, and applied symmetric whitening to the remaining 120 AC components. No low pass filtering or further dimensionality reduction was applied. All computations were done in the 120-dimensional whitened space and the optimized filters then projected back into the original pixel space.
|
10.1371/journal.pntd.0002682 | Inferring Plasmodium vivax Transmission Networks from Tempo-Spatial Surveillance Data | The transmission networks of Plasmodium vivax characterize how the parasite transmits from one location to another, which are informative and insightful for public health policy makers to accurately predict the patterns of its geographical spread. However, such networks are not apparent from surveillance data because P. vivax transmission can be affected by many factors, such as the biological characteristics of mosquitoes and the mobility of human beings. Here, we pay special attention to the problem of how to infer the underlying transmission networks of P. vivax based on available tempo-spatial patterns of reported cases.
We first define a spatial transmission model, which involves representing both the heterogeneous transmission potential of P. vivax at individual locations and the mobility of infected populations among different locations. Based on the proposed transmission model, we further introduce a recurrent neural network model to infer the transmission networks from surveillance data. Specifically, in this model, we take into account multiple real-world factors, including the length of P. vivax incubation period, the impact of malaria control at different locations, and the total number of imported cases.
We implement our proposed models by focusing on the P. vivax transmission among 62 towns in Yunnan province, People's Republic China, which have been experiencing high malaria transmission in the past years. By conducting scenario analysis with respect to different numbers of imported cases, we can (i) infer the underlying P. vivax transmission networks, (ii) estimate the number of imported cases for each individual town, and (iii) quantify the roles of individual towns in the geographical spread of P. vivax.
The demonstrated models have presented a general means for inferring the underlying transmission networks from surveillance data. The inferred networks will offer new insights into how to improve the predictability of P. vivax transmission.
| The transmission of Plasmodium vivax has induced enormous public health problems at the global level. Natural transmission of P. vivax depends on interactions between anopheles mosquitoes and human beings. There are two important factors that influence its geographical spread. First, different locations may have different risks of infection due to their heterogeneous environmental and demographical profiles. Second, human mobility may bring pathogens from high-transmission locations to low-transmission locations. In view of this, to effectively and efficiently control the geographical spread of P. vivax, it would be desirable for us to characterize how it transmits from one location to another. To achieve this, we first build a spatial transmission model to characterize both the heterogeneous infection risks at individual locations and the underlying mobility of infected populations. By doing so, we can further infer the underlying P. vivax transmission networks from tempo-spatial surveillance data by using a machine learning method (i.e., based on a recurrent neural network model). Our study offers new insights into malaria surveillance and control from the viewpoint of both system modeling and machine learning.
| As one of the malaria parasites that can infect and be transmitted by human beings, Plasmodium vivax has induced enormous challenges to the public health of human population. It has been estimated that 2.5 billion people all over the world are at risk of infection with this organism, among which China accounts for 19% of the global populations at risk [1]. To control, eliminate or even eradicate malaria, WHO has suggested that the most important measure is a timely response with the implementation of strategic intervention [2]. This requires the establishment of effective and efficient monitoring or surveillance systems [3]. Moreover, in practice, human mobility can introduce malaria into previously low-transmission or malaria-free areas, which has been cited amongst the significant causes of the failure of the Global Malaria Eradication Programme [4]. Therefore, it would be desirable to investigate the underlying geographical spread of malaria, which is not apparent from surveillance data. In this paper, the transmission networks of P. vivax characterize how the parasite transmits from one geographical location to another due to human mobility. By focusing on the malaria transmission in Yunnan province, People's Republic of China, we pay special attention to the problem of how to infer the underlying transmission networks of P. vivax based on tempo-spatial patterns of observed/reported cases.
Natural transmission of P. vivax depends on the interactions between female anopheles mosquitoes and human beings. On the one hand, the ability of mosquitoes to transmit P. vivax within a geographical location is dependent upon a series of biological factors, such as the daily survival rate of mosquitoes and the sporogonic cycle length of sporozoits in their bodies [5], [6]. On the other hand, human mobility between geographical locations in various temporal (e.g., daily or monthly) and spatial (e.g., intra-urban or inter-urban) scales may result in P. vivax transmission from high-transmission to low-transmission or malaria-free locations [7]–[9]. Generally speaking, the geographical spread of P. vivax has the following characteristics:
In view of this, to infer the underlying transmission networks of P. vivax, it would be desirable to address the following two computational issues:
Mathematically speaking, the problem can be defined as follows: Let be a directed network with self-links, where and represent the sets of nodes and links, respectively. Each node stands for a geographical location in a malaria transmission area, and each link stands for the possible P. vivax transmission from to . For each node , let be the set of nodes that have links from , i.e., , and be the set of nodes that have links to , i.e., . Note that does not belong to either or . Moreover, we denote the weight of link as to represent the proportion of infected populations transmitting from to . Specifically, refers to the proportion of infected populations in that do not transmit. In this case, the objective is to estimate the link weights of based on surveillance data, which are formulated as tempo-spatial series (corresponding to nodes or geographical locations, such as villages or towns). For each node , the tempo-spatial series take the form of 3-tuple , which indicates that cases are observed/reported at time step at with attribute set . In this paper, the attribute set consists of the dynamically-changing temperature and rainfall over time at node , which reflects the heterogeneity of the nodes concerning the transmission potential of P. vivax.
In this paper, we focus on the problem of how to infer the underlying transmission networks of P. vivax among 62 towns located in four adjacent counties (i.e., Teng Chong, Long Ling, Ying Jiang, and Long Chuan) in Yunnan, China (see Figure 1), where the IDs and names of these towns are listed in Table 1. All these towns have been experiencing high P. vivax transmission in the past three years, with at least one year having the annual incidence rate larger than 1/10,000. Figure 2 presents the reported P. vivax cases of the 62 towns in 2005 grouped by every two weeks. It can be observed that different towns has different patterns of infections. There are three major reasons: First, due to the environmental and demographical heterogeneity of these towns, the transmission potential of P. vivax at each individual town is different. Figure 3 shows the heterogeneous transmission potential (i.e., vectorial capacity) estimated by the average temperatures and accumulated rainfall at each town based on the method proposed by Ceccato et al. [6]. Second, human mobility from one location to another may result in geographical spread of P. vivax. Third, a large number of malaria cases in Yunnan are imported from Myanmar [10], which is a high-transmission country for malaria and contiguous with Yunnan.
Imported cases in this work are defined as malaria infections whose origin can be traced to an area outside the country. Based on the annual case reporting system in P.R. China, the fraction of imported cases of P. falciparum in Yunnan was about 69.0% in 2005 [11]. While in 2011, among totally 301 reported P. falciparum cases in Yunnan, 269 of them were imported cases (i.e., the fraction of imported cases was about 89.4%) [12]. It was also reported that the fraction of imported cases of P. vivax in China in 2011 is about 62.9% [12]. Along this line, in this paper, we study several transmission scenarios with respect to different percentages of imported cases (i.e., 60%, 70%, 80%, and 90%) among all the reported P. vivax cases in the 62 towns. Specifically, we present a spatial transmission model and a recurrent neural network model to (i) infer the transmission networks of P. vivax from tempo-spatial surveillance data, (ii) estimate the fraction of imported cases in all reported cases for each individual town, and (iii) examine the roles of individual towns on P. vivax transmission.
Due to the complex nature of P. vivax transmission, to infer the underlying transmission networks, appropriate spatial transmission model should first be formulated. In this paper, we aggregate the tempo-spatial series of surveillance data for each individual town based on a time step with duration . In reality, may be related to the incubation period of malaria (i.e., the period from the point of infection to the appearance of symptoms of the disease). In doing so, we assume that the observed/reported infections at time step are more likely to be infected at previous time step . Generally speaking, the causes of geographical spread of P. vivax are twofold. First, within a town/node , the number of malaria infections at a time step is determined by multiple factors, such as temperature, rainfall, population size, as well as the number of infections at previous time step . Second, human mobility may introduce P. vivax from one town to another. Specifically, we focus mainly on the mobility of infected populations among different towns because patients with typical malaria symptoms will be rapidly diagnosed and treated in Yunnan, P.R. China. It is seldom for a diagnosed patient to cause further malaria infection.
To model P. vivax transmission at a node, we use the notion of vectorial capacity (VCAP), which is defined as “the number of potentially infective contacts an individual person makes, through vector population, per unit time [13].” The VCAP is adapted from the basic reproductive number calculated based on the Macdonald model [14]. At each node , the value of VCAP is given by:(1)where represents the equilibrium mosquito density per person, is the expected number of bites on human beings per mosquito per day, is the probability of a mosquito surviving through one whole day, and is the entomological incubation period of malaria parasites. Based on the study of Ceccato et al. [6], all these parameters are dynamically dependent on temperature () and rainfall () at node . Table 2 shows the detailed parameter descriptions and settings in this work for calculating the vectorial capacity of each individual town in Yunnan. It should be noted that the values of relevant parameters are based on a certain degree of assumptions and estimates, and they could be adjusted once more accurate values are available in the future.
To further estimate the number of infections at a node , we introduce another notion of entomological incubation rate (EIR), which is defined as the number of infectious bites received per day by a human being [15]. Let denote the proportion of infected populations among all human populations at at time step , i.e., . Here, is the number of observed/reported infections at at time step , and is the population size of . Figure 4 shows a schematic diagram illustrating various data sources utilized (i.e., physiological, environmental, demographical, and surveillance data) for characterizing the infection risks of P. vivax at each individual town based on the notion of EIR. Mathematically, can be calculated through as follows:(2)where denotes the probability of the disease transmitting from an infectious person to an uninfected mosquito, represents the daily death rate of a mosquito [15].
Based on the definition of EIR, the estimated number of infections without considering human mobility at time step can be estimated as follows:(3)where represents the probability of the disease transmitting from an infectious mosquito to an uninfected person, and represents control impact of malaria transmission at node . Here, the control impact measures the efficiency of various intervention strategies implemented at node , such as insecticide treated nets, and long-lasting insecticide-treated nets. Although according to Equation 3, the estimated number of human infections at is a linear function of EIR at , the nonlinear interactions of infected mosquitoes and susceptible human beings and vice versa are taken into account in Equations 1 and 2 associated with VCAP and EIR, respectively. Specifically in this paper, since all of the 62 towns are within Yunnan, we assume the malaria control strategies over them have the same impact. Without loss of generality, we can set , which corresponds to perfect malaria transmission between human beings and mosquitoes. In reality, these parameters can be estimated by assessing biting habits of mosquitoes at different locations and conducting virological and serological analysis on infected individuals [16]–[18].
In the following, we introduce how to model the mobility of infected populations with respect to the geographical spread of P. vivax. Since human mobility among the 62 towns in Yunnan mainly relies on road transportation, in this paper, we assume that the transmission networks of P. vivax have the same topology (i.e., connectivity) with the transportation network. By doing so, we can quantify the transmission of P. vivax from one node to another by learning the link weight between them, which stands for the proportion of infected populations moving from to (Note that in this paper, the weight only characterizes the mobility of infected populations, where the population size of each node indirectly contributes to the weight via EIR). Accordingly, taking into consideration the mobility of infected populations, the number of increased infections at node can be calculated as follows:(4)which represents the difference between the number of cases transmitted from neighboring nodes and the number of cases transmitted to neighboring nodes. In summary, the estimated number of new infections of node at time step should be:(5)
After modeling the spatial transmission of P. vivax, we further introduce a recurrent neural network model, which allows for reflecting both structural (or spatial) and temporal dependencies of the nodes in the network by creating interdependent internal states in the model [19]. Specifically, we build the model by taking into consideration the control impact at individual nodes, the road transportation network, as well as the total number of imported cases to the towns from the outside. Figure 5 illustrates the internal states of the model within a time step. There are totally hidden layers in the network, and the links between two hidden layers are determined by the connectivity of the transportation network. Each hidden layer describes one stage of disease transmission between two neighboring towns. In doing so, to guarantee the possibility that one infected person may travel to any other towns at a time step, should be equal to the diameter of the road transportation network. The diameter of a network refers to the greatest distance between any pair of nodes in the network. To reflect the impact of P. vivax control at individual nodes, a vector is associated to the out-links of the nodes in the input layer. In addition, the total number of imported cases (i.e., ) of all the towns is linked to the nodes in the output layer of the neural network, where a vector () is associated with to represent the proportion of imported cases each town received in all the imported cases.
For each time step , we have a vector of reported infections , which represents the number of P. vivax infections at each individual town. Based on the proposed spatial transmission model, we can estimate the number of infections at time step by treating as an input. In other words, when an input pattern is presented to the network, it produces an output , which is usually different from the number of reported cases at time step . Suppose that we totally have a number of time steps, that is to say, we have a training set consisting of ordered pairs of dimensional vectors (i.e., input-output patterns). In this case, the problem of inferring underlying transmission networks of P. vivax is to learn the parameters , , and link weights (i.e., ) of by minimizing the sum of squares of error between the estimated numbers of infections (i.e., ) and the observed numbers of infections (i.e., ) for all towns and time steps, that is,(6)To solve the problem, we can use the backpropagation algorithm. The algorithm consists of three steps: (i) feed-forward computation, (ii) backpropagation computation, and (iii) weight updates.
Step 1: Feed-forward computation. Given an initial and the input vector , the estimated output at layer can be calculated as follows:(7)Accordingly, the final output at the output layer can be calculated by(8)where is a diagonal matrix with diagonal entries .
Step 2: Backpropagation computation. The vector of backpropagation error at the output layer is computed by . Then, the vector of backpropagation error at layer can be calculated as follows:(9)
Step 3: Weight updates. After the backpropagation error has been computed for all nodes in the network, we start to update the link weights. Based on the backpropagation algorithm, the update for any link weight between layer and is given by:(10)where is a learning constant defining the step length of the update. Since each link has the same weight at different layers, backpropagation is performed as usual for each link and the results are simply added, i.e., . For the situation that there are input-output patterns, the necessary update will be(11)The update of and can be done in a similar way, where and .
In summary, the objective of the backpropagation algorithm is to gradually adjust the link weights so as to minimize Equation 6 by treating each time step as an input-output pattern. Theoretically speaking, the global minimum cannot be guaranteed due to the nonlinearity of the optimization problem. In this case, the step length for weight updates is set to be a small value, i.e., . Moreover, the algorithm will be stopped when there are successive 10 times that the change of is less than 1.
The following data are involved in constructing our spatial transmission model and recurrent neural network model to infer the underlying transmission networks of P. vivax among 62 towns in Yunnan, P.R. China.
Since the available MODIS and TRMM data have different spatial resolutions, we first project the TRMM data into the same resolution with MODIS data (i.e., 1 km spatial resolution). In doing so, many spatial grids may have the same values of daily precipitation. Such a deficiency can be addressed if more accurate estimates are available in the future. Then, we aggregate the daily precipitations on an 8 day basis to match the temporal resolution of the MODIS data. Finally, by respectively averaging the aggregated MODIS and TRMM data in a time duration , we can calculate the value of VCAP for each individual town based on the model proposed by Ceccato et al. [6].
The proposed models have presented a general way to investigate the geographical spread of P. vivax based on surveillance data, which involve both the heterogeneous transmission potential of P. vivax and a machine learning algorithm. Based on the available one-year surveillance data, the demonstrated models are able to arrive at some informative results. Accordingly, if more malaria cases are collected from surveillance data across multiple years, the accuracy of our models will be further improved.
The number of reported P. vivax cases for each individual town shows a certain degree of spatial heterogeneity. Figure 7 demonstrates a smoothed surface map with respect to the number of reported cases in individual towns in Yunnan, P.R. China. The map is generated using ArcGIS version 10.0 (ESRI; Redlands, CA, USA), where the kernel density estimator with search radius 0.2 is employed. The size of a node in blue corresponds to the total number of reported cases in 2005, while the colored surface represents the hotspot density magnitude of the P. vivax cases after smoothing. Four obvious hotspots can be observed, that is, the areas in red around the towns of Wuhe, Gudong, Pingyuan, and Jinghan.
Based on the annual case reporting system in P.R. China over the last several years [11], [12], we assume that the fraction of imported cases among all the reported P. vivax cases in the 62 towns is at least 60%. Accordingly, we can estimate the proportion of imported cases for each individual town, that is, the vector for the 62 towns. Figure 8 shows the estimated proportion of imported cases for each individual town under four scenarios with different percentage of imported cases in the total number of reported cases (i.e., 60%, 70%, 80%, and 90%). The error bars demonstrate the standard deviations, which refer to the variation of the estimated results for the four scenarios. It can be observed that for most towns, the proportion of imported cases does not vary too much. This is reasonable because international labor/tour mobility may have certain regular temporal or spatial patterns [26]. Specifically, it can also be observed that the town Wuhe has the largest proportion of imported cases among the 62 towns. This is consistent with the situation that Wuhe is the hotspot of malaria transmission (see Figure 7). From the viewpoint of active surveillance and intervention, we can pay special attention to those towns with a larger proportion of imported cases, namely, Wuhe, Tuantian, Mingguang, Tengyue, and Longjiang.
Figure 9 illustrates the values of weight matrices for the four scenarios with different percentages of imported cases. It seems that the inferred transmission networks of P. vivax (i.e., the weight matrices) show different patterns when the total percentage of imported cases changes. Particularly, it can be observed that as the total percentage of imported cases increases, the values of the diagonal entries vary dramatically. Note that the diagonal entries in a weighted matrix represent the severity of P. vivax transmission within individual towns (i.e., self-propagation of malaria) associated with their local transmission potential. This is because there is only little change about the proportion of imported cases for each individual town as shown in Figure 8. In this case, as the total percentage of imported cases increases, the total number of P. vivax cases caused by local infections will decrease. In other words, the P. vivax cases of individual towns will become geographically sparse. In this case, some towns with high malaria transmission risks may need to contribute more to the number of reported P. vivax cases in other towns to minimize the sum of squares for error, which makes them much easier to be identified.
Give the total percentage of imported cases in the 62 towns in Yunnan, we can further assess the roles of individual towns during the P. vivax transmission. Based on the estimated weight matrix for the scenario with 80% imported cases, the towns can be classified into two typical categories: the self-propagating towns and the diffusive towns (see Figure 10). A self-propagating town has a relatively larger , which means that fewer new infections in this town will transmit to other towns. While a diffusive town has a relatively smaller , which means that new infections in this town will be more likely to transmit to other towns. Figure 10 shows an example of classification with two specific thresholds, i.e., 0.5 and 0.8. The towns with the proportion of self-propagation larger than 0.8 (respectively, less than 0.5) are classified into the category of self-propagating towns (respectively, diffusive towns). The names of the corresponding towns can be found in Table 1. In reality, the thresholds can be defined by domain experts based on their work experiences.
With respect to the vector-borne pathogen (i.e., malaria), existing studies have shown that human mobility from one location to another, which exhibits various spatial and temporal scales, is a key behavioral factor for its geographical spread. This is because human mobility influences their exposure to infectious vectors (i.e., mosquitoes), and further the malaria transmission [8], [27], [28]. Extensive studies have been conducted attempting to quantify human mobility patterns so as to indirectly predict the underlying malaria transmission networks. Such human mobility patterns can be constructed from various available data, such as survey [29], census data [30], airline transportation [31], mobile phone [9], [32], [33], or even by certain computational methods, such as the gravity model or its extension [34]. However, most of them emphasize only the impacts of human mobility, which cannot reflect the complex properties of malaria transmission. To step forward to understand the underlying transmission networks of P. vivax, in this paper, we have considered both the dynamics of P. vivax transmission and the impact of human mobility.
Another research direction focuses on understanding the critical features of host-vector-parasite interactions by building explicit mathematical models, which assume homogeneous mixing of the population [13]. Starting from the Ross model [35], a variety of differential equation models with different levels of complexity have been proposed to investigate the roles of demographic, socio-economic, and environmental factors (e.g., age, immunization, and migration), which are helpful to predict the effects of interventions on the model parameters. Along this line, to assess the effects of human mobility on the persistence of malaria, many spatial transmission models have been proposed [28], [36], [37]. One common limitation of these conceptual models is that the population of both human beings and mosquitoes are assumed to be fixed. However, researchers have shown that environmental factors (e.g., temperature and rainfall) have a significant impact on mosquito population as well as their biological cycles [38], [39]. In this paper, we have adopted the notion of vectorial capacity (VCAP) to characterize the heterogeneous transmission potential of P. vivax at different locations. Specifically, a vectorial capacity model proposed by Ceccato et al. [6] is used to monitor changing malaria transmission potential within a town by taking into consideration the impact of temperature and rainfall on the bionomics of mosquitoes and the parasite extrinsic incubation period in mosquitoes.
The last decade has witnessed a great upsurge in studying and revealing the unifying principles of real-world systems by modeling them as complex networks [40]–[42]. Since then, lots of efforts have been made to investigate disease transmission in populations by integrating epidemic modeling with complex networks analysis (e.g., human contact heterogeneity [43]). Each node in a network can represent either an individual or a group of individuals to model disease transmission at the individual/metapopulation level [44]. Accordingly, the transmission dynamics on the network can be formulated by stochastic models on regular networks [45] or irregular networks [46]. The mean-field versions of stochastic models on regular networks correspond to the deterministic models for which the homogeneous mixing of the population is a good approximation. One major concern of these studies is to investigate the impacts of realistic network topologies (e.g., random networks [47], small-world networks [47], [48], and scale-free networks [49]) on the process and results of disease transmission. Different from these studies, in this paper, we have focused on inferring the underlying P. vivax transmission based on a small-scale actual network (i.e., the road transportation network among the 62 towns in Yunnan). In the future, the proposed model may be considered for larger networks, in which a complex networks approach will be suitable.
Regarding the machine learning procedure, Liu et al. [50] have stated that the methods to infer underlying networks of disease transmission from observed incidences could be significantly different from those to infer the structures of diffusion networks from information flows due to the unique nature of disease transmission dynamics [51], [52]. Existing methods consider merely temporal information to infer diffusion networks, and most of them are based on the assumption of independent cascading of information. On the contrary, malaria may spatially propagate due to human mobility in two ways: (i) infected persons may bring the pathogen from one location to another, and (ii) susceptible persons can become infected while traveling to high-transmission locations. Therefore, geographical malaria transmission is not independent cascading. Reasonable transmission networks can be discovered only when appropriate transmission models are formulated.
As for the predictability, it is always expected that there is a powerful model that can provide accurate predictions on the malaria transmission patterns. However, it is extremely challenging due to the complicated dynamics of malaria transmission. Based on surveillance data for scenarios with various percentages of imported cases among all reported P. vivax cases, the hybrid model (i.e., the spatial transmission model and the recurrent neural network model) presented in this paper can help infer (i) the the proportion of imported cases for individual towns, and (ii) the transmission networks of P. vivax among the 62 towns. The results have shown that the proportion of imported cases for individual nodes (i.e., the value of vector ) is relatively stable for different percentages of imported cases (Figure 8), while the underlying transmission networks depend heavily on the total number of imported cases (Figure 9). In P.R. China, the number of imported P. falciparum cases at the county level is released every year through an annual case reporting system. To further implement our models, it would be necessary to continuously monitor the imported P. vivax cases. By doing so, our models may provide public authorities with new insights into active surveillance and control of P. vivax transmission. Specifically, this can be achieved by (i) identifying whether or not a particular P. vivax case is imported during data collection in the front line, and (ii) analyzing the tempo-spatial patterns of imported P. vivax cases across multiple years.
Last but not the least, this work is novel in that it provides a way to investigate the underlying malaria transmission patterns from the real-world malaria surveillance data [53], [54]. Figure 11 illustrates a machine learning framework, which consists of the interactions between malaria transmission models and machine learning models. The framework consists of three interactive components:
The integration of the spatial transmission model and the recurrent neural network model in this paper provides a typical implementation of this framework.
Finally, due to the data availability at the moment, the proposed models still have several limitations that are worthy of being improved and investigated in the future:
|
10.1371/journal.ppat.1005882 | IL-1 Coordinates the Neutrophil Response to C. albicans in the Oral Mucosa | Mucosal infections with Candida albicans belong to the most frequent forms of fungal diseases. Host protection is conferred by cellular immunity; however, the induction of antifungal immunity is not well understood. Using a mouse model of oropharyngeal candidiasis (OPC) we show that interleukin-1 receptor (IL-1R) signaling is critical for fungal control at the onset of infection through its impact on neutrophils at two levels. We demonstrate that both the recruitment of circulating neutrophils to the site of infection and the mobilization of newly generated neutrophils from the bone marrow depended on IL-1R. Consistently, IL-1R-deficient mice displayed impaired chemokine production at the site of infection and defective secretion of granulocyte colony-stimulating factor (G-CSF) in the circulation in response to C. albicans. Strikingly, endothelial cells were identified as the primary cellular source of G-CSF during OPC, which responded to IL-1α that was released from keratinocytes in the infected tissue. The IL-1-dependent crosstalk between two different cellular subsets of the nonhematopoietic compartment was confirmed in vitro using a novel murine tongue-derived keratinocyte cell line and an established endothelial cell line. These data establish a new link between IL-1 and granulopoiesis in the context of fungal infection. Together, we identified two complementary mechanisms coordinating the neutrophil response in the oral mucosa, which is critical for preventing fungal growth and dissemination, and thus protects the host from disease.
| The opportunistic pathogen Candida albicans is a major risk factor for immunosuppressed individuals, and oropharyngeal candidiasis (OPC) is a frequent complication in patients with weakened cellular immunity. The cytokine interleukin-17 (IL-17) plays a critical role for antifungal host defense and was proposed to act by regulating neutrophil recruitment to the oral mucosa. However, although IL-17 can promote neutrophil trafficking in some situations, we recently showed in a mouse model that this is not the case during OPC. Thus, the mechanism governing the neutrophil response to C. albicans remained to be determined. Here, we demonstrate an essential role of IL-1 receptor (IL-1R) signaling in the recruitment of neutrophils from the circulation to the infected tissue via enhanced secretion of chemokines and increased output of neutrophils from the bone marrow. We found that IL-1α is released from keratinocytes upon invasion of C. albicans and acts on endothelial cells to induce the production of granulocyte colony-stimulating factor (G-CSF), a key trigger of emergency granulopoiesis. Thereby, IL-1R signaling translates the local response to the fungus in the oral mucosa into a systemic response that critically contributes to protection from infection.
| The opportunistic fungal pathogen Candida albicans has emerged as a significant cause of morbidity and mortality worldwide, particularly in immunocompromised individuals [1]. Of the diverse forms of disease manifestations, mucosal infections with C. albicans are by far most abundant [2]. The symptoms reach from mild forms of infection to chronic or recurrent diseases. No licensed fungal vaccines are currently available to prevent disease, and toxicity and resistance to available drugs compromise the effective management of patients. With the ever-increasing population of immunocompromised patients, C. albicans infections thus represent an important socio-economic challenge worldwide.
The epithelium constitutes the first point of contact between the fungus and the host [3]. It provides an important physical barrier to prevent fungal invasion. Moreover, it has the capacity to sense and respond to the fungus. By producing inflammatory mediators and antifungal defense molecules the epithelium actively participates in the host response and together with leukocytes, including neutrophils and IL-17-producing lymphocytes, contributes to limiting fungal (over)growth. Diverse mutual interactions between leukocytes and the epithelium are critical for mounting a broadly protective response against C. albicans. The epithelium elicits signals in response to the fungus that can promote the inflammatory response [3], while cytokines such as IL-17, produced by leukocytes, act on the epithelium to enhance its barrier function and antimicrobial activity [4,5]. Neutrophils have been shown to rapidly accumulate in the oral mucosa in response to C. albicans infection, and they critically contribute to prevent invasion of the fungus in underlying tissues and dissemination to the circulation and visceral organs as was shown in a model of acute oropharyngeal candidiasis (OPC) [6,7]. The relevance of neutrophils in protection from oropharyngeal candidiasis is also evidenced by the high incidence of the disease in hemato-oncological patients with bone marrow aplasia [8,9].
Neutrophils comprise a major proportion of circulating peripheral blood leukocytes. They are generated from granulocyte-macrophage progenitors in the bone marrow under the control of granulopoietic growth factors, primarily granulocyte colony-stimulating factor (G-CSF) [10]. During acute infection, granulopoiesis is massively enhanced to comply with the increased demand for neutrophils in host defense [11]. Control mechanisms of this demand-adapted hematopoiesis involve long-distance regulatory feedback loops induced at the site of infection where neutrophils act, which is usually distant from the production site of neutrophils in the bone marrow. Increased release of G-CSF in response to infectious and/or inflammatory insult plays a key role in this process [11]. Given the potentially harmful effects of dysregulated neutrophils, granulopoiesis and neutrophil trafficking is under tight control and regulated in a tissue-specific manner [12].
With the discovery of interleukin-17 (IL-17) and the realization of its critical role in defense against mucocutaneous candidiasis [5,13], it was postulated that IL-17 mediates protection by promoting the neutrophil response. Indeed, IL-17 signaling can enhance expression of neutrophil cytopoietic and chemotactic factors in response to C. albicans [14]. However, we recently demonstrated that neutrophils are recruited normally to the site of infection in IL-17 receptor-deficient mice, thus that the IL-17 pathway is not required for the neutrophil response during OPC [6]. Therefore, although neutrophil trafficking may be regulated by IL-17 in some tissues and in response to certain pathogens [15–19]—this is not the case during C. albicans infection in the oral mucosa.
Alternative candidate factors regulating the neutrophil response include IL-1. In fact, secretion of both IL-1α and IL-1β are efficiently induced in dendritic cells and macrophages when stimulated with C. albicans [20–23] and both IL-1 family members were shown to contribute to protection from systemic infection [24]. Epithelial cells also secrete IL-1 cytokines when triggered with C. albicans [25–27], although in mice, unlike in humans, keratinocytes produce only IL-1α but no IL-1β [28]. Whether and how IL-1 cytokines contribute to antifungal defense in barrier tissues remains poorly defined.
Using a mouse model of oropharyngeal candidiasis, we found that IL-1R signaling is critical for host defense by regulating the neutrophil response. We show here that IL-1 acts by two complementary mechanisms. First, it regulates the production of neutrophil-chemotactic factors by epithelial cells for the recruitment of neutrophils from the circulating pool to the site of infection. Second, it induces G-CSF production by the endothelium for enhanced neutrophil output from the bone marrow to meet the increased demand in response to infection. Release of IL-1α from keratinocytes upon contact with C. albicans is critical for mediating this crosstalk between the epithelium and the endothelium in the oral mucosa.
Infection of mice with C. albicans via the oropharyngeal route induces a rapid inflammatory response, characterized by a massive recruitment of neutrophils to the oral mucosa (Fig 1A, S1 Fig) [6,7,29], which accumulate in proximity to where C. albicans hyphae invade the keratinocyte barrier on the dorsal and ventral side of the tongue [6]. Following from our previous findings that neutrophil trafficking during OPC was independent of the IL-17 pathway [6], we sought after factors responsible for controlling neutrophil recruitment in response to C. albicans. We found that mice lacking the IL-1 receptor (IL-1R) recruited significantly less neutrophils to the site of infection than their wild type (WT) counterparts (Fig 1B). In consequence Il1r1-/- mice were unable to control the fungus and displayed an increased fungal load in the tongue on day 3 post-infection (Fig 1C). These data indicated clearly that IL-1R signaling was critical for the neutrophil response during OPC.
The rapid accumulation of neutrophils in the oral mucosa of infected mice was paralleled by a strong induction of the neutrophil-recruiting chemokines CXCL1, CXCL2 and CXCL5 (Fig 2A). Consistent with the role of IL-1 in neutrophil recruitment, chemokine expression in the oral epithelium was impaired in absence of IL-1R signaling (Fig 2B), while basal levels were comparable (S2 Fig). To determine the cellular compartment responsible for chemokine production, we sorted cellular subsets from the tongue of naïve and C. albicans-infected mice on day 1 post-infection, the peak of the neutrophil response, including CD45+ leukocytes, CD45- EpCAM+ CD31- keratinocytes and CD45- EpCAM- CD31+ endothelial cells (S3 Fig) and analyzed chemokine expression at the transcriptional level. The keratinocyte fraction displayed the highest RNA levels, and the induction in response to infection was most prominent in this population (Fig 2C). Consistent with this result, a cell line of tongue-derived keratinocytes (TDKs) also secreted neutrophil-attracting chemokines in an IL-1-dependent manner in vitro (Fig 2D). These data suggested that the neutrophil response to OPC is initiated locally in the infected mucosa. While C. albicans may directly induce chemokine expression in keratinocytes, their production is strongly enhanced by IL-1R signaling.
In addition to the strong induction of the neutrophil-recruiting chemokines during OPC, the expression of Csf3, the gene coding for G-CSF, was also markedly induced in the infected oral mucosa (Fig 3A), suggesting that it may contribute to the overall neutrophil response during OPC by boosting granulopoiesis and neutrophil egress from the bone marrow. We therefore analyzed typical surrogate hallmarks of emergency granulopoiesis in the bone marrow. While total numbers of CD45+ cells in the bone marrow were unchanged in infected mice compared to naïve controls (Fig 3B), we observed a reduction in mature Ly6Ghi CD11b+ Ly6Cint neutrophils (Fig 3C and 3D). This response was paralleled by an increase in Ly6Glo CD11b+ Ly6C- immature neutrophils (Fig 3C and 3E). These reciprocal changes in mature and immature neutrophils in the bone marrow indicated that G-CSF, which was induced by the local infection with C. albicans, acted at a distance and thereby elicited a systemic response. Indeed, neutralization of G-CSF impaired the induction of emergency granulopoiesis during infection (Fig 3F and 3G). The increased demand for neutrophils in the infected tissue was thus compensated by an increased rate of granulopoiesis and mobilization of neutrophils from the bone marrow.
The increased neutrophil output from the bone marrow was relevant for the local antifungal response as indicated by the fact that depletion of the circulating neutrophil pool with an anti-Ly6G-specific antibody (1A8) was not sufficient to blunt the neutrophil response to C. albicans [6]. Instead, the combination of anti-Ly6G and anti-G-CSF was required for efficient depletion of neutrophils in the blood and in the infected tissue (Fig 3H and 3I) resulting in a total loss of fungal control (Fig 3J). Note that the reduction of neutrophils by ~1 log (Fig 3I) resulted in a ~4-log increase in fungal load over a course of 3 days (Fig 3J), while the ~0.5 log reduction in neutrophils in absence of IL-1 signaling (Fig 1B) lead to a ~2-log increase in fungal load over the same period of time (Fig 1C).
Given the prominent role of G-CSF for the overall neutrophil response during OPC we wanted to understand the regulation of this growth factor in more detail. For this, we isolated again CD45+ leukocytes, CD45- EpCAM+ CD31- keratinocytes, and CD45- EpCAM- CD31+ endothelial cells from the tongue of C. albicans-infected mice and uninfected controls as above (S3 Fig) and analyzed Csf3 transcript levels in each population. Surprisingly, the most prominent expression was observed in the endothelial cell population (Fig 4A). Although keratinocytes were previously shown to secrete G-CSF in response to C. albicans in vitro [26], their contribution in vivo was minor (Fig 4A). Similarly, leukocytes expressed only very low levels of G-CSF during OPC (Fig 4A). To corroborate this unexpected finding, we examined G-CSF protein production by the different cell subsets. Because we were unable to visualize intracellular G-CSF by flow cytometry, we prepared cell lysates from sort-purified tongue cell populations and quantified their G-CSF content by ELISA. Again, by far the highest production of G-CSF was detected in the endothelial cell fraction isolated from infected mice (Fig 4B). Further separation of the CD45- EpCAM- CD31+ population in podoplanin-positive and podoplanin-negative subsets confirmed that blood endothelial cells rather than lymph endothelial cells were responsible for G-CSF production during OPC (S4 Fig). Production of G-CSF by endothelial cells was supported by the observation that high levels of G-CSF could be detected in the serum of infected mice (Fig 4C). Together, these data suggested that G-CSF, which is induced in response to local infection with C. albicans, acts at a distance to promote granulopoiesis and neutrophil mobilization in the bone marrow and thereby sustains a systemic neutrophil response that meets the increased demand of these cells during OPC.
Next, we investigated how G-CSF secretion by endothelial cells is regulated during OPC. Direct stimulation of endothelial cells by C. albicans appeared unlikely given their spatial distribution in the oral mucosa. We made use of VE-cadherin-cre x ROSA26-RFP reporter mice [30] to visualize the blood and lymphatic vessels in situ and infected them with GFP-expressing C. albicans (Fig 5A). In immunocompetent animals, fungal hyphae were restricted to the avascular tongue epithelium without penetrating the basal epithelial layers. The absence of direct contacts of C. albicans with endothelial cells suggested that G-CSF production was regulated indirectly.
Given the important role of IL-1 for the overall neutrophil response to C. albicans, we assessed the IL-1 dependence of G-CSF during OPC. The induction of Csf3 transcripts was less pronounced in the oral mucosa of Il1r1-/- mice (Fig 5B), while basal levels were unchanged (S2 Fig). Likewise, G-CSF protein expression by endothelial cells that were sorted from infected tongues was drastically diminished in absence of IL-1R signaling (Fig 5C). This translated in strongly diminished G-CSF levels in the serum of C. albicans-infected Il1r1-/- mice compared to WT controls (Fig 5D). In contrast, TNF, which had been proposed to regulate G-CSF expression in vitro [31], was not involved in G-CSF production during OPC (S5 Fig). As a consequence of the G-CSF defect in Il1r1-/- mice, emergency granulopoiesis was strongly impaired during OPC in these mice, as indicated by the higher ratio of mature to immature neutrophils in the bone marrow of infected Il1r1-/- mice compared to WT controls (Fig 5E and 5F). Importantly, Il1r1-/- mice responded to G-CSF treatment, and the administration of recombinant G-CSF was sufficient to fully overcome the defect in emergency granulopoiesis in these mice (Fig 5G and 5H).
Together, these data demonstrated, that G-CSF secretion during OPC was controlled by the IL-1 pathway. G-CSF production appeared to underlie a different regulatory mechanism compared to neutrophil chemokine production during OPC given their distinct cellular sources, despite the fact that both, chemokines and G-CSF, depended on IL-1R signaling.
Next, we examined the expression of IL-1α and IL-1β, the two activating ligands of the IL-1R. Both were found strongly induced during OPC (Fig 6A). IL-1α and IL-1β contribute to G-CSF production during OPC because the induction of G-CSF expression in the oral mucosa and its release into the serum were strongly impaired in absence of IL-1α or IL-1β (Fig 6B and 6C).
IL-1β was expressed predominantly by the hematopoietic compartment as assessed by flow cytometric analysis of intracellular pro-IL-1β (S6 Fig). To determine the cellular source of IL-1α in the murine tongue, we applied an immunofluorescence approach, which allowed us to detect IL-1α with high specificity on tissue sections. Basal expression levels observed in the keratinized epithelium of the tongue in naïve mice were strongly enhanced upon infection (Fig 6D). Co-staining with antibodies specific for keratinocytes of the tongue (keratin-6) or those of stratum basale (keratin-14) [32] revealed that IL-1α was predominantly produced by differentiated keratinocytes (Fig 6E). Note that the IL-1α signal was absent in neutrophil-rich areas, which were identified by DAPI staining.
The availability of preformed IL-1α (but not IL-1β) protein in steady-state (Fig 6D, S6 Fig), which can be rapidly released in response to stimulation, together with the strategic position of keratinocytes as the first contact point between the host and the infecting fungus, suggested that keratinocyte-derived IL-1α likely functions as an ‘alarmin’ right at the onset of infection to alert the host about fungal invasion and to initiate a protective antifungal response. As such, keratinocyte-derived IL-1α may also act on endothelial cells and thereby contribute to the neutrophil response during OPC.
To delineate the putative crosstalk between keratinocytes and endothelial cells in more detail, we made use of a newly generated cell line of mouse tongue-derived keratinocytes (TDKs) (S7 Fig). TDKs expressed high levels of EpCAM and keratin-6 indicating that they represented differentiated oral keratinocytes [32]. Their concurrent expression of keratin-14 was consistent with their origin of basal keratinocytes with stem cell properties. Consistent with published data with keratinocytes from other sources [26], TDKs released IL-1α when stimulated with C. albicans (Fig 7A). Notably, and in contrast to human keratinocytes [28], mouse keratinocytes released no IL-1β (Fig 7B). This response was dependent on live and hyphenating fungus because heat-killed or a yeast-locked C. albicans did not induce IL-1α (Fig 7C). Likewise, IL-1α was not induced with zymosan, a yeast cell wall extract, or curdlan, a pure β-glucan preparation (Fig 7A). Consistent with our in vivo data, the overall amount of IL-1α detected from TDKs in response to live C. albicans resulted from the release of preformed cytokine and de novo biosynthesis of IL-1α (Fig 7D).
Next, we tested the effect of TDKs and TDK-derived IL-1α on endothelial cells for G-CSF induction (Fig 8A). For this, we employed an established endothelial cell line, MS1 [33]. TDKs and MS1 cells both did not produce G-CSF when directly stimulated with C. albicans nor with curdlan or zymosan, although they responded strongly to LPS, which was included as a positive control (Fig 8B and 8C). However, MS1 cells secreted high amounts of G-CSF when stimulated with the sterile-filtered supernatant of C. albicans-stimulated TDKs (Fig 8D). This response was dose-dependent (Fig 8E) and only observed when TDKs were stimulated with life and hyphenating C. albicans, but not with heat-killed C. albicans, a yeast-locked strain of C. albicans or inert fungal cell wall components such as zymosan and β-glucan (curdlan) (Fig 8D–8F). Similar results were obtained when supernatant of freshly isolated mouse oral keratinocytes stimulated with C. albicans was added to MS1 cells (S8 Fig). This indicated that a C. albicans-induced TDK-derived soluble factor was responsible for G-CSF production by endothelial cells. To test whether this factor was IL-1α, we added anakinra (IL-1R antagonist) or a neutralizing anti-IL-1α antibody into the supernatant-transfer assay. This resulted in a complete abolishment of the response (Fig 8G–8I), while adding an anti-IL-1β antibody had no effect on G-CSF induction by TDK-derived factors (Fig 8I), consistent with the notion that IL-1β was not produced by murine keratinocytes (Fig 7B). IL-1α was not only necessary but also sufficient for triggering G-CSF production in endothelial cells, because MS1 cells secreted large quantities of G-CSF when stimulated with recombinant IL-1α (Fig 8J). In summary, these results revealed a novel IL-1α-dependent crosstalk between epithelial and endothelial cells that mediates the induction of G-CSF by C. albicans.
In conclusion, our data corroborate the notion that the epithelium takes an active part in host defense in barrier tissue through its strategic location and by alerting the immune system about the presence of a pathogenic threat. By demonstrating the relevance of IL-1α release (for G-CSF induction) and chemokine secretion, we provided an example in a physiologically relevant system.
In this study, we describe an essential function of the IL-1 pathway in antifungal immunity in the murine oral mucosa. We demonstrated how IL-1R signaling regulates the neutrophil response against C. albicans in two ways to prevent fungal growth and dissemination. Specifically, it promotes the production of chemokines by oral keratinocytes for neutrophil recruitment from the circulating pool, and it induces G-CSF secretion from endothelial cells to enhance granulopoiesis in the bone marrow to meet the rapid demand for neutrophils in the tissue. IL-1R signaling thereby translates the local response to a tissue-specific infection into a systemic response. The availability of preformed IL-1α from keratinocytes, which are the first cells to be exposed to C. albicans during infection, is critical for the initiation of the response. Together, our data demonstrate how signaling through the IL-1R coordinates a cellular crosstalk between keratinocytes, endothelial cells and neutrophils for optimal control of C. albicans in the oral mucosa.
Rapid infiltration of neutrophils to the site of infection is a hallmark of the inflammatory response to OPC and critical for the confinement of the fungus in the mucosal epithelium [6]. The neutrophil response in the oral mucosa was originally thought to be controlled by the IL-17 pathway, which itself is highly critical for fungal control during OPC [14]. However, previous work from our laboratory showed that the key protective function of IL-17 is uncoupled from the neutrophil response [6]. Although IL-17 signaling can enhance the expression of neutrophil chemokines and granulopoietic factors [14], IL-17 is not required for neutrophil chemotaxis and function during OPC [6]. Instead, neutrophil trafficking to the oral mucosa during acute infection is under the control of IL-1R signaling as our data demonstrate. IL-1R deficiency is associated with impaired neutrophil recruitment and defective fungal control in response to OPC. Our data are consistent with previous reports demonstrating a role for IL-1R signaling in antifungal defense in different settings including systemic candidiasis [24], a model of mixed oral and systemic candidiasis [34], A. fumigatus keratitis [35] and invasive pulmonary aspergillosis [36]. The mechanism of IL-1R-mediated protection, however, was not addressed in most of these studies. Here, we used the model of OPC to dissect the impact of IL-1R signaling on neutrophil mobilization and recruitment in response to C. albicans infection at a cellular and molecular level.
Keratinocytes take center stage in the coordination of the IL-1-mediated neutrophil response during OPC. They act as the major producers of neutrophil-recruiting chemokines. Chemokine production by keratinocytes was greatly enhanced by IL-1R signaling and at least in part through the autocrine activity of IL-1α. A similar mechanism for enhanced chemokine secretion by keratinocytes was described before in the context of Staphylococcus aureus skin infection [37]. A second mechanism, by which keratinocytes promote the neutrophil response during OPC, is by the induction of G-CSF via their capacity to produce IL-1α. Our data thus link IL-1 signaling and G-CSF production. G-CSF regulates granulopoiesis and neutrophil mobilization in the bone marrow, which is critical to meet the rapidly increasing demand of neutrophils in the infected tissue. Administration of recombinant G-CSF in mice is sufficient to drive emergency granulopoiesis as we and others have shown [38], and recombinant G-CSF is widely used therapeutically to treat neutropenia [39]. G-CSF has also been used successfully in the treatment of chronic mucocutaneous candidiasis [40]. Besides its host-beneficial role in hematopoiesis and protection from infection, dysregulated production of G-CSF has been linked to autoinflammatory disorders such as psoriasis and inflammatory arthritis [41–43] and targeting G-CSF has been proposed as a therapeutic approach against these diseases [44].
We identified endothelial cells in the oral tissue to be the major producers of G-CSF during OPC, while we were unable to detect G-CSF expression by murine keratinocytes in response to C. albicans. Direct secretion of G-CSF into the bloodstream by endothelial cells facilitates its delivery to the distant bone marrow and the induction of a systemic response. During systemic bacterial infections, G-CSF production by the endothelium was shown to be a direct and TLR4-dependent response of the vasculature to endotoxin stimulation which lead to the induction of emergency granulopoiesis [38]. During OPC, we found no evidence for direct response of the endothelium to C. albicans, and endothelial cells were not activated by the fungus in vitro. Instead, G-CSF was produced as a result of an IL-1-dependent crosstalk between keratinocytes and endothelial cells. IL-1α was critical for the secretion of G-CSF into the circulation during OPC, and it was sufficient for stimulating G-CSF production by endothelial cells in culture, consistent with a published report [31]. In contrast, TNF, which was also reported to induce G-CSF production by endothelial cells in vitro [31], did not regulate G-CSF production during OPC.
In addition to IL-1α, IL-1β also contributes to the antifungal response in vivo and at least partially compensates in absence of IL-1α. IL-1β was also induced in the oral mucosa during infection. Consistent with the notion that murine keratinocytes are unable to produce IL-1β [28], we found this cytokine to be expressed by the hematopoietic compartment. IL-1β induction in response to C. albicans was shown previously to depend on the NLRP3 inflammasome in different infection models [21,34,45] and NLRP3 was required in the hematopoietic compartment [29]. In addition to the NLRP3 inflammasome, the NLRC4 inflammasome was also shown to contribute to protection from OPC [29]. Whether its function at the level of the mucosal stroma is linked to the IL-1 pathway and whether it is involved in IL-1α production in keratinocytes was not addressed.
We found that the overall IL-1α response resulted from the release of preformed and intracellularly stored IL-1α on the one hand, and from the induction of de novo synthesis on the other hand. The release of IL-1α from keratinocytes correlates with the induction of cell damage [46], suggesting that IL-1α release may be a consequence of cell death induced by fungal invasion. It remains to be determined whether IL-1α secretion by oral keratinocytes in response to C. albicans infection depends on caspase-1 as was shown to be the case for bone marrow-derived MNPs differentiated with GM-CSF [20]. The dependence of IL-1α secretion on life and filamenting fungus is consistent with the implication of candidalysin in this process. This fungal peptide toxin has recently been identified as a critical virulence factor promoting epithelial cell damage and the release of cytoplasmic molecules, some of which may act as alarmins [47]. Importantly, our data demonstrate the biological relevance of keratinocyte-derived factors induced by C. albicans for the induction of inflammation and protection from infection in vivo, and they dissect the mechanism how IL-1 coordinates the neutrophil response against the fungus in the oral epithelium.
All mouse experiments described in this study were conducted in strict accordance with the guidelines of the Swiss Animal Protection Law and were performed under protocols approved by the Veterinary office of the Canton Zurich, Switzerland (license number 201/2012 and 183/2015). All efforts were made to minimize suffering and ensure the highest ethical and humane standards.
WT C57BL/6J mice were purchased from Janvier Elevage. Il1a-/- mice (a kind gift from Manfred Kopf) [48], Il1b-/- mice [48] (a kind gift from Manfred Kopf and Hans-Dietmar Beer) and Il1r1-/- mice [49] were bred at the Laboratory Animal Service Center (University of Zürich, Switzerland). VE-cadherin-cre x ROSA26-RFP reporter mice [30] and Tnf-/- mice (a kind gift from Annette Oxenius) [50] were bred at the Rodent Center HCI at ETH Zürich. All mice were on the C57BL/6 background, kept in specific pathogen-free conditions and used at 6–15 weeks of age. For neutrophil depletion, mice were treated with anti-Ly6G (clone 1A8, BioXCell, 150μg per mouse i.p. on day -1) and/or anti-G-CSF (clone 67604, R&D Systems, 10μg per mouse per day i.p. starting from day -1), as indicated. For inhibition of emergency granulopoiesis, mice were treated with anti-G-CSF (10μg per mouse per day i.p. on day -1 and day 0). For induction of emergency granulopoiesis, mice were injected with human recombinant G-CSF (Filgrastim, Amgen, 5 μg/mouse i.p. at 5h and 17h post-infection).
The C. albicans strain SC5314 was used for all experiments except where stated otherwise. The yeast-locked strain hgc1Δ/Δ and its revertant hgc1Δ/Δ:HGC1 [51] were obtained from N. Gow (Aberdeen). The strain pACT1-GFP [52] was obtained from C. Reis e Sousa. All strains were grown in YPD medium at 30°C for 15–18 hours. Mice were infected with 2.5x106 cfu C. albicans sublingually as described [53] without immunosuppression. Mice were monitored for morbidity and euthanized in case they showed severe signs of pain or distress. For determination of fungal burden, the tongue of euthanized animals was removed, homogenized in sterile 0.05% NP40 in H2O for 3 minutes at 25 Hz using a Tissue Lyzer (Qiagen) and serial dilutions were plated on YPD agar containing 100 μg/ml ampicillin. The detection limit corresponds to one colony divided by the mean weight of the tongues in the experiment. For heat-killing 108 yeast cells were boiled for 45 minutes. Preformed hyphae were generated in keratinocyte medium at 37°C, 5% CO2 for 24 hours.
Mice were anaesthetized with a sublethal dose of Ketamine (100 mg/kg), Xylazin (20 mg/kg) and Acepromazin (2.9 mg/kg), and perfused by injection of PBS into the right heart ventricle prior to removing the tongue and/or the bones. Tongues were cut into fine pieces and digested with DNase I (200 μg/ml, Roche) and Collagenase IV (4.8 mg/ml, Invitrogen) in PBS at 37°C for 45–60 minutes. Single cell suspensions were passed through a 70 μm strainer using ice-cold PBS supplemented with 1% FCS and 2 mM EDTA and analyzed by flow cytometry (see below). For intracellular cytokine staining, tongue leukocytes were enriched over a 40% percoll gradient before antibody staining and analysis. Bone marrow was flushed from femurs using PBS and passed through a 70 μm strainer using ice-cold PBS supplemented with 1% FCS and 2mM EDTA. Erythrocytes were lyzed using erythrocyte lysis buffer (0.3 M NH4Cl, 28 μM NaHCO3, 125 μM EDTA), and leukocytes were analyzed by flow cytometry.
Tongue-derived keratinocytes (TDKs) were obtained from WT mice. The tongue was cut to obtain ventral and dorsal parts. The dorsal part of the tongue was freed from muscle tissue with a scalpel and floated on DMEM (PAA) containing 0.8% trypsin (Thermo Fischer) for 40 minutes with the epithelial side facing upwards. After incubation, the epithelial tissue was separated from the lamina propria, and keratinocytes were isolated by incubating in DMEM medium containing DNase I (Roche, 200 μg/ml) for 30 minutes and occasional vortexing and filtering through a 70 μm strainer. The cells were pelleted, resuspended in keratinocyte medium (S1 Table, adapted from [54]) and cultivated in cell culture plates coated with collagen (Bornstein and Traub Type IV, Sigma Aldrich, 25μg/ml). The blood endothelial cell line MS1 [33] was kept in DMEM medium supplemented with 5% FCS and Penicillin/Streptomycin. Prior to stimulation experiments, TDK and MS1 cells were rested for 48 hours and then stimulated with recombinant IL-1α (Peprotech, 20 ng/ml), IL-1β (Peprotech, 20 ng/ml), anakinra (250 μg/ml), zymosan (20 μg/ml), curdlan (200 μg/ml), LPS (100 ng/ml) or C. albicans at MOI = 3 for 24 hours. Amphotericin B was added 8 hours post-stimulation where stated explicitly. Supernatants were collected for analysis. In some cases, cell lysates were generated by adding 0.1% Triton-X 100 in PBS to the supernatant-free cells. To determine the total amount of IL-1α in cells and supernatant, Triton-X 100 was added to separate culture wells at a final concentration of 0.1%. For supernatant transfer experiments, TDKs were stimulated for 24 hours as described, supernatants were removed, sterile-filtered and added to MS1 cells at a 1:3 dilution. G-CSF production by MS1 cells was determined by ELISA after 24 hours (see below). In some experiments, TDK supernatants were treated with anti-IL-1α (clone ALF-161, BioXCell, 20 μg/ml) and/or anti-IL-1β (clone B122, BioXCell, 20 μg/ml) before they were added to MS1 cells, or MS1 cells were pretreated with anakinra for 30 minutes before addition of TDK supernatant.
All antibodies were from BioLegend, if not stated otherwise. For flow cytometry analyses of neutrophils, single cell suspensions of tongues, bone marrow or blood were stained in ice-cold PBS supplemented with 1% FCS, 5 mM EDTA, and 0.02% NaN3 with LIVE/DEAD Fixable Near-IR Stain (Life Technologies), anti-CD45.2 (clone 104), anti-CD11b (clone M1/70), anti-Ly6C (clone AL-21, BD Biosciences) and anti-Ly6G (clone 1A8). For flow cytometry analyses of tongue keratinocytes and endothelial cells, single cell suspensions were stained in ice-cold PBS supplemented with 1% FCS, 5 mM EDTA, and 0.02% NaN3, with LIVE/DEAD Fixable Near-IR Stain or 7AAD (BD Pharmingen), anti-CD45.2 (clone 104), anti-EpCAM (clone G8.8) and anti-CD31 (clone MEC13.3). In some experiments, anti-podoplanin (clone 8.1.1) was included. For intracellular cytokine staining, tongue cells were first incubated in LIVE/DEAD Fixable Near-IR Stain and surface marker antibodies. After fixation and permeabilization using BD Cytofix/Cytoperm (BD Biosciences) the cells were incubated in Perm/Wash buffer (BD Biosciences) containing anti-pro-IL-1β (clone NJTEN3) or the respective isotype control antibody. Cells were acquired on a FACS LSRII (BD Biosciences) or on a FACS Gallios (Becton Coulter), and data were analyzed with FlowJo software (Tristar). For all experiments, the data were pre-gated on live single cells. For isolating tongue cell subsets by FACS sorting, single cell suspensions of five tongues were pooled, stained as described above, and sorted on a FACS AriaII (BD Biosciences) using FCS as a collection medium.
Isolation of total RNA from bulk tongues or sorted cell populations was carried out according to standard protocols using Trizol Reagent (Sigma) or Trizol LS Reagent (Life Technologies). cDNA was generated by RevertAid (Thermo Scientific). Quantitative PCR was performed using SYBR Green (Roche) and a Rotor-Gene 3000 (Corbett Research) or a QuantStudio 7 Flex (LifeTechnologies). The primers were Actb fwd 5'-CCCTGAAGTACCCCATTGAAC-3', Actb rev 5'-CTTTTCACGGTTGGCCTTAG-3'; Cxcl1 fwd 5'-CCGCTCGCTTTCTGTG-3', Cxcl1 rev 5'-GCAGCTCATTGGCGATAG-3'; Cxcl2 fwd 5'-AGTGAACTGCGCTGTCAATGC-3', Cxcl2 rev 5'-GCAAACTTTTTGACCGCCCT-3'; Cxcl5 fwd 5'- GAAAGCTAAGCGGAATGCAC-3', Cxcl5 rev 5'-GGGACAATGGTTTCCCTTTT-3'; Csf3 fwd 5'-CTTAAGTCCCTGGAGCAAGTG-3', Csf3 rev GTGGCCCAGCAACACCAG; Il1a fwd 5'-GGGAAGATTCTGAAGAAGAG-3', Il1a rev 5'-TAACAGGATATTTAGAGTCG-3'; Il1b fwd 5'-TACAGGCTCCGAGATGAACA-3', Il1b rev 5'-AGGCCACAGGTATTTTGTCG-3'. All qPCR assays were performed in duplicate and the relative gene expression (rel. expr.) of each gene was determined after normalization with β-actin transcript levels.
For the preparation of cell lysates from sorted tongue cell populations, Brefeldin A (250 μg / mouse) was injected i.p. 6 hours before removal of the tongue tissue. Brefeldin A was also added during the antibody staining and supplemented in the collection medium during cell sorting. Sorted cells were washed with PBS, and 400’000 cells of each population (60’000 for naïve leukocytes, 25’000 for LECs) were lysed in 100 μl 0.1% Triton-X 100 in PBS containing cOmplete protease inhibitor cocktail (Roche) using a Tissue Lyzer (Qiagen) for 3 minutes at 25 Hz. The G-CSF content of the lysates was analyzed immediately by ELISA (see below).
G-CSF protein in the serum, in cell lysates or in cell culture supernatants was determined by sandwich ELISA using purified anti-G-CSF (clone 67604, R&D Systems) for coating and biotinylated polyclonal rabbit anti-G-CSF (Peprotech) for detection according to standard protocols. For determination of IL-1α and IL-1β levels in cell culture supernatants, cytometric bead array assays (BD Biosciences) were performed according to the manufacturer’s instructions. The detection limits are indicated by a dotted line.
Tongues were embedded in OCT (Sakura), snap-frozen in liquid nitrogen and stored at -20°C. For immunofluorescence staining, sagittal cryosections (9 μm), were fixed with acetone for 10 minutes at room temperature and stained with anti-IL-1α (clone ALF-161, BioXCell), anti-K6 (BioLegend, poly19057), anti-K14 (BioLegend, poly19053) and 4′,6’-Diamidino-2-phenylindole dihydrochloride (DAPI, Sigma-Aldrich). Slides were mounted with Mowiol (VWR International AG), scanned with a NanoZoomer 2.0 HT (Hamamatsu Photonics K.K.) using 10x and 20x objectives, and analyzed using NDP.scan 2.5.88. All scale bars indicate 100 μm.
For whole mount confocal microscopy, tongues were longitudinally cut with a scalpel and fresh tissue was placed on 35mm microscopy dishes (IBIDI) with PBS. Z stack tile scans were acquired with an SP8 confocal microscope (Leica) equipped with a 20x APO objective (0.7 N.A.) covering large areas of the tongue tissue. 3D maximum projections and mosaic reconstructions of the acquired data were generated with IMARIS (Bitplane).
Statistical significance was determined by Student’s t-test with Welch’s correction, one-way ANOVA with Dunnett’s or Tukey’s multiple comparison test, or two-way ANOVA with Tukey’s multiple comparison test where appropriate using GraphPad Prism software with *, p < 0.05; **, p < 0.01; ***, p < 0.001; **** p < 0.0001. For data plotted on a logarithmic scale the geometric mean is indicated, and data were log-transformed before statistical analysis.
|
10.1371/journal.ppat.1006764 | Exosomes serve as novel modes of tick-borne flavivirus transmission from arthropod to human cells and facilitates dissemination of viral RNA and proteins to the vertebrate neuronal cells | Molecular determinants and mechanisms of arthropod-borne flavivirus transmission to the vertebrate host are poorly understood. In this study, we show for the first time that a cell line from medically important arthropods, such as ticks, secretes extracellular vesicles (EVs) including exosomes that mediate transmission of flavivirus RNA and proteins to the human cells. Our study shows that tick-borne Langat virus (LGTV), a model pathogen closely related to tick-borne encephalitis virus (TBEV), profusely uses arthropod exosomes for transmission of viral RNA and proteins to the human- skin keratinocytes and blood endothelial cells. Cryo-electron microscopy showed the presence of purified arthropod/neuronal exosomes with the size range of 30 to 200 nm in diameter. Both positive and negative strands of LGTV RNA and viral envelope-protein were detected inside exosomes derived from arthropod, murine and human cells. Detection of Nonstructural 1 (NS1) protein in arthropod and neuronal exosomes further suggested that exosomes contain viral proteins. Viral RNA and proteins in exosomes derived from tick and mammalian cells were secured, highly infectious and replicative in all tested evaluations. Treatment with GW4869, a selective inhibitor that blocks exosome release affected LGTV loads in both arthropod and mammalian cell-derived exosomes. Transwell-migration assays showed that exosomes derived from infected-brain-microvascular endothelial cells (that constitute the blood-brain barrier) facilitated LGTV RNA and protein transmission, crossing of the barriers and infection of neuronal cells. Neuronal infection showed abundant loads of both tick-borne LGTV and mosquito-borne West Nile virus RNA in exosomes. Our data also suggest that exosome-mediated LGTV viral transmission is clathrin-dependent. Collectively, our results suggest that flaviviruses uses arthropod-derived exosomes as a novel means for viral RNA and protein transmission from the vector, and the vertebrate exosomes for dissemination within the host that may subsequently allow neuroinvasion and neuropathogenesis.
| In this study we have demonstrated that cells from the medically important vector tick, secretes exosomes that mediate transmission of tick-borne Langat (LGTV) viruses from arthropod to human and other vertebrate host cells. This study not only provides evidence that suggest tick-borne pathogens use arthropod-derived exosomes for transmission from vector to mammalian cells but also use exosomes for dissemination within the vertebrate host.
| Exosomes are small membranous extracellular microvesicles (30 to 250 nm in diameter) of endocytic origin formed in late endosomal compartments (as multivesicular bodies; MVBs) of several different cell types [1–5]. Initially, exosomes were considered as garbage bins to discard the unwanted cellular or molecular components or membranous proteins from reticulocytes [6–9]. Other studies have suggested that exosomes are mere cell debris or apoptotic blebs and signs of cell death [10–12]. Recently, the role of exosomes has been highlighted in important medical research on cancer and autoimmune diseases and they are now recognized as novel therapeutic targets for neurological disorders such as Parkinson’s disease [11,13–16]. Over the past 10 years, exosomes have been given potential biological significance by identifying a variety of their specific roles [3,5,11,17–20]. Exosomes derived from several different cells have been shown to function as signaling related vesicles, transporting cell-specific collections of several proteins, lipids and nucleic acids such as DNA, RNA and microRNA [12,20–28]. Exosomes are released into circulation after the fusion with the plasma membrane and these vesicles serve as mediators of molecular transmission [3,10,18,29]. Cell-derived exosomes have been shown to be important modes of intercellular communication and as transmitters of information over longer distances for e.g., between different tissues or multiple organs [2,15,27,30,31].
Studies have also shown that exosomes are vehicles of transmission for a variety of microorganisms and that some pathogens uses exosomes to manipulate their environments [10,15,32–34]. As an example, malaria parasites, Plasmodium falciparum, uses exosomes for communication between infected red blood cells [35]. Hepatitis C virus (HCV), an enveloped RNA virus, associates with exosomes isolated from cell culture supernatants and from infected patients [36,37]. Recent findings of HCV transmission through hepatic exosomes establish infection provides new insight into hepatitis drug discovery [38,39]. Exosomes also function in the transfer of immuno-stimulatory viral RNA from HCV-infected cells to co-cultured plasmacytoid dendritic cells [32]. In addition, exosomes facilitate receptor-independent transmission of replication-competent HCV viral RNA that was found to be in complex with Ago2-miR122-HSP90 in HCV-infected individuals or infected hepatocytes [36]. Interestingly, exosomes have been shown to play dual roles in transmitting Hepatitis A virus (HAV) and HCV, thereby evading antibody-mediated immune responses [40]. It has been demonstrated that Toll-like receptor 3 (TLR-3) activated macrophages release exosomes containing anti-HCV micro (miRNA)-29 family members that suggest a novel antiviral mechanism against HCV infections [41]. Herpes Simplex-1 virus and Epstein-Barr virus also use exosomes for transmission [42,43]. Several studies have suggested exosomes as important players in HIV-1 pathogenesis [33,34,44]. HIV Nef protein secreted in exosomes has been shown to trigger apoptosis in CD4+ T cells and the Gag p17 coding RNA is also targeted to the exosomes [45,46]. HIV-infected cell-derived exosomes have been shown to contain the TAR (Trans-Activation Response Element) miRNA that facilitates production of pro-inflammatory cytokines [47,48]. Moreover, a recent but very highlighting study showed that exosomes from uninfected cells activates the transcription of latent HIV-1 [49].
Ixodes ticks transmit several viruses belonging to the family Flaviviridae such as tick-borne encephalitis virus (TBEV), Powassan virus (POWV) and Langat virus (LGTV) [50–53]. LGTV is considered as a model biosafety level 2 (BSL2) pathogen to study pathogenesis of TBEV, due to its significant genome homology with the later. Transmission modes of these arthropod-borne flaviviruses (with positive sense single-stranded RNA) are poorly understood [37,54]. Our study shows for the first time that exosomes facilitate transmission of flavivirus RNA and proteins from arthropod to human cells. We have demonstrated that cells from the medically important vector tick, Ixodes scapularis, secretes exosomes that mediate transmission of tick-borne LGTV RNA and proteins from arthropod to human. Our study shows the presence of abundant amounts of LGTV RNA and proteins in exosomes isolated from arthropod and neuronal cells. We also found that LGTV-infected tick cell-derived exosomes were capable of transmigrating and infecting naïve human skin keratinocytes (the initial barrier lining the human cells that comes in contact during bites from infected ticks) and human vascular endothelial cells (that comes in contact during arthropod blood feeding). Our data show that vertebrate exosomes mediate transmission of tick-borne LGTV RNA and proteins from infected-brain microvascular endothelial cells (a component of the blood-brain barrier; BBB) to neuronal cells. In addition, we have demonstrated that exosomes containing tick-borne LGTV and mosquito-borne West Nile virus (WNV) facilitate transmission of viral RNA and proteins from one neuronal cell to others suggesting their novel role in neuropathogenesis. Dihydrochloride hydrate, GW4869 (a selective inhibitor for neutral sphingomyelinase; N-SMase, an enzyme that regulates production and release of exosomes), reduced LGTV loads in exosomes and inhibited the transmission of LGTV RNA and proteins in both arthropod and vertebrate host cells. Overall, our study suggests that exosomes are not only the mediators for transmission of arthropod-borne flavivirus RNA and proteins from arthropod to the vertebrate host, but also facilitate dissemination of these infectious RNA and proteins within the vertebrate host, including crossing of BBB cells and allowing neuroinvasion and neuropathogenesis in the Central Nervous System (CNS).
Despite the significance of ticks as important medical vectors, we know little about the transmission modes of tick-borne viruses and other tick-borne pathogens to the vertebrate host. We first analyzed whether tick cells secrete extracellular vesicles (EVs) and exosomes and if tick-borne flaviviruses use those exosomes as modes of pathogen transmission. LGTV, a flavivirus closely related to TBEV, readily infected Ixodes scapularis ISE6 tick cells, with increased viremia at 72 h post-infection (p.i.) (S1A Fig), similar to the viral infection kinetics observed in Vero cells (S1B Fig). We selected 72 h p.i. as the time point for the isolation of exosomes from tick cells due to the higher viral loads. First, we isolated exosomes by density gradient centrifugation technique using OptiPrep (DG-Exo-iso) as described in [55]. This isolation method used in our settings with a floor ultracentrifuge unit is shown as a schematic representation in (S1C Fig). Exosomes were also independently isolated by differential ultracentrifugation with slight modifications and longer spin times for 155 minutes (S2 Fig) [22,29,32,56,57]. We also isolated arthropod-derived exosomes using commercially available exosome isolation reagent following manufacturer's instructions (Invitrogen/ThermoScientific). Notably, all preparations contained 30 to 200 nm vesicles and these techniques have been used extensively in several studies. Cryo- Electron Microscopy (cryo-EM) performed on tick cell-derived exosomal fractions showed the presence of purified arthropod exosomes with the size range of 30 to 200 nm in diameter (Fig 1A), similar to exosomes isolated from mammalian cells [1–3]. Exosomes isolated from arthropod cells showed a heterogenous population of vesicles in the cryo-EM analysis. In order to understand such heterogeneity in exosome populations, we did quantitative analysis using images collected from both uninfected and LGTV-infected tick cell-derived exosomes. We noted that majority of the exosomes were of sizes between 50–100 nm in both uninfected and infected groups (Fig 1B and 1C). However, exosomes of other sizes 100–150 and 150–200 nm were evenly distributed in infected group when compared to the uninfected group (Fig 1B and 1C). Fewer vesicles from sizes of 200–250 nm were slightly more in uninfected (10.1%) in comparison to the infected group (6.5%). The large exosomes were very few and were from 0–1.5% in both the uninfected and infected groups (Fig 1B and 1C). Counting of exosomes per image showed higher number of exosomes in LGTV-infected (n = 14) in comparison to the uninfected (n = 27) group (Fig 1D). This data suggested that LGTV-infection (72 h p.i.) might enhance the production and/or release of exosomes.
The OptiPrep (DG-Exo-iso) method yielded purified exosomes in six different fractions. Immunoblotting analysis (with highly cross-reactive 4G2 monoclonal antibody that recognizes the viral Envelope (E)- protein) of these fractions (20 μl) showed presence of LGTV E-protein in all six fractions but enriched amounts of E-protein were present in fractions four and five in comparison to the other fractions (Fig 1E). These results correlated with the size analysis data (Fig 1B, 1C and 1D). Enhanced detection of LGTV-E protein in fractions four may correspond to the 50–100 nm (fraction 4) size exosomes that are highly populated (Fig 1E). As expected, we did not detected E-protein in the fractions from uninfected control. Total cell lysates (20 μg) from uninfected and LGTV-infected groups were used as internal controls to compare the amounts of E-protein detected in the 20 μl of different fractions used (Fig 1E). The PonceauS images showing the protein profile serve as control (Fig 1E). Quantitative Real-Time PCR (QRT-PCR) analysis revealed presence of LGTV total mRNA in exosomes isolated from infected tick cells (Fig 1F). The copy numbers of viral RNA in exosomes derived from LGTV-infected (72 h p.i.) tick cells is shown in (S3A Fig). In addition, we also determined the presence of both positive and negative sense LGTV RNA strands in tick cell-derived exosomes (Fig 1G). LGTV mRNA was also evident in exosomes from tick cells cultured and infected in exosome-free FBS medium (with no cross-contaminating bovine exosomes present in regular commercial FBS), further suggesting the presence of viral RNA in tick cell-derived exosomes (S3B Fig). Presence of LGTV E-protein in tick cell-derived exosomes was further recognized by SDS-PAGE followed by immunoblotting with 4G2 antibody (Fig 1H). Higher E-protein loads were detected (at ~50kDa) in total cell lysates in comparison to exosomal preparations (Fig 1H). Immunoblotting with monoclonal anti-Langat virus NS1 (Clone 6E11) antibody (obtained from BEI resources) also showed the presence of NS1 in both tick cell-derived exosomes and total cell lysates (Fig 1H). Although, higher NS1 protein loads were evident in total lysates, but the presence of NS1 in tick-cell derived exosomes (Fig 1H) further confirmed that these arthropod exosomes contain LGTV proteins. Remarkably, we also detected the presence of tick HSP70 (heat-shock cognate protein 70, a specific exosomal marker in mammalian cells) in exosomal lysates (Fig 1H). No differences were noted in HSP70 loads between uninfected and infected exosomal lysates (Fig 1H). Presumably due to low amount in cell lysates, no HSP70 was detected in the tested condition (Fig 1H). Total protein lysates prepared from same batch of uninfected or LGTV-infected tick cell-derived exosomes or from whole tick cells served as loading control for all immunoblots (Fig 1H). It was noted that some of the bands in the total protein profile gel were enhanced in LGTV-infected tick exosome lysates in comparison to the uninfected controls (Fig 1H).
Furthermore, native-PAGE followed by immunoblotting with 4G2 antibody, showed enhanced levels of LGTV E-protein (at <250kDa; in native state) in exosomes treated (30 min, RT) with Triton-X-100 (a detergent that lyses the exosomal lipid bilayer membranes) in comparison to the exosomes treated for three rounds of freeze-thaw cycle (1 h, at -80°C) or the untreated exosomes held at 4°C (Fig 1I). Total protein lysates prepared from uninfected or LGTV-infected tick cell-derived exosomes with similar treatments served as controls in this immunoblotting analysis (Fig 1I). Detection of LGTV E-protein inside exosomes (but not outside in the PBS suspensions) was further analyzed by E-protein-4G2-antibody-beads binding assay as described in methods. No significant (P>0.05) differences in viral loads were observed in LGTV-infected (72 h p.i.) exosome samples that were either untreated or treated with 4G2 antibody (that binds to LGTV E-protein) or relevant isotype control antibody (Fig 1J). Similar results were obtained with LGTV-infected exosomal preparations derived from GW4869 inhibitor treated tick cells collected at 72 h p.i., (Fig 1J). Native-PAGE and the beads assay clearly suggest that exosomes contain viral RNA and proteins inside exosomes. To further evaluate if viral E-protein is indeed (totally) inside the exosomes, we performed the protease-resistance assay with Proteinase K that generally digest proteins in all biological samples. We found that treatment with Proteinase K (0.5 μg/μl or 50 μg/ml, 15 min at 37 ºC) at typical and suggested working concentrations (50–100 μg/ml) digested all proteins (S3C Fig). We detected E-protein in untreated infected samples but not in treated infected samples. Uninfected samples either treated or untreated served as internal controls (S3C Fig). The Ponceau S stained blot showed no proteins in infected or uninfected proteinase K-treated samples (S3C Fig). During isolation of tick exosomes, pellet fraction (containing exosomes) and supernatant fraction (generated after pelleting exosomes and before PBS wash; See S2 Fig) was tested in plaque assays to determine infectivity and replication of viral RNA and titers as described in methods. Plaque assays performed with the tick cell-derived exosome pellet fractions yielded plaques at dilutions of 1:10 and 1:100 that were too numerous to count, and around 20–22 plaques at a dilution of 1:1000 (Fig 1K and S3D and S3E Fig). No plaques were detected in plates where Vero cells were treated with the supernatant fractions at any tested dilution (Fig 1K and S3D and S3E Fig). Plaque assays indicated the presence of infectious viral RNA or proteins in LGTV-infected exosomes that resulted in high loads of LGTV in Vero cells (6.6 x 104 pfu/ml) and increased formation of viral plaques. Plaque assays further confirmed that tick cell-derived exosomes contain LGTV RNA and proteins capable of replication and forming viable plaques that are highly infectious to mammalian cells (Fig 1K and S3D and S3E Fig). No detection of viral plaques in the supernatant fractions suggests presence of abundant amounts of LGTV RNA and proteins in exosomes (Fig 1K and S3D and S3E Fig). Overall, these results suggest that majority of the LGTV RNA and proteins exit tick cells via exosomes and that exosomes could mediate transmission of these and possibly the other closely related viruses such as TBEV and POWV.
As tick-borne viruses (including TBEV, LGTV and Powassan virus) are transmitted by an infected tick bite to the vertebrate hosts, we tested whether exosomes isolated from LGTV-infected tick cells are infectious to human cells. In an infection kinetics assay, LGTV readily infected human keratinocytes (HaCaT cells) at all tested time points (24, 48 and 72 h p.i.) and there were no changes in viral loads at different times p.i. (S3F Fig). Infection of HaCaT cells with exosome fraction prepared from LGTV-infected tick cells (72 h p.i.) showed significantly (P<0.05) increased levels of viral loads at 72 h p.i. in comparison to HaCaT cells treated with supernatant fractions prepared from 72 h post-infected-tick cells (Fig 1L). Tick cell-derived exosomes containing LGTV grown in the presence of exosome-free FBS medium were also found to be infectious to HaCaT cells (S3G Fig). However, LGTV was not detectable in HaCaT cells (grown in exosome-free FBS medium) treated with the supernatant fractions (S3G Fig). Our data also showed that LGTV (laboratory viral stocks, prepared from Vero cells) was capable of infecting human vascular endothelial (HUVEC) cells with no differences in viral loads at 24 h p.i. in comparison to later tested time points (48 and 72 h p.i.) (S3H Fig). HUVEC cells treated with exosomes-containing-LGTV showed significantly (P<0.05) increased viral loads at 48 h p.i. in comparison to the cells treated with supernatant fractions, suggesting that LGTV RNA is enriched in exosomes (S3I Fig). We then performed transwell assays (as described in the methods) to test whether tick exosomes mediate transmission of LGTV from infected tick cells (plated in upper inserts) to uninfected/naïve human keratinocytes (seeded into the lower well). We found that tick cells treated with infected tick-cell-derived exosomes (that were isolated from independent batch of LGTV-infected tick cells) readily transmitted infectious exosomes to uninfected HaCaT cells (Fig 1M). However, upon incubations with tick cell-derived exosomes collected from GW4869 (5 μM) treated cells, significantly (P<0.05) reduced transmission of viral RNA to HaCaT cells was noted (Fig 1M). Infection of arthropod cells with laboratory virus stocks with known titers (MOI 1) served as control in this assay (Fig 1M). Taken together, these results suggest that LGTV infectious RNA and proteins are transmitted to human cells via arthropod exosomes.
Upon transmission to the vertebrate host, arthropod-borne neurotropic encephalitis viruses are known to first replicate in the blood and peripheral tissues (spleen and liver), cross the BBB and invade the CNS [58,59]. We used mouse brain-microvascular endothelial cells (bEnd.3 cells; that constitutes the BBB) to test whether LGTV infectious RNA and viral proteins are transmitted to neuronal cells via bEnd.3 cell-derived exosomes. LGTV readily infected and replicated in bEnd.3 cells at all tested time points (48 and 72 h p.i.) (S4A Fig). In addition, we found that the viral loads in brain endothelial cells were not significantly different over the infection period as revealed by the viral loads at much later time points (96 and 120 h p.i.) (S4B Fig). QRT-PCR analysis revealed significantly (P<0.05) increased viral burden and copy numbers in exosomes isolated from bEnd.3 cells at 24 h p.i. in comparison to the other tested time points (48, 72, 96 and 120 h p.i.) (Fig 2A and S4C Fig). We also detected higher loads of LGTV positive and negative sense RNA strands at 24 and 48 h p.i., in comparison to the other tested time points (72, 96 and 120 h p.i.) (Fig 2B). LGTV infected and replicated in neuronal cells (mouse N2a cells) in a time-dependent manner with peak level of infection at 72 h p.i. (S4D Fig). N2a cells were then infected with bEnd.3 cell-derived exosomes collected at different time points (24 and 48 h p.i.). LGTV RNA and proteins containing exosomes from bEnd.3 cells were found to be infectious to N2a cells with peak level of infection observed with exosomes isolated from endothelial cells at 48 h (p.i.) (Fig 2C). N2a cells treated with supernatant fractions (collected at the indicated time points) derived from endothelial cells resulted in significantly (P<0.05) lower viral loads in comparison to the treatments with exosome fractions isolated from the bEnd.3 cells (Fig 2C). Transwell assays performed with exosomes isolated from LGTV-infected-brain endothelial cells showed transmission of viral RNA and proteins from bEnd.3 cells (plated in upper inserts) to uninfected/naïve N2a cells seeded in the lower well (Fig 2D). Presence of exosome inhibitor significantly reduced transmission of LGTV infectious RNA from bEnd.3 cell-derived exosomes to N2a cells (Fig 2D). Infection of bEnd.3 cells with laboratory virus stocks with known titers (6 MOI) showed transmission of LGTV to N2a cells (by crossing the membrane barriers in transwell plates) and served as control in this assay (Fig 2D). These results suggest that exosomes derived from brain-endothelial cells are perhaps the mediators for BBB permeability (crossing of infectious exosomes from infected-endothelial cells lining the BBB and transmission to the neuronal cells) that may facilitate neuroinvasion of tick-borne LGTV and possibly TBEV and POWV.
Upon entry in to the brain, tick-borne neuroinvasive viruses (such as TBEV) infects neuronal cells [60]. To test whether transmission of these viruses within the brain from one neuronal cell to another is mediated by exosomes; we first infected N2a cells with LGTV (S4C Fig). Cryo-EM showed the presence of purified exosome preparations from neuronal cell-derived exosomal fractions with the size range of 30 to 200 nm in diameter (Fig 3A), similar to exosomes isolated from tick cells. Also, we isolated exosomes by precipitation using the commercially available kit isolation reagent following the manufacturer’s protocol (S5A Fig). Cryo-EM images (generated using this method) showed the presence of purified exosome preparations from neuronal cell-derived exosomal fractions with the similar size range of 30 to 200 nm in diameter (S5B Fig). Like arthropod exosomes, neuronal cell-derived exosomes also showed a heterogenous population of vesicles. In a very similar way, we did quantitative analysis using cryo-EM images collected from both uninfected and LGTV-infected N2a cell-derived exosomes. Majority of these exosomes were also of sizes between 50–100 nm in both uninfected and infected groups (Fig 3B and 3C). Smaller exosomes of sizes 0–50 nm were of slight higher percentages in infected exosomes when compared to the uninfected group (Fig 3B and 3C). Fewer vesicles from sizes of 150–200 (9–11%) or 200–250 (6.3%) were found in both infected and uninfected groups. Less than 1% of larger vesicles (250–350 nm sizes) were found in infected group (Fig 3B and 3C). Counting of exosomes per image showed higher number of exosomes in LGTV-infected (n = 13) in comparison to the uninfected (n = 9) group (Fig 3D). This data suggested that LGTV-infection (72 h p.i.) might enhance the production of exosomes. The OptiPrep density gradient exosome separation (that separates exosomes from viruses and large microvesicles) yielded purified exosomes at six different fractions. Immunoblotting analysis (using 4G2 antibody) of these fractions (20 μl) showed presence of LGTV E-protein in all fractions but enriched amounts of E-protein were present in fractions four and five in comparison to the other fractions (Fig 3E). We did not detect E-protein in the fractions from uninfected control. The cell lysates (20 μg) from uninfected and infected groups were used as controls to compare the amounts of E-protein detected in the different fractions volume (Fig 3E). Immunoblotting with anti-HSP70 antibody detected enriched amounts of HSP70 (exosomal marker) in fourth fraction of both uninfected and infected samples (Fig 3E). HSP70 levels were also detected in three and five of infected fractions but not in uninfected fractions (Fig 3E). In addition to the HSP70, we also analyzed the CD9 (a protein enriched in the mammalian cell-derived exosomes and recognized as exosomal marker) levels in uninfected and infected fractions. CD9 was detected in all six of the uninfected fractions in an increasing manner, with higher levels in fractions four and five (Fig 3E). However, CD9 was detected in 2–5 of infected fractions with higher-level detection in fractions three and four (Fig 3E). It was interesting to note that LGTV E-protein was enhanced in similar exosomal fractions (fractions 3–5) that had enhanced loads of both HSP70 and CD9, suggesting that infectious exosomes in fraction four have higher levels of exosomal markers. OptiPrep DG-isolation of exosomes using 0.1 μm filter (culture supernatants were filtered before concentration and processing for gradient steps) detected E-protein also in the fraction 4, suggesting that these infectious exosomes have sizes of 50–100 nm (Fig 3E). This data also correlated with the quantitative analysis from cryo-EM images. In order to address, where the intact LGTV particles may run on the parallel gradients, we performed OptiPrep DG-isolation on the laboratory stocks of LGTV (prepared in Vero cells, collected at 7–14 days post-infection and stored at -80°C). We noted a differential pattern in E-protein loads when density gradients were performed on LGTV-infected exosomal fractions from N2a cells (Fig 3E) or on LGTV laboratory stocks containing viruses (S5C Fig). An enhanced E-protein signal was detected in fraction 6 (indicating presence of virions in this fraction) and not in fraction 5. Detection of E-protein in fractions 4, 3 and 2 from the laboratory virus stock suggested the presence of infectious exosomes containing viral E-protein (S5C Fig). This data indicated that the viral stocks are not just the virions but are perhaps mixtures of infectious exosomes containing viral E protein.
Upon LGTV infection of N2a cells, exosomes were collected at different time points (24, 48, 72 h p.i.) and analyzed for viral loads. QRT-PCR analysis revealed an increased total viral RNA load and copy numbers at 72 h p.i. in comparison to the other tested time points (24 and 48 h p.i.) (Fig 3F and 3G). Both positive- and negative-sense RNA was detected at higher levels in the exosomes isolated from N2a cells at 72 h p.i. in comparison to the other tested time points (Fig 3H). Exosomes collected from the kit reagent also yielded similar results with increased LGTV loads in exosomes (S5D Fig). Next, we addressed the possibility that viral RNA could be binding to the outside of the exosomes and may be transmitted to the recipient cells. In order to test this possibility, we treated freshly prepared LGTV-infected (72 h p.i.)- N2a cell-derived exosomes with RNase A (5 μg/ml, for 15 min, at 37°C). We did not find any differences in LGTV loads from infected treated or untreated groups (Fig 3I). The uninfected group treated with RNase A was kept as internal control (Fig 3I). In addition, we treated freshly derived exosomes isolated from LGTV-infected N2a cells, with Triton-X-100 (0.1%; for 45 min, at RT), followed by treatments with RNaseA (5 μg/ml, for 15 min, at 37°C). QRT-PCR analysis showed that exosomes treated with both Triton-X-100 and RNaseA has lower LGTV loads in comparison to exosomes not treated with RNaseA (S5F Fig). Immunoblotting analysis further suggested the presence of LGTV E-protein in the exosomes isolated from N2a cells (Fig 3J). The E-protein loads were one-or two-fold higher in total lysates in comparison to the exosomal lysates derived from N2a cells (Fig 3J). Reduced molecular mass of LGTV E protein was found in exosomes derived from N2a cells in comparison to the total lysates (Fig 3J), suggesting a possible de-glycosylation of the viral E protein in neuronal cell-derived exosomes. We found similar de-glycosylation of the viral E protein in immunoblots performed on laboratory virus stocks (S5E Fig). A high level of CD9 was detected in the LGTV-infected N2a cell-derived exosomes in comparison to low levels in the uninfected control and the total cell lysates prepared from LGTV-infected or uninfected N2a whole cells (Fig 3J). Total protein lysates used in the immunoblot analysis served as loading control (Fig 3J). Enhanced levels of LGTV- E protein in neuronal exosomes treated with Triton-X-100 (0.03%; for 30 min, RT) in comparison to the exosomes treated after freeze-thaw cycle (thrice frozen and thawed at -80°C) or untreated exosomes held at 4°C was detected by native-PAGE followed by immunoblotting with 4G2 antibody (Fig 3K). We noticed that E-protein was detected at higher molecular mass (<250kDa) in neuronal exosomes when samples were processed for native-PAGE analysis under non-reducing and non-denaturation conditions. Detection of NS1 protein in independent samples at the similar molecular mass suggests presence of other LGTV proteins or polyprotein in exosomes (Fig 3K). Exosomes derived from uninfected N2a cells served as control (Fig 3K). Total protein lysates prepared from uninfected or LGTV-infected neuronal cell-derived exosomes after freeze-thaw or Triton-X-100 treatments or untreated samples served as loading control (Fig 3K). ELISA corroborate results of the native-PAGE, where higher loads of LGTV E-protein were detected when exosomes were treated with 0.1% of Triton-X-100 in comparison to untreated exosomal fractions (Fig 3L). Lower level of E-protein in LGTV-infected untreated neuronal exosomes was considered as background signal due to non-specific antibody binding (Fig 3L).
Furthermore, we analyzed the presence of E-protein inside neuronal exosomes by a 4G2-antibody-coated bead-binding assay as described in methods (Fig 3M). No significant (P>0.05) differences in viral loads were observed in LGTV-infected (72 h p.i.) neuronal exosome samples that were untreated/treated with either 4G2 antibody or isotype control antibody (Fig 3M). GW4869 inhibitor treated exosomes from LGTV-infected neuronal cells collected at 72 h p.i., followed by treatments with either 4G2 or isotype control also showed no significant (P>0.05) differences in viral load in comparison to untreated samples (Fig 3M). However, a significant decrease in LGTV loads were observed in the inhibitor treated group in comparison to no-inhibitor treated group (Fig 3M). In addition, we found that exosomes treated with Proteinase K (100 μg/μl, 15 min at 37°C) may be digested all proteins on the surface, thereby, lysing the vesicles and allowing degradation of the exosomal luminal proteins (S5G Fig). We detected E-protein in infected- untreated samples but not in treated samples. Untreated, uninfected samples serve as internal controls (S5G Fig). The Ponceau S stained blot showed no proteins upon Proteinase K treatment (S5G Fig). Plaque assays further confirmed that exosomes isolated from LGTV-infected N2a cells contain infectious viral RNA, with a significantly higher number of plaques, evident upon infection with exosome fractions in comparison to the infection with supernatant fractions (Fig 4A and S6A and S6B Fig). Furthermore, infectious exosomes containing LGTV RNA and proteins prepared from N2a cells at different time points (24, 48, 72 h p.i.) were capable of re-infecting naïve N2a cells (Fig 4B). Significantly higher level of viral burden was evident in N2a cells freshly infected with LGTV-containing exosome fractions prepared from 48 or 72 h (p.i.) in comparison to the infection with exosome fractions prepared from 24 h p.i. (Fig 4B). Re-infection with supernatant fractions showed undetectable levels of LGTV (Fig 4B). Similar levels of viral re-infection kinetics were observed upon incubations with LGTV-infected N2a cell-derived exosomes isolated using commercially available isolation reagent that were used to infect naïve/fresh N2a cells (S6C Fig). To find, if mosquito-borne flaviviruses such as WNV viral RNA is also present in exosomes, mouse N2a cells were infected with WNV. Viral infection kinetics showed that WNV readily infected N2a cells with increased viremia at 72 h p.i. (Fig 4C). Also, exosomes derived from WNV-infected N2a cells showed a peak in viral burden at 72 h p.i (Fig 4D), suggesting that WNV RNA is also present in exosomes.
We treated (4 h) N2a cells with 5 μg of 4G2 monoclonal antibody, followed by infection with exosomes from LGTV-infected (72 h p.i.) N2a cells to analyze if treatment with 4G2 antibody affects viral transmission. No differences were found in antibody treated or untreated groups (Fig 4E). Next, we determined if exosome mediated viral transmission is receptor-dependent and requires clathrin-mediated endocytosis. We treated N2a cells with clathrin specific inhibitor (Pitstop-2; 30 μM and 15 min), and infected these clathrin-inhibitor treated cells with infectious (LGTV; 72 h p.i.) exosomes derived from N2a cells. We noted significant (P<0.05) reduction in LGTV loads (72 h p.i.) in Pitstop-2 treated cells in comparison to the DMSO (vehicle) treated controls (Fig 4F). These results suggest that exosome-mediated LGTV transmission to naïve cells is receptor-dependent endocytosis that requires clathrin.
Presence of exosome-inhibitor at different concentrations (1, 5 and 10 μM) significantly (P<0.05) reduced LGTV loads in exosomes (from N2a cells) in comparison to DMSO-treated controls (Fig 5A). In addition, exosomes prepared from inhibitor-treated (1 μM) N2a cells were significantly reduced in re-infecting naïve N2a cells in comparison to DMSO-treated control group (Fig 5B). Furthermore, we found that exosomes isolated from N2a cells pre-treated with 5 μM inhibitor before LGTV infection had significantly lower viral loads in comparison to exosomes isolated from cells post-treated with inhibitor after infection (Fig 5C). However, viral loads in exosomes were significantly reduced in N2a cells irrespective of pre- or post- inhibitor treatment in comparison to the infection performed with LGTV from laboratory viral stocks with known titers (5 MOI) (Fig 5C). Plaque assays performed with LGTV-infected N2a cell-derived exosomes isolated from DMSO-treated group yielded significantly (P<0.05) increased number of plaques in comparison to the number of plaques with exosomes isolated from inhibitor-treated group (Fig 5D and 5E). Also, plaque assays performed with exosome fractions from N2a cells revealed the viral titers for both N2a-DMSO control group (8 x 103 pfu/ml) and N2a 1 μM-inhibitor treated group (2.3 x 103 pfu/ml). We also determined the effects of GW4869 inhibitor on LGTV viral particles from laboratory virus stocks. Immunoblotting with 4G2 antibody showed no differences in 5 or 10 μM treated (4 h) groups, in comparison to the DMSO control (Fig 5F). This data suggested that GW4869 has no effect on viral particles.
To analyze whether exosomes are the mediators of viral transmission from one neuronal cell to other in an in vivo model, LGTV infections were performed on primary neuronal cultures of murine cortical neurons (isolated from embryonic day E16 brains, as described in methods). Infection of cortical neurons with LGTV (MOI 4) showed time dependent kinetics of LGTV infection with increased viral burden at 72–96 h p.i. (Fig 6A). QRT-PCR analysis revealed significantly (P<0.05) increased LGTV total loads and copy numbers in exosomes isolated from murine cortical neurons at 72 h (p.i.) when compared to exosomes isolated from other tested time points (24 and 48 h p.i.) (Fig 6B and 6C). We also detected higher loads of LGTV positive- and negative- sense RNA strands, suggesting presence of both viral genomes in the exosomes derived from infected-cortical neurons (Fig 6D). Immunoblotting showed abundant LGTV E-protein amounts (2–3 folds) in exosomes isolated from cortical neurons in comparison to the loads found in total cell lysates (Fig 6E). Similar to N2a cells, possibly de-glycosylated LGTV E protein (with low molecular mass) was detected in exosomes isolated from cortical neurons in comparison to the total cell lysates (with high molecular mass) prepared from cortical neurons (Fig 6E). Elevated levels of CD9 (exosomal enriched marker) were found in the exosomes derived from LGTV-infected cortical neuronal cells and in total cell lysates in comparison to their respective uninfected controls (Fig 6E). In addition, levels of CD9 were dramatically elevated in exosomes from LGTV-infected cortical neuronal cells in comparison to the levels in total cell lysates (Fig 6E), supporting that LGTV infection may regulate the enrichment of CD9 in neuronal exosomes. Also, exosomes derived from LGTV-infected cortical neurons showed higher amounts of CD9, when compared to the N2a cell-derived exosomes containing LGTV (Figs 6E and 3J). Total protein profiles served as loading control in the immunoblotting analysis (Fig 6E). Immunoblotting showed presence of NS1 in both exosome fractions and in total cell lysates suggesting that exosomes from cortical neuronal cells contain LGTV proteins (Fig 6F). Plaque assays confirmed that exosomes isolated from cortical neurons carry infectious and replicative viral RNA, since significantly increased number of plaques were evident upon infection with exosome fractions (in different dilutions; 1:10, 1:100, 1:1000) in comparison to the infection with supernatant fractions (Fig 6G and S7A and S7B Fig). Similar observations were previously noted for N2a cells, suggesting that LGTV is enriched in neuronal exosomes. Additionally, exosome fractions prepared from LGTV-infected cortical neurons at different time points (24, 48, 72 h p.i.) were capable of re-infecting naïve primary cultures of cortical neurons (Fig 6H). A significant higher viral burden was evident in the cortical neurons infected with exosome fractions (prepared from 24, 48, 72 h p.i.) in comparison to the infection with the supernatant fractions prepared from respective time points (Fig 6H). These data suggest that exosomes derived from LGTV-infected neuronal cells are potential mediators for spreading infection to other neurons. Furthermore, presence of exosome-inhibitor at concentrations of 10 or 20 μM significantly (P<0.05) reduced viral infection in cortical neurons in comparison to DMSO-treated controls (Fig 7A). However, no differences in the viral burden of cortical neurons were noted upon treatment with 1 μM exosome inhibitor in comparison to DMSO-treated control (Fig 7A). This data suggested that neuronal cells in in vivo might produce higher number of exosomes that could not be inhibited with less concentration (1 μM) of inhibitor. Exosomes isolated from 10 μM-treated cortical neurons showed significantly (P<0.05) reduced re-infection of naïve cortical neurons in comparison to the infections performed with exosomes isolated from DMSO-treated control group (Fig 7B). Plaque assays confirmed that LGTV-containing exosomes isolated from DMSO-treated neurons contained viable and increased LGTV loads in comparison to exosomes isolated from 10 μM exosome inhibitor-treated group (Fig 7C and 7D). Collectively, these results suggest that LGTV and perhaps TBEV, uses exosomes as novel modes of transmission from one neuronal cell to the other.
Exosomes contribute to the transmission of intracellular information from one cell to another and from one tissue to the other [2,30,61]. Several biological implications and medical applications have been associated with the exosomes as potential mediators of communication between cells and tissues [3,20,62,63]. For the first time our study shows that exosomes are novel mediators for transmission of arthropod-borne flaviviruses that infect a wide variety of vertebrate hosts including humans. Our discovery that tick cells secrete exosomes and that these exosomes are the carriers of tick-borne LGTV (Fig 1) suggest that other tick-borne flaviviruses such as TBEV and POWV might also use this novel mode of transmission from arthropods. Cryo-EM data showed that arthropod or neuronal cell-derived exosomes are of variable sizes and were in the ranges of 30–250 nm (Figs 1A and 3A). Exosomes isolated from both arthropod and neuronal cells had majority of the exosome sizes between 50–100 nm and fewer vesicles from sizes of 200–250 nm in both uninfected and infected groups (Figs 1B, 1C, 3B and 3C), suggesting purity in isolation methods. Increased number of exosomes in LGTV-infected in comparison to the uninfected groups (Figs 1D and 3D), suggested higher production and release of exosomes. Our immediate future avenue determines the loads and activity of the neutral sphingomyelinase in LGTV-infected arthropod and neuronal cells. To make virus preparations for structural studies, concentrated supernatants or titers with 109 to 1012 PFU/ml and centrifugal forces of 200,000g are used [64], that are not similar in exosomal preparation methods. However, in order to minimize the viruses and large protein aggregates that co-sediments during ultracentrifugation, we adopted the buoyant density of exosomes for purification purposes. Continuous or discontinuous sucrose density gradient centrifugation has been used extensively to purify exosomes. However, this method does not allow separation of exosomes from viruses and macro vesicles or large microvesicles with comparable sedimentation velocities [55]. Substituting sucrose with iodoxanol (OptiPrep) in the velocity gradients using 5–40% density gradients has been shown to overcome the limitations and result in purified exosomal preparations [55]. Detection of tick HSP70 in exosomal fractions (Fig 1H), suggested it to be a novel arthropod marker that may be present in exosomes from saliva and facilitate tick feeding on vertebrate host. Our recent study reported that arthropod HSP70 may aid in the host fibrinogenolysis at the tick bite site [65]. Detection of CD9 in all uninfected fractions and enrichment in fractions four and five suggested these fractions to be exosomes. The observed shift in enrichment of CD9 in LGTV-infected fraction three and four and no detection in fractions one and six suggested presence and enrichment of other proteins or cargo (including viral E-protein in fraction four) in those fractions. Our findings showing the enrichment of both arthropod and neuronal E-protein in exosomal fractions four and five confirmed the presence of viral E-protein in exosomes (Figs 1E and 3E). We hypothesize that due to space limitation and tightly regulated cargo sorting mechanisms, exosomes are certainly filled with viral RNA and proteins that are trafficked to extracellular space and later recycled back through fusion with plasma membranes. If virions or entire viral particle are perhaps exported through exosomes, we could anticipate enclosure (or packaging) of few LGTV viruses of 40–60 nm size in ~150–200 nm diameter of arthropod/neuronal cell-derived exosomes. We did not detect any viral particles or fully assembled virions inside of the exosomes, in several of our preparations processed for cryo-electron microscopy. However, we do not exclude the possibility of viral particles presence in the exosomes. Based on our findings, we believe that if viral RNA (both positive and negative strands) and proteins are loaded into exosomes, they can be exported and subsequently transmitted to the neighboring cells and distant tissues for pathogenesis in short times. The matured virions containing positive sense RNA strand exit cells through membrane budding. On the other hand, the replicative viral RNA genome will have a negative RNA strand and are cytosolic [17,54,66–68]. Detection of both positive and negative-sense RNA strands in tick/neuronal cell-derived exosomes suggest that exosomes facilitate transmission of both negative and positive-strand RNA genomes. The higher loads of negative-strand RNA in the exosomes derived from neuronal cells implied that LGTV negative strand RNA may simply get trafficked during endocytosis/uptake by these cells. The negative-strand of RNA generally exists as dsRNA with positive-strand. Thus, it seems that dsRNA may be present inside exosomes rather than single-stranded positive or negative strand of LGTV. In addition, entry of more viral RNA and proteins inside cells via receptor-mediated endocytosis may simply force the replicative viral RNA to exit the host cell and seek other neighboring cells through exosome-mediated transmission. Our finding that exosome mediated viral transmission is dependent on clathrin (Fig 4F) further suggest an important role for exosomes as viral RNA and protein transporters.
Up-regulation of some proteins in LGTV-infected tick exosomal lysates in comparison to the uninfected controls suggests the importance of these proteins in facilitating the transmission of tick-borne flaviviruses from tick cell-derived exosomes (Fig 1H). Our current efforts are focused in identification and characterization of these important cargo proteins on arthropod exosomes that could be candidates for the development of novel transmission-blocking vaccine(s) [69]. The presence of LGTV RNA (as determined by RNase A treatment studies) and E- protein inside exosomes but not outside in suspensions of exosomal fractions, suggest that exosomes not only securely carry the viral RNA (both positive and negative strands), but also transport the essential viral E-protein into host endosomal membranes for release of viral content inside cells. Our finding that exosome mediated viral transmission is clathrin-dependent suggests a possible receptor-mediated endocytosis uptake of infectious exosomes into naïve cells. Our transwell assays with tick cell-derived exosomes and human keratinocytes (Fig 1M) suggest that tick spit/secreted saliva (that could contain exosomes loaded with LGTV viral RNA and proteins) could facilitate transmission of this virus from the bite site to the vertebrate host skin cells. No differences in time course of LGTV infection in human keratinocytes suggested that these cells may not keep persistent infection, but may transmit viruses to dendritic/other migratory immune cells in the skin at their earliest. We also assume that keratinocytes are probably highly immune tolerant and may maintain viral infections to lower peaks. Abundance of LGTV infectious RNA and proteins in exosomes also suggests that exosomes may readily facilitate the dissemination of these viral factors within the tick body (for example, from midgut, upon entry, and through hemolymph to salivary glands during transmission) or transmission through saliva to the vertebrate host upon infected arthropod bite or blood feeding. Infection of vascular endothelial cells with tick cell-derived exosomes containing LGTV infectious RNA and proteins suggests that upon tick blood feeding, arthropod exosomes facilitate infection of the blood endothelium in vertebrate host. It is noteworthy that GW4869 inhibitor significantly (P<0.05) lowered LGTV in exosomes derived from tick, bEnd.3, N2a and neuronal cells (Figs 1M, 2D, 5A and 7A). These data suggest a common pathway shared in the production and release of exosomes in both arthropod and vertebrates. Overall, these studies revealed a novel mode of flavivirus transmission from the arthropod vector to the vertebrate host via arthropod exosomes that could be envisioned as transmission-blocking strategies.
Most of the flaviviruses can infect and replicate in the vertebrate brain microvascular endothelial cells that line and guard the BBB. Infected endothelial cells allow these flaviviruses to enter and cause neuroinvasion of the CNS [70–73]. We hypothesize that initial entry of few infected exosomes derived from endothelial cells, lining the BBB may lead to virus transmission into the CNS. Infection of neuronal cells and secretion of abundant loads of infectious exosomes by neuronal cells may promote the breaching of the BBB, thereby allowing entry of higher peripheral viral loads, in addition to trafficking of immune cells from the periphery. Based on our results (Fig 2), we assume that initial batch of infected-brain microvascular endothelial (bEnd.3) cell-derived exosomes containing higher loads of infectious LGTV RNA and proteins may enter into the CNS at an early time point (24 h p.i. of endothelial cells). The higher viral loads in brain endothelial cell-derived exosomes from early time points (24 h p.i.) in comparison to lower loads in exosomes at later time points further suggest earlier transmission of viral RNA through exosomes that infects neighboring neuronal cells. We also noted that bEnd.3 cells (in infection kinetics assays) were more resistant to LGTV infection with no severe cytopathological effects when compared to neuronal cells. We hypothesize that the brain endothelial cells may not support the higher rate of viral replication or persistent infection for longer times. This could result in higher packaging of LGTV viral RNA and perhaps proteins in bEnd.3 cell-derived exosomes that would lead to dissemination of flaviviruses to neuronal cells at earlier times. The transwell assay data (Fig 2D) mimic in vivo scenario, where exosomes derived from infected-bEnd.3 cells might transmigrate through astrocyte foot layer and infect neurons in the CNS. This data could be directly related to the in vivo situation that proposes virus transmission from infected-brain microvascular endothelial cells (lining the BBB) to the interior of the CNS. Taken together, these studies imply that infected-brain endothelial cells may not entertain flavivirus replication for longer times and hence transmit these viral RNA and proteins to their neuronal counterparts at the earliest and via exosomes.
Our study also suggest that exosomes derived from neuronal cells likely able to mediate transmission of tick-borne flavivirus RNA and proteins from one neuronal cell to the other in the CNS. Higher loads of E-protein (2–4 folds more) in exosomes derived from murine cortical neurons in comparison to the in vitro cultures of N2a cell-derived exosomes suggest higher packaging of viral RNA and proteins in cortical neurons (Fig 6E). We believe that the observed lower mass for LGTV E-protein in both N2a cells and cortical neurons is due to possible de-glycosylation of the E- protein. The de-glycosylated E-protein in laboratory viral stocks suggested mixture of virions with exosomes in those frozen supernatants. However, this effect was not evident in arthropod cell-derived exosomes, suggesting that E-protein in arthropod exosomes may not undergo protein modification. Glycosylated form of LGTV E- protein in arthropod cells is maintained possibly to facilitate exosome fusion and viral infection of host cells immediately upon host seeking and tick blood feeding on vertebrate host. It has been also observed that in mosquitoes, WNV E-protein is heavily glycosylated and is required for pathogen transmission to the vertebrate host [74]. We assume that presence of de-glycosylated form of viral E-protein in neuronal and other mammalian cell-derived exosomes may allow viral E-protein to maintain its stability in these small vesicles during transmission from one cell to other. Alternatively, we also hypothesize that higher packaging of viral E-protein in vertebrate neuronal exosomes may be feasible only if E-protein exist in de-glycosylated form with less molecular mass in comparison to the glycosylated form. Also, arthropod and vertebrate host may require different conformations of E-protein that may aid when contents from exosomes are delivered to the host cytosol. The de-glycosylation of E protein in neuronal exosomes may also facilitate higher infectious ability to form matured virions in new hostile environment. Higher loads of CD9 in exosomes derived from neurons suggest that cortical neuronal cells might have greater production of exosomes in comparison to the in vitro cultured N2a cell line (Figs 3J and 6E). It is reasonable to consider that neurons have complex ways of cell communication such as synaptic transmission and neurotransmitter release that might require greater production of exosomes in the CNS. Low total protein content observed in the N2a and cortical neuronal exosomes compared to the protein content in the whole cells also suggest that only few essential proteins are imported as cargo in LGTV-infected neuronal cell-derived exosomes. Total protein profiles in N2a cells showed absence of some exosomal proteins upon LGTV infection, implying that these proteins may affect or inhibit viral proteins in N2a cell-derived exosomes. Our future studies in identifying these reduced exosomal proteins upon LGTV infection would assist in identifying novel therapeutic targets against transmission. The detection of E-protein at higher molecular mass (<250kDa) in native-PAGE gels, suggested that exosomes might contain higher order structures of E-protein as oligomers. The presence of NS1 in the same samples at similar molecular mass further indicated that exosomal fractions might contain polyprotein (Fig 3K). Presence of highly infectious LGTV RNA and proteins in exosomes from neuronal cells suggests that these cells upon infection, mediate dissemination in the CNS. Also, detection of WNV in neuronal cell-derived exosomes, further suggest exosomes as novel transmission modes for both tick- and mosquito-borne flaviviruses in neuronal cells. We assume that exosomes may maintain viability of these viral RNA and proteins that may favor persistent pathogenesis.
GW4869 (dihydrochloride hydrate) is a cell permeable but selective inhibitor for neutral sphingomyelinase (an important enzyme required for the exosome production and release). Effect of this inhibitor on both arthropod and mammalian cells used in this study suggests, presence of neutral sphingomyelinase in these cells. Treatment with GW4869 affected LGTV- replication, loads and transmigration from one cell type to other suggesting that LGTV or other tick-borne flaviviruses may use neutral sphingomyelinase or its related pathway(s) for packaging into exosomes. Future studies would unravel the role of neutral sphingomyelinase on packaging of LGTV and other flaviviruses in arthropod or mammalian exosomes. In N2a cells, 1 μM of inhibitor was sufficient to inhibit the loads of LGTV (as revealed by infection, reinfection and plaque formation) in contrast to higher doses (10 and 20 μM) of GW4869 that was required for inhibition of viral loads in primary cortical neuronal cells (Fig 7A). Higher sensitivity of N2a cells to GW4869 could be explained by the possibility of low number of exosomes or less neutral sphingomyelinase in these cells in comparison to cortical neurons. The effects of GW4869 on LGTV infection in N2a cells implied that inhibition of exosomes either before or after infection would affect LGTV loads and transmission (Fig 5C). Our data suggested that inhibition of exosomes reduced LGTV loads in both arthropod and mammalian cells and that infection with tick-borne flaviviruses was affected when exosome production and release was hampered with GW4869 treatment. No effects of GW4869 on laboratory viral stocks suggested that it is specific for blocking release of exosomes and has no direct effect on viral particles. It would be interesting to determine whether GW4869 or other novel exosome inhibitor(s) could serve as potential therapeutic approaches for treating flaviviral infections. The proposed model (Fig 8) summarizes the role of exosomes in transmission of tick-borne flavivirus RNA and proteins from the arthropod vector to human cells and dissemination of these infectious exosomes within the vertebrate host. Taken together, our study suggests that exosomes play following important roles: 1) In the transmission of tick/mosquito-borne flaviviruses from infected arthropod vector to the vertebrate host cells, 2) In the infection of the human skin keratinocytes and vascular endothelial cells during tick bite/blood feeding, 3) In mediating the infection of brain microvascular endothelial cells (lining the BBB) and crossing these infectious exosomes to allow neuroinvasion and 4) In the infection of neuronal cells resulting in high production of exosomes containing infectious viral RNA and proteins, necessary for the dissemination and infection of naïve neuronal cells in the CNS that leads to neuropathogenesis and severe neuronal loss.
Ixodes scapularis ISE6 tick cell line was obtained from Dr. Ulrike Munderloh, University of Minnesota. ISE6 cells were grown as per the culture methods provided by Dr. Munderloh [75]. Human keratinocytes (HaCaT cells) or Human Umbilical Vein Endothelial Cells (HUVEC) were obtained from Drs. Loree Heller and John Catravas laboratories, respectively. Vero (African Green Monkey kidney), mouse brain endothelial (bEnd.3 cells) and mouse neuroblastoma Neuro-2a or N2a cells were purchased from ATCC and were grown according to Company guidelines. Briefly, HaCaT, Vero, bEnd.3 and N2a cells were grown in complete DMEM medium containing 5–10% heat-inactivated FBS (Invitrogen/ ThermoScientific). HUVEC cells were grown in human lung MVEC medium (M199 medium containing 150 mg ECGF- bovine brain extract and 20% FBS) kindly provided by Dr. Catravas laboratory. To determine infection kinetics, 1 x 105 cells were seeded in a 12-well plate, infected with various multiplication of infections (MOI 1; tick cells), (MOI 6: Vero, HaCaT, HUVEC, bEnd.3 and N2a cells) of LGTV. Wild type LGTV (LGT-TP21) strain used in this study was obtained from Dr. Alexander G. Pletnev, NIAID, NIH. Cells were collected at different time points (24, 48, and 72 or 96 and 120 h post infection, p.i.) and processed for RNA or protein extractions. Details for infection studies corresponding to the data shown in different figures is mentioned in their respective Figure legends. Briefly, for infection experiments (or re-infection studies) with exosome fractions containing infectious LGTV RNA and proteins, we infected 1 x 105 HaCaT/HUVEC cells or N2a cells with 20 μl (from 150 μl) of tick (6.6 x 104 pfu/ml) or bEnd.3/N2a cells (3.5 x 103 pfu/ml) derived exosomal fractions, respectively. We used same ratio of supernatant fractions (collected from the step before PBS wash during exosome isolation) from tick or bEnd.3/N2a cells. Titers were determined after plaque countings and calculations. For studies with exosomes and exosome-free supernatant fractions, infected cells (infected with exosomes or supernatant fractions collected at different time points) were either collected at 24 or 48 or 72 h p.i. and processed for RNA extractions. For infection of mouse N2a cells with WNV, we used CT2741 wild-type strain (MOI 5) and analyzed cells for WNV loads in cells and exosomes at different time points (24, 48 and 72 h p.i.).
Exosomes were vitrified as previously described [76,77] on carbon holey film grids (R2x2 Quantifoil; Micro Tools GmbH, Jena, Germany; or C-flat, Protochips, Raleigh, North Carolina). Briefly, purified concentrated suspensions of exosomes in PBS were applied to the holey films in a volume of ca. 3 μl, blotted with filter paper, and plunged into liquid ethane cooled in a liquid nitrogen bath. We used computerized Vitrobot plunger (FEI, Hillsboro, OR) for freezing. Frozen grids were stored under liquid Nitrogen and transferred to a cryo-specimen holder (70 deg. 626, Gatan, Inc., Pleasanton, CA, or 2550 cryo-tomography holder, E.A. Fischione Instruments, Inc., Export, PA) under liquid Nitrogen before loading into a JEOL 2200FS, or a JEOL 2100 electron microscopes (JEOL Ltd., 3-1-2 Musashino, Akishima, Tokyo 196–8558, Japan). JEOL 2200FS was equipped with in-column energy filter (omega type) and a field emission gun (FEG); JEOL 2100 had a LaB6 filament, both were operating at 200 keV. Grids were maintained at near-liquid Nitrogen temperature (-172–-180°C) during imaging. Preliminary screening and imaging of exosomes was done using a 4k x 4k Gatan US4000 CCD camera (Gatan, Inc., Pleasanton, CA), and final imaging was done at indicated 40,000x magnification with a 5k x 4k Direct Electron Detector camera (DE-20, Direct Electron, Inc., San Diego, CA) using a low-dose imaging procedure. An in-column omega electron energy filter was used during imaging with a zero-loss electron energy peak selected with a 20 eV slit. Images were acquired with a ca. 20 electrons/Å2 dose; the pixel size corresponded to 1.5 Å on the specimen scale. We used a 2.0–2.3 μm defocus range for imaging. Overall, individual exosome images were acquired from two-three independent batches of exosomes from tick and N2a cells. For quantitation of exosomes size ranges, we manually analyzed the sizes using scale bar from cryo-EM images and counted exosomes per image in each group. Three independent estimations and countings were performed without any bias. Percentages (for size determination) were calculated based on the total number of exosomes in each size range. In addition, total number of exosomes/cryo-EM images were counted and analyzed.
Tick cells (1.2 x 107 cells cultured in 12 of Nunc tubes; ThermoScientific) or N2a neuronal cells (8 x 107 cells cultured in 8 different T75 flask; Greiner) were infected with either 1 MOI (tick cells; six of each tube) or 5 MOI (five of each flask with N2a cells) of LGTV. Remaining tubes (6) or flasks (4) were maintained as uninfected controls. The detailed protocol is shown as S1C Fig. Supernatants (20-50ml) from uninfected/infected cells of respective cell type were collected and centrifuged at 4°C (480g for 10 min followed by 2000g for 10min to remove cell debris and dead cells). Cell culture supernatants were either first filtered (using 0.1 μm filtering devices; VACUCAP filter for conical tubes; Pall Laboratory/VWR) and concentrated to 2–2.5ml using the Corning Spin-X UF concentrators or centrifugal filter device with a 5 k nominal molecular weight limit (NMWL). The concentrated culture medium were processed for OptiPrep (DG-Exos) isolation as described [55]. In case of OptiPrep (DG-iso) on laboratory virus stocks (7.4 x 106 pfu/ml), we added concentrated stocks of 1.5 ml supernatants directly on the gradient cushion. Briefly, discontinuous gradient of 40% (w/v), 20% (w/v), 10% (w/v) and 5% (w/v) solutions of iodixanol was prepared from the stock solution of OptiPrep 60% (w/v) of aqueous iodixanol (Axis-Shield PoC, Norway) with 0.25M Sucrose/10mM Tris, pH 7.5. We used the polycarbonate bottles with cap (Beckman Coulter) and maximum volume capacity of 26.3 ml to load the discontinuous gradient of iodixanol (4ml each of 40% (w/v), 20% (w/v), 10% (w/v) and 3 ml of 5% (w/v) from bottom to top). The cell culture supernatants (2–2.5ml) was overlaid onto the top of the gradient, and centrifuged at 100,000g for 18 h at 4°C. Six individual uninfected or infected fractions of ~3ml were collected (from top to bottom) manually (with increasing density) and diluted with 5ml of sterile PBS. Fractions were centrifuged at 100,000g for 3 h at 4°C, and followed by one more wash with 5ml of PBS and resuspended in 80 μl PBS. DG-Exos were stored in -80°C and used for analysis.
Exosomes were isolated and purified by either DG-Exo gradient method as described before or differential ultracentrifugation method as described by [29]. Details for exosome isolation procedure and modifications (used in this study) are also schematically shown and discussed in S1 and S2 Figs and in corresponding figure legends. Briefly, cells were seeded for exosomal- RNA (5 x 106 tick cells; 1 x 105 of either bEnd3.1 or N2a or murine cortical neurons) or protein (1 x 106 tick cells, 2 x 106 N2a cells or 2 x 107 cortical neurons) extractions in either 12/6-well or 10cm2 plates in complete L15, DMEM or Neurobasal medium with FBS for overnight, respectively. Next day, cells were changed to respective medium containing bovine exosome-free FBS (Systems Biosciences Inc; SBI). Tick cells plated in commercially available exosome-free FBS medium showed severe loss of cells but infectious loads were not affected. After 4–6 hour of medium replacement, cells were infected with LGTV (tick cells MOI 1; bEnd3.1 and N2a cells MOI 6; and cortical neurons MOI 4). Tick cells were susceptible to 2 or 3 MOI of infection and showed massive death, hence we used 1 MOI dose for tick cell infection studies. Cell culture supernatants were spun at 300 x g, for 10 min, cell pellet was discarded and the supernatant was spun again at 2000 x g for 10min. The pellet containing dead cells was discarded and the supernatant was spun again at 10,000 x g for 30 min to remove cell debris. Increased centrifugation times and rotor types is shown to improve exosome yield and purity [57] and, hence we used these modifications for isolation of exosomes from tick cells. Either supernatants were spun at 100,000 x g, for 70 min (for bEnd.3, N2a and cortical neurons) or for 155 min (for ISE6 tick cells). Supernatants collected after this spin step served as supernatant fractions and were used as controls in our study (indicated with * in S2 Fig). For plaque assays performed in this study 600, 60 and 6 μl and for infection studies 400 μl of supernatant fractions were used for all except HUVEC cells (300 μl). The pellets containing exosomes and any contaminants were washed one- time with ice-cold PBS and spun again at 100,000 x g, for either 70 min (for bEnd.3, N2a, cortical neuronal cells) or 155 min (for tick cells), respectively. Resulting exosomes pellet is referred as exosome fractions in this study. Freshly prepared exosome pellets were collected in PBS (and stored frozen at -80°C for re-infection studies performed on uninfected cells or for plaque assays or other tested evaluations) or resuspended in RNA lysis buffer for total RNA extractions, or in modified RIPA buffer (G-Biosciences, BioExpress) for total protein extractions. We also isolated exosomes from N2a cells using the total exosome isolation reagent and extracted total RNA and proteins using total exosome RNA and Protein Isolation kit (Invitrogen/ThermoScientific) as per the manufacturer’s instruction.
Total RNA from ISE6 tick cells, HaCaT, Vero, HUVEC, bEnd.3, N2a cells or murine cortical neurons infected with various MOI of LGTV or WNV or uninfected controls were extracted using Aurum Total RNA Mini kit (BioRad) following manufacturer’s instruction. Using BioRad iScript cDNA synthesis kit, 1 μg RNA was converted to cDNA and the generated cDNA was used as template for the amplification and determination of the viral burden. For determination of positive- or negative-sense strands of LGTV, we used the iTaq Universal SYBR Green One-Step kit (BioRad) and followed manufacturer’s instructions. For detection of positive- and negative-sense strands of LGTV RNA, we used published forward and reverse primers for Langat prM-E [78]. For WNV detection, published primers for E gene were used [73]. To normalize the amount of templates, either tick or mouse or human beta actin amplicons were quantified with published primers [73,79]. Equal amounts of tick/mouse/human cDNA samples were used in parallel for beta actin and Langat prM-E. The ratio of Langat prM-E gene copy/beta actin gene copy was used as an index to determine the rate of infection in each analyzed sample. QRT-PCR was performed using iQ-SYBR Green Supermix (BioRad, USA). Standard curves were prepared using 10-fold serial dilutions starting from standard 1 to 6 of known quantities of actin or Langat prM-E gene fragments and QRT-PCR reactions were performed as described [72,73,79]. To determine the copy number of viral RNA in exosomes, we used the LGTV RNA values with standards and converted to copy numbers using the formula: Number of copies (molecules) = (amount of amplicon) ng x 1023 molecules per mole/(length of dsDNA amplicon * 660g per mole)† *1 x 109 ng per g. Alternatively, we also used the online calculator to convert to copy numbers (http://scienceprimer.com/copy-number-calculator-for-realtime-pcr). For RNase A treatment, we isolated fresh exosomes from either uninfected or LGTV-infected N2a cells (2 x 107), distributed the infected exosomes as treated (5 μg/ml RNase, 37°C for 15 min) or untreated groups. Exosomes were also treated with Triton X-100 (0.1%, for 45 min at RT) and then followed by treatment with RNaseA as before. N2a cells (2 x 105) were infected (72 h p.i.) with these treated or untreated LGTV-infected exosomal samples were processed for RNA extractions and QRT-PCR. Untreated exosomal samples from uninfected group served as internal controls.
Western blotting was performed as described [72,73]. For DG-Exos samples, equal volume (20 μl) of each fraction from 1–6 or 20 μg of total protein lysates from uninfected and infected cells or 10 μl of each fraction from virus stock samples were loaded onto 12% SDS-PAGE, followed by immunoblotting and labeling with highly cross-reactive 4G2 monoclonal antibody to detect LGTV E-protein or exosomal specific markers such as HSP70 (rabbit polyclonal; Cell Signaling Technologies, Inc) or CD9 (mouse monoclonal; Invitrogen/ThermoScientific) and respective secondary antibodies (Santa Cruz Biotechnologies, Inc). For immunoblotting using cell lysate and exosome lysates, briefly, 5 x 106 ISE6 tick cells, or 2 x 106 N2a cells or 2 x 107 cortical neurons were seeded in 10 cm2 plates and allowed to settle/adhere for overnight. Next day, we changed the media on N2a cells and neurons to DMEM or Neurobasal medium, respectively containing bovine exosome free FBS (Systems BioSciences, Inc; SBI). ISE6 cells were retained with complete L-15 media containing 5% regular FBS (to avoid massive cell death and loss observed when processed for exosome isolation using commercially available exosome-free FBS; SBI). After 4–6 hours of media replacement, cells were infected with LGTV (tick cells, MOI 1; N2a cells, MOI 6 and cortical neurons, MOI 4). After 72 h (tick cells) or 24, 48, 72 h p.i. (N2a cells and neurons), cell culture supernatants were collected and processed for exosome isolation by ultracentrifugation (See S2 Fig). The exosome fractions collected after PBS wash and the adherent cells collected from same plates (washed twice with 1 x PBS), were resuspended in modified RIPA buffer. Total protein amounts were estimated using BCA kit (Pierce/ThermoScientific). We loaded 25 μg (tick cells) or 30–35 μg (N2a and cortical neurons) of total cell lysates or total exosomal proteins and separated them on either 12% (Laboratory casted) or precasted 4–20% SDS-PAGE gradient stain-free gels (NuSep; BioExpress). Followed by gel electrophoresis, blots were blocked in buffers and probed with either highly cross-reactive 4G2 (obtained from Dr. Michel Ledizet, L2 Diagnostics; under non-reducing conditions) or CD9 (Invitrogen/ThermoScientific; under non-reducing conditions) or monoclonal anti-Langat virus NS1 (Clone 6E11; BEI Resources) antibodies, followed by mouse monoclonal HRP-conjugated secondary antibodies (Santa Cruz Technologies, Inc). Total protein profiles (images obtained from stain free gels after running or imaged from Coomassie stained gels) serve as loading controls. For protease-resistance assay using proteinase K (that generally digest proteins in biological samples), we used typical working concentrations of 50–100 μg/ml (for tick cells; 50 μg/ml) or much above the concentrations (for N2a cells; 100 μg/μl). Briefly, we isolated fresh exosomes (ultracentrifugation methods) from tick (2 x 106) or N2a cells (2 x 107), and treated with Proteinase K for 15 min at 37°C. Samples were then heat-inactivated at 60°C for 10 min and loaded on SDS-PAGE gels and processed for immunoblotting with 4G2 antibody followed by relevant secondary antibody. Antibody binding was detected with WesternBright ECL kit (Advansta, BioExpress). Blots were imaged using Chemidoc MP imaging system and processed using Image Lab software from the manufacturer (BioRad).
For the native-PAGE analysis, we seeded ISE6 tick cells (2 x 106) in regular L15 medium for overnight and infected with LGTV (MOI 1). For N2a cells (5 x 106), we plated them in regular complete DMEM medium and allowed them to adhere for overnight, cells were then replaced with exosome-free FBS medium. After 4 h of media change, N2a cells were infected with LGTV (6 MOI). Post 72 h of infection, tick/N2a cell culture supernatants were processed for isolation of exosomes. Exosomes collected from uninfected or LGTV-infected tick/N2a cells were resuspended in PBS and distributed into three groups (from the same preparations), that were either held as untreated group on ice, treated with Triton-X-100 (0.03%; 30 min, RT), or processed for three cycles of freezing at -80°C (for each freezing cycle samples were incubated for 1 h). After treatment and processing, protein lysates were prepared in a non-reducing and non-denaturating sample buffer (62.5 mM Tris-HCL, pH 6.8, 25% Glycerol and 1% Bromophenol blue), that maintained the proteins secondary structure and native charge density. Gels were pre-run for 60 min in gel running buffer (25 mM Tris and 192 mM Glycine). Uninfected or LGTV-infected exosomal preparations with different treatment or untreated samples were separated on 12% native-PAGE gels. Gels were transferred on to nitrocellulose membranes followed by immunoblotting using 4G2 or NS1 monoclonal antibodies followed by mouse monoclonal HRP-conjugated secondary antibodies (Santa Cruz Technologies, Inc). Total protein profiles (Coomassie blue stained gel) serve as loading controls. Antibody binding was detected with WesternBright ECL kit (Advansta, BioExpress). Blots were scanned using Chemidoc MP imaging system and instructions from the manufacturer (BioRad).
We collected N2a cell-derived exosomes from 5 x 106 uninfected or LGTV-infected (MOI 6; 72 h p.i.) cells and resuspended in PBS (250 μl/sample). Exosomal fractions were grouped as untreated or treated with 0.1% of Triton-X-100 for 30 min. Nunc grade ELISA plates were coated with 50 μl of untreated or Triton-X-100 treated- uninfected or infected samples for overnight and incubated at 4°C. Samples were incubated with 4G2 antibody for 1 h, followed by HRP-conjugated mouse monoclonal secondary antibody for another 1 h as described [72]. We used SureBlue TMB Microwell Peroxidase substrate and Stop solution (KPL) and followed manufacturer’s instructions. After stopping the reactions with TMB Stop solution, optical density was measured from triplicate samples at an absorbance of 450nm using a Multimode infinite M200 Pro Microplate reader (Tecan).
LGTV-infected tick or N2a cell-derived exosomes (from 72 h p.i.) were freshly isolated from 2 x 106 tick cells (infected with MOI 1) or 5 x 105 N2a cells (infected with 6 MOI). We also isolated exosomes from GW4869 inhibitor (5 μM) treated tick or N2a cells. For inhibitor treatment, cells were seeded in plates for overnight, changed to exosome-free FBS medium (in case of N2a cells) and after 4 h, treated with exosome release GW4869 inhibitor for 4h, followed by infection with LGTV for 72 h p.i. The exosomes collected from untreated or inhibitor treated cells were resuspended in PBS and grouped into three categories for both inhibitor treated or untreated samples as; untreated, treated with 4G2 antibody (that recognizes LGTV E-protein) or relevant isotype control antibody (R & D Systems) groups. Exosomal fractions were incubated for 1 h (RT) with respective antibodies followed by incubation (4°C) with protein A/G agarose beads (Pierce/ThermoScientific) for another 30 min. The antibody-beads complexes were spun (13k rpm) at 4°C for 30 min and supernatants were collected and lysed in RNA lysis buffer, processed for RNA extractions, followed by cDNA synthesis and QRT-PCR to detect LGTV loads.
Assays were performed to analyze the trans-migration of infectious exosomes from infected cells (seeded in inserts; upper chamber) to uninfected cells seeded in 12-well plates (lower chamber). Sterile, polycarbonate tissue culture-treated transwell inserts (12mm insert size) with 0.4 μm microporous membrane pore size were used in our assays (Corning). We plated, 1 x 105 ISE6 tick or bEnd.3 cells in inserts (upper chamber) and 1 x 105 HaCaT or N2a cells were seeded in 12-well plates (lower chamber). Inserts with tick or bEnd.3 cells were first kept in a separate 12-well plates containing 0.5 ml (in order to keep microporous membranes moist/wet) of L-15 (tick cells) or DMEM complete medium (bEnd.3 cells), respectively. Inhibitor-treated group in transwell assays was treated with 5 μM of GW4869 inhibitor, and at 24 h post treatment, tick cells or bEnd.3 cells were either infected with exosomes containing LGTV (25 μl of the exosome fraction collected from infected tick or bEnd.3 cells) or with LGTV from laboratory viral stocks (MOI 1 for tick cells or MOI 6 for bEnd.3 cells) prepared from infected Vero cell culture supernatants. Four hours post-infection, inserts with tick or bEnd.3 cells (with change of new media) were moved to 12-well plates containing HaCaT or N2a cells, respectively. Exosomes containing viral RNA and proteins produced from tick or bEnd.3 cells were allowed to transmigrate and infect HaCaT or N2a cells (that were kept uninfected). After, 48 h post incubation with inserts (containing either infected tick or bEnd.3 cells in inserts or upper chambers), HaCaT or N2a cells from lower chamber were washed with ice-cold PBS (3x) and collected for RNA extractions, cDNA synthesis and QRT-PCR to determine viral loads from cells.
Plaque assays were performed as described [72]. To determine infectious and replicative viruses after incubation with exosome and exosome-free supernatant fractions, we seeded Vero cells in 6-well plates at densities of 1 x 106 cells per well, allowed them to adhere and grow as monolayers to reach 65–85% confluency (for ~24 h). Exosome fractions containing unknown PFU (plaque forming units) of LGTV viral genomes were collected from tick cells (5 x 106 cells) or N2a cells or murine cortical neurons (1 x 105 cells) and resuspended in 250 μl of PBS, 30 μl of this suspension (exosome fraction) was used for plaque assays. Exosome free supernatants (600 μl) that corresponds to the same ratio of exosome fractions were used as controls. Serial dilutions (1:10, 1:100 and 1:1000) of the exosomes or supernatant fractions were prepared in duplicate (shown are the representative plate images from two-three independent experiments). Monolayers of Vero cells were infected with exosomes or supernatant fractions or with exosomes from 1 μM or 10 μM (LGTV-infected N2a cells or mouse cortical neurons) of GW4869 inhibitor-treated or DMSO-treated controls. Four hours post infection, medium was removed and warm 2% Seaplaque agarose (Lonza) overlay with complete DMEM media (1:1 ratio) containing antibiotic and antimycotics solution (1% each; Sigma) was added. Plates were incubated for 6–7 days, at 37 °C, 5% CO2. After incubation period, plaques were stained with 0.03% of Neutral Red (Sigma) for 4 h, and the stain was removed to either count plaques on the same day or otherwise plates were incubated (inverted and covered in foil) for overnight, and then plaques were counted next day to determine the viral titers from LGTV-infected exosomal fractions from tick/neuronal cells.
Gestation period (day13) wild-type female C57BL/6 (Charles River Laboratories) mice were purchased and allowed to reacclimatize. All animal experiments were done in accordance with the University Animal Care and Use committee regulations. Primary cortical neurons were isolated from embryonic day-16 (E16) brains [72,73]. Murine cortical neurons (1 x 105) were seeded in a 12-well plate coated with poly-L-Lysine and cultures were established in neurobasal complete medium with FBS. After 24h of plating, half of the medium was replaced with FBS-free neurobasal media, to slow growth of glial cells. For infection kinetics, cortical neurons were infected with LGTV (MOI 4) (after 48 h of post-seeding), neurons were collected at different time points (24, 48, 72 and 96 h p.i.) and processed for RNA extractions. For infection with neuronal cell-derived exosomes or supernatant fractions, 1 x 105 murine cortical neurons were infected with 20 μl of neuronal exosomes (2.2 x 103 pfu/ml) or 400 μl of exosome free supernatant fractions (collected from the step before PBS wash during exosome isolation, See S2 Fig). Cells were harvested at 48 h p.i. and processed for RNA extraction. Protein extractions were collected from uninfected or LGTV-infected cortical neurons (seeded at 1 x 107 cells) or from exosomes isolated from these cell culture supernatants.
For exosome inhibition studies, we used GW4869 a cell permeable, selective inhibitor for Neutral Sphingomyelinase (N-SMase) (Santa Cruz Biotechnologies, Inc) and DMSO as controls. Cells did not show any toxicity at tested doses. For both transwell assays, inhibitor-treated group was treated with 5 μM exosome inhibitor. N2a cells or murine cortical neurons were seeded at 1 x 105 cells in a 12-well plate. Next day, before treatment with inhibitor, cells were replaced with bovine exosome free-FBS (Systems BioSciences, Inc.) containing DMEM (N2a cells) or neurobasal medium (neurons). Cells were treated with either 1, 5 or 10 μM (N2a cells) or with 1, 10 or 20 μM (neurons) of inhibitor for 4 h, followed by infection with LGTV (N2a cells, MOI 6; cortical neurons MOI 4). Plaque assays were performed with 30 μl of exosome fractions derived from N2a cells or cortical neurons to determine the unknown titers for both DMSO control or inhibitor treated groups, respectively. N2a cells were either pre- or post- treated with inhibitor, where cells were first treated with inhibitor (5 μM) for 4 h followed by infection for 72 h or vice versa, respectively. Supernatants collected from uninfected controls and cells infected with LGTV (laboratory virus stocks with known titers) (48 h p.i.) were processed for exosome isolation. Purified exosomes were resuspended in PBS and processed for either RNA extraction or used for infection of new cells to determine re-infection kinetics or used to determine viral titers by plaque assays. For GW4869 treatment on laboratory virus stocks, we treated the viral supernatants (collected from Vero cells) with known titers (7.4 x 106 pfu/ml). We used 30 μl of the virus stocks and treated with either DMSO or inhibitor (5 and 10 μM for 4 h at 37°C) followed by immunoblotting with 4G2 antibody. For 4G2 functional blocking antibody studies, we plated N2a cells (2 x 105), treated with 5 μg of antibody for 4 h and infected cells with freshly isolated exosomes from LGTV-infected (72 h p.i.) N2a cells. N2a cells were infected through infectious exosomes for 72 h p.i. and collected for RNA extractions and QRT-PCR analysis. Untreated samples serve as control. For Pitstop-2 inhibitor treatment, N2a cells (2 x 105) were treated with 30 μM Pitstop-2 (dissolved in DMSO) for 15 min followed by infection through freshly isolated exosomes from LGTV-infected (72 h p.i.) N2a cells. Cells were collected for RNA extractions after 72 h p. i. and further processed for QRT-PCR. DMSO treated cells served as controls.
Statistical difference observed in data sets was analyzed using GraphPad Prism6 software and Microsoft Excel. The non-paired, two-tail Student t test was performed (for data to compare two means) for the entire analysis. Error bars represent mean (+SD) values, P values of <0.05 were considered significant in all analysis. Statistical test and P values are indicated for significance.
All animal work in 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 Institute of Health. The approved protocol from the Institutional Animal Care and Use Committee (Animal Welfare Assurance Number: A3172-01) was used in this study (permit number: 16–017).
|
10.1371/journal.ppat.1005739 | PET CT Identifies Reactivation Risk in Cynomolgus Macaques with Latent M. tuberculosis | Mycobacterium tuberculosis infection presents across a spectrum in humans, from latent infection to active tuberculosis. Among those with latent tuberculosis, it is now recognized that there is also a spectrum of infection and this likely contributes to the variable risk of reactivation tuberculosis. Here, functional imaging with 18F-fluorodeoxygluose positron emission tomography and computed tomography (PET CT) of cynomolgus macaques with latent M. tuberculosis infection was used to characterize the features of reactivation after tumor necrosis factor (TNF) neutralization and determine which imaging characteristics before TNF neutralization distinguish reactivation risk. PET CT was performed on latently infected macaques (n = 26) before and during the course of TNF neutralization and a separate set of latently infected controls (n = 25). Reactivation occurred in 50% of the latently infected animals receiving TNF neutralizing antibody defined as development of at least one new granuloma in adjacent or distant locations including extrapulmonary sites. Increased lung inflammation measured by PET and the presence of extrapulmonary involvement before TNF neutralization predicted reactivation with 92% sensitivity and specificity. To define the biologic features associated with risk of reactivation, we used these PET CT parameters to identify latently infected animals at high risk for reactivation. High risk animals had higher cumulative lung bacterial burden and higher maximum lesional bacterial burdens, and more T cells producing IL-2, IL-10 and IL-17 in lung granulomas as compared to low risk macaques. In total, these data support that risk of reactivation is associated with lung inflammation and higher bacterial burden in macaques with latent Mtb infection.
| Asymptomatic infection with Mycobacterium tuberculosis, often called latent tuberculosis, affects more than 2 billion people. Reactivation of latent infection to active TB occurs in only a minority of those infected, yet can lead to deadly disease and transmission. Here we show, using a non-human primate model, that imaging using PET/CT can identify certain features that are associated with a higher risk of reactivation. These factors include overall lung inflammation, individual granulomas in the lung with higher bacterial burden, and a site of infection outside the lungs. Using these parameters may allow discovery of peripheral biomarkers regarding risk of reactivation from latent TB. Such biomarkers could identify those people who would benefit most from treatment of latent TB to prevent reactivation.
| The vast majority of people infected with Mycobacterium tuberculosis (Mtb) develop asymptomatic, latent infection (LTBI). It is increasingly recognized that there is a spectrum of LTBI in humans, and this spectrum may correlate with the risk of reactivation [1]. Although reactivation risk is estimated at 10% per lifetime in HIV-negative LTBI humans, this is a population level estimate. Instead, it seems more likely that a small percentage of those with LTBI are at higher risk of reactivation. However, it has been challenging to identify the small fraction of the more than 2 billion latently infected humans who are at greatest risk of reactivation, so that therapy can be targeted to that population.
As in humans, LTBI in macaques is a stable, asymptomatic infection without clinical signs [2]. Reactivation of LTBI can be triggered in macaques by immune suppression due to SIV infection, TNF neutralization and CD4 depletion [3–6], but variable rates of reactivation are observed, similar to humans. We hypothesize that the spectrum of LTBI is associated with susceptibility to reactivation [1, 2]. Here we develop criteria based on 18F-fluorodeoxyglucose (FDG) positron emission tomography coupled with computed tomography (PET CT) imaging of macaques with LTBI to predict reactivation risk due to TNF neutralization. These criteria were then applied to latently infected macaques (without TNF neutralization) to identify biologic features that correlate with higher risk of reactivation. Macaques at high reactivation risk had greater cumulative lung bacterial burden, higher bacterial burden within an individual granuloma, more Mtb-infected mediastinal lymph nodes, and more T cells producing IL-2, IL-10 and IL-17 in lung granulomas compared to low risk macaques. Our results support the model of a spectrum of latency, suggesting that the extent and quality of bacterial control as well as lung inflammation in latency determines risk of reactivation after TNF neutralization.
We have previously published criteria for determining whether cynomolgus macaques with M. tuberculosis infection are “active” or “latent” by 6 months post-infection, based on clinical and microbiologic tests, as in humans [2, 7]. These clinical classifications were confirmed at necropsy, where those classified as active TB had significantly more pathology and bacterial burden than those classified as latent [2]. In this study, our aim was to determine whether we could identify latently infected macaques that would be more susceptible to reactivation. To do this, we employed serial FDG PET CT imaging, prior to and during neutralization of TNF, which we have shown previously can induce reactivation in macaques [5]. A cohort of cynomolgus macaques with LTBI (n = 26) was PET CT imaged at least 6 months post-infection, immediately prior to being randomly assigned to receive either TNF neutralizing antibody for 5–8 weeks or no treatment. Each macaque was evaluated for reactivation which was strictly defined here as dissemination, determined by the appearance of at least one new granuloma in lungs or extrapulmonary sites by PET CT during anti-TNF antibody treatment (Fig 1). Of 26 animals with TNF neutralization, 50% (n = 13) developed new lesions. At necropsy, macaques that developed new granulomas during TNF neutralization had greater disease pathology and higher total bacterial burden in lungs (Fig 2A and 2B) as well as within individual granulomas and lymph nodes (S1 Fig). Animals that developed reactivation had a significantly smaller proportion of sterile (or greater proportion with Mtb growth) among granulomas and mediastinal lymph nodes compared to animals that did not reactivate (Fig 2C and 2D). Thus, these data support the use of dissemination, the formation of new lesions in lungs or extrapulmonary sites, as a primary metric of reactivation.
While we defined reactivation in terms of bacterial dissemination, we postulated that we might also see evidence of loss of bacterial control among pre-existing lesions from the animals that reactivated. We examined lesion specific changes in metabolic activity (FDG avidity of each granuloma by PET, reported as standard uptake value, SUV) and/or size (by CT) during TNF neutralization. Granulomas were classified as “stable” if they remained similar in SUV (change < 5 units) or size (change < 1mm) or “dynamic” if they increased in SUV (≥ 5 units) or size (≥1mm) (S2 Fig). At least 1 dynamic lesion was observed during TNF neutralization in 69% (9 of 13) of reactivated monkeys compared to only 31% (4 of 13) among non-reactivated animals. Dynamic lesions were less likely to be sterile and had significantly higher bacterial burdens (measured as colony forming units, CFU) compared to stable lesions (Fig 2E) among all TNF-neutralized animals. The frequency of sterile lesions among new, stable and dynamic lesions was statistically different (Fig 2E) with the lower proportion of sterile lesions among the dynamic and newly developed granulomas. The total number of granulomas per monkey among reactivated (median = 12, IQR25 = 8,IQR75 = 24.5) and non-reactivated (median = 8, IQR25 = 4.5, IQR75 = 19.5) animals was similar (Mann-Whitney, p = 0.3). These data suggest that the increased bacterial burden observed in reactivation is not solely driven by the number of new lung granulomas but likely a combination of granuloma types and MLN burden.
We also compared the ratio of live Mtb CFU to chromosomal equivalents (CEQ) (the cumulative burden of live and dead Mtb) to estimate bacterial killing [8] in dynamic and stable lesions. Dynamic granulomas had higher CFU/CEQ ratios (i.e., less bacterial killing) than stable granulomas among all animals undergoing TNF neutralization (Fig 2F). Granulomas from reactivated animals had higher CFU/CEQ ratios compared to non-reactivated animals and LTBI controls (that did not receive TNF antibody) (S3 Fig). Importantly, however, many lesions in reactivated animals did not increase in metabolic activity or size or display reduced killing after TNF neutralization. This supports our previous data that demonstrates marked heterogeneity of lesions within an individual animal [8–11].
We next sought to identify PET CT characteristics of macaques prior to TNF neutralization that are predictive of reactivation risk. The number of lung granulomas observed before TNF neutralization was similar between animals that reactivated and those that did not (Fig 3A). We have previously shown that overall lung inflammation (total lung FDG avidity) detected by PET is loosely associated with lung bacterial burden in macaques with active TB, and decreased dramatically with anti-TB drug treatment [10, 12]. In this study, total lung FDG avidity immediately prior to anti-TNF treatment was significantly higher in animals that would later reactivate (Fig 3B) compared to those that did not. We then sought to define the distinguishing characteristics of individual lung granulomas prior to anti-TNF treatment that correlated with reactivation risk. Animals that would develop reactivation had a higher proportion of FDG avid (defined as SUV ≥ 5) granulomas (67.2%) compared to non-reactivator animals (30.8%) (Fisher’s Exact, p<0.0001) before TNF neutralization. Comparing the single most FDG avid or the largest size granuloma in each animal prior to anti-TNF treatment showed that those that would develop reactivation had a granuloma that was larger or had higher FDG avidity (Fig 3C and 3D). Although dynamic lesions had higher FDG avidity (median SUV 8.8, IQR 25–75: 5.0, 13.9) compared to stable lesions (median SUV 4.5, IQR25-75: 2.9, 7.5, Mann-Whitney, p<0.0001) before TNF neutralization, we were unable to predict which lesions would become dynamic as only 23.7% of the lesions with greatest SUV and 31.6% of lesions of maximum size were dynamic lesions after TNF neutralization.
FDG avid mediastinal lymph nodes (MLN) seen on PET CT are usually associated with Mtb involvement. Previous data from our lab and from human studies suggest that infected MLN are a potential source of reactivation [5, 6, 13]. The cumulative FDG avidity of hot MLNs was significantly higher among animals that would later reactivate (reactivators cumulative SUV median = 9.6, IQR25-75: 5.8, 31.5 vs. non-reactivators cumulative SUV median = 5.5, IQR25-75: 0, 13.9, Mann-Whitney, p = 0.04). Lastly, the number of extrapulmonary sites of infection before TNF neutralization was significantly higher in animals that would later develop reactivation compared to those that would not (Fig 3E).
To assess which PET CT variables could distinguish reactivators from non-reactivators before TNF neutralization, we ran several different simple logistic regression models (and a contingency analysis) to narrow down the best predictors using the following variables: total lung FDG activity, number of “hot” lymph nodes, total number of lymph nodes, total SUV of lymph nodes, number of lesions, and the presence of extrapulmonary lesions. We chose total lung FDG activity and the presence of extrapulmonary lesions as best predictors based on goodness of fit. We then used recursive partitioning, splitting the data into a decision tree to define an optimal cut-off (947.2 SUV) for total lung FDG activity. Combining the presence of extrapulmonary involvement and total lung FDG avidity resulted in favorable receiver operator curve results (area under the curve = 0.94) and a high sensitivity (92.3%) and specificity (92.3%) (Fig 3F).
It was not feasible to administer TNF neutralizing antibody to another large set of latently infected macaques to validate our predictions or identify biologic features associated with reactivation risk. Therefore we leveraged a set of latent control macaques (N = 25) that were necropsied concurrently with the TNF neutralized group, categorizing them being as at high or low risk for reactivation based on our PET CT defined parameters (i.e., total lung FDG avidity and extrapulmonary involvement) using the scan prior to necropsy. Animals who had evidence of extrapulmonary involvement on scan or a total lung FDG avidity of greater than or equal to 947.2 were classified as high risk. We then analyzed lesions from these animals to investigate the bacterial and immunological factors associated with risk of reactivation. While the median CFU per granuloma was the same between high and low risk animals (S4 Fig), a wide range of CFU per granuloma was observed within each individual monkey (S5 Fig). Because low risk animals appeared to have a lower peak bacterial burden compared to high risk, we then compared the maximum CFU per single granuloma within an in individual monkey to limit the bias. Interestingly, the maximum CFU per single granuloma within an animal was also higher among high risk compared to low risk animals (Fig 4A, S6 Fig). The total lung CFU (cumulative CFU of all lesions in the lung) was also greater among high-risk LTBI animals (Fig 4B, S6B Fig). High-risk LTBI animals also had a trend toward greater CFU per MLN (Fig 4C), which was surprising given the large range of CFU per MLN on an individual animal level (S7 Fig). High risk animals also had a smaller proportion of sterile MLN (i.e., greater proportion of MLN with Mtb growth) compared to low risk animals (Fig 4D). Together these data indicate that individual lesional characteristics (i.e., high bacterial burden in one granuloma or lymph node) are associated with high risk of reactivation.
We then examined Mtb-specific T cell cytokine production within individual lung granulomas and blood of high- (n = 10) and low-risk (n = 10) LTBI control animals (without TNF neutralization). Granulomas from high-risk animals had higher frequencies of IL-17, IL-10 or IL-2 producing CD3+ T cells as compared to granulomas from low risk animals (Fig 5, S9 Fig). While most T cells from granulomas were single cytokine producers, as we previously described [9], we found that high-risk animals had a higher percentage of granuloma T cells producing more than one cytokine. There were no differences in frequencies of Mtb specific cytokine production by CD4+ and CD8+ T cells or by memory subsets in peripheral blood (S9, S10 and S11 Figs). These data suggest that granulomas from high-risk LTBI animals are more immune stimulated, possibly due to higher bacterial activity, although the specific factors driving this are unknown.
In this study, we strictly defined reactivation following TNF neutralization based on dissemination (formation of at least one new granuloma) and validated that definition at necropsy, where reactivated macaques had higher bacterial burden than those that did not reactivate. Together the data in this study support the hypothesis that the spectrum of latency has implications for risk of reactivation. Here we provide evidence that lung inflammation and/or evidence of extrapulmonary involvement detected by PET CT is associated with reactivation risk following TNF neutralization in macaques. In addition, reactivation risk is correlated with at least one granuloma of larger size or higher inflammation (measured by PET).
This unique set of data provided the opportunity to investigate individual granulomas in macaques predicted to be at high or low risk of reactivation. Using the parameters we developed based on PET CT characteristics of LTBI macaques that did or did not reactivate following TNF neutralization, we predicted the risk of reactivation of 25 latent control macaques. Neutralizing TNF would change the bacterial and immunologic features in these animals, which would prevent investigation of these factors in risk of reactivation. Therefore, instead of testing our prediction by treating the animals with anti-TNF treatment, we compared several factors in our predicted high or low risk animals without further intervention. Animals predicted at high risk had higher total lung and lymph node bacterial burden. In addition, an individual granuloma in high risk animals had a high bacterial burden, suggesting that a single poorly contained granuloma can contribute to reactivation. T cell cytokine production in granulomas was higher in high risk compared to low risk macaques in the absence of TNF neutralization. This could be due to more bacterial replication and antigen production in high risk animals. Alternatively, the combination of cytokines in certain granulomas from high risk animals may provide a less stable host immune environment. Further studies of immune responses in granulomas from low or high risk animals is necessary to differentiate cause and effect of cytokine responses in predisposing animals to reactivation. The data from this study suggests that only one or a few granulomas need to fail in bacterial containment to lead to dissemination and reactivation.
Not all granulomas were equally affected by TNF neutralization, suggesting that reactivation and dissemination can occur from as few as one unstable granuloma. For example, TNF neutralization led to dynamic granulomas (increasing in size or FDG avidity), but this was restricted to a subset of granulomas in reactivating monkeys. This is consistent with the independent and dynamic nature of lung granulomas in this model [8–11]. However, given the lack of current technology for tracking individual bacilli, it is not possible to confirm that dynamic lesions, or those with higher bacterial burden, are the source of dissemination. A limitation of this study is that we are unable at this time to identify direct causes of increased bacterial burden or instability of lesions, which are both associated with reactivation.
The ability of mediastinal lymph nodes (MLN) to control infection during clinical latency also appears to contribute to risk of reactivation. In LTBI, a Ghon complex refers to the combination of a lung granuloma and an involved lymph node, suggesting lymph node involvement is common in humans. Studies from the pre-antibiotic era also demonstrate substantial lymph node involvement in humans soon after infection [14]. Similarly, mediastinal lymph nodes have been detected by PET CT in humans with Mtb infection [15–18]. Thus, the role of lymph nodes in susceptibility to reactivation is likely important [19] and should be targeted in the development of drugs to treat latent infection. We previously published an association between extent of CD4 depletion in MLN and reactivation in latently infected macaques treated with CD4-depleting antibody [6]. In this current study, the latent controls predicted to be at high risk of reactivation a greater proportion of MLNs with Mtb growth compared to those that were at low risk. However, in a previous study, macaques vaccinated with BCG plus the protein fusion H56 vaccine were protected against TNF-neutralization induced reactivation [20]. Examination of our data from that study shows that protection against reactivation was associated with significantly fewer Mtb positive MLN (S12 Fig). Thus, it is likely that the MLN play an important role in reactivation risk. Even less is known about the presence of extrapulmonary sites of infection during LTBI. While it occurs in humans [21, 22], the actual prevalence has not been well described. We speculate that the events that result in extrapulmonary infection being established are due to poor initial control and early dissemination, which is then associated with reactivation risk.
In summary, we have provided evidence that the spectrum of clinically defined LTBI, specifically that associated with inflammation detected by PET CT and the presence of extrapulmonary disease, is associated with reactivation risk. It is likely that this occurs in humans with LTBI and similar lung lesions have been described in LTBI humans by PET CT [16, 23–25]. Importantly, this is the first assay that can functionally distinguish those at high and low risk for reactivation induced by TNF neutralization. We recognize that using PET CT to stratify reactivation risk is not feasible in most human settings. However, the use of PET CT in this well characterized animal model provides an opportunity to identify potential biomarkers in blood, including transcriptional signatures, which may correlate with reactivation risk. Prioritizing treatment to those patients at increased risk of reactivation (especially those with HIV infection) is a more efficient strategy in our current efforts to eradicate TB, as most programs are unable to provide treatment to all LTBI patients.
Adult (≥ 4 years of age) cynomolgus macaques (Macacca fasicularis) (Valley Biosystems, West Sacramento, CA) were screened with standard tests for co-morbidities prior to infection with Mtb as previously published[26]. Animals were maintained in a Biosafety Level 3 facility for primates after M. tuberculosis infection.
Cynomolgus macaques were infected with low dose (~25 CFU) of M. tuberculosis (Erdman strain) via bronchoscopic instillation into a lower lung lobe and subsequent serial clinical, microbiologic and immunologic parameters were followed until outcome was determined as previously described [2, 7]. Once animals were classified as latent, a subset was randomized to receive TNF neutralizing agent Adalimumb (Humira, Abbott Labs, Abbott Park, IL) at 4 mg/kg/dose subcutaneously every 7–10 days [5]. In general, PET-CT scans were performed at baseline before TNF neutralization and every 2 weeks after treatment until 5–8 weeks. A pre-necropsy scan was performed on all animals to facilitate harvesting scan-identified lesions. Serial analysis of these lesions before and during treatment was performed (see below). At necropsy, animals were maximally bled and gross pathology was assessed using our previously published quantitative scoring system in which a number is given for the size, number and pattern of granulomas in each lung lobe, mediastinal lymph node and extrapulmonary sites (e.g., liver, spleen). Harvested lung granulomas are characterized, measured and processed into single cell suspension for bacterial burden and flow cytometry as previously reported [2, 9].
All animal protocols and procedures were approved by the University of Pittsburgh’s Institutional Animal Care and Use Committee (protocol assurance number A3187-01.) Our specific protocol approval numbers for this project are 0808244, 0906877, 1011342,1105870 and 11080037. The IACUC adheres to national guidelines established in the Animal Welfare Act (7 U.S.C. Sections 2131–2159) and the Guide for the Care and Use of Laboratory Animals (8th Edition) as mandated by the U.S. Public Health Service Policy.
At predetermined time points, animals were sedated, intubated and imaged by 2-deoxy-2-18F-D-deoxyglucose (FDG) PET (microPET Focus 220 preclinical PET scanner, Siemens Molecular Solutions, Knoxville, TN) and CT (Neurologica Corp, Danvers, MA) imaging within our biosafety level 3 facility as previously described [10, 12, 27]. Lesions were identified by two analysts (M.T.C. and P.M.) and size was measured by CT. FDG avidity was measured by drawing a region of interest (ROI) in the axial view and SUVs (standard uptake volume, SUV = counts/(injected activity/body weight), normalized to muscle to reduce variability between scans, were calculated using OsiriX (Pixmeo, Geneva, Switzerland) as previously published [27].
The total lung FDG avidity was analyzed and calculated using Osirix viewer, an open-source PACS workstation and DICOM viewer. The whole lung was segmented on CT by using the Growing region algorithm on the OsiriX viewer to create a ROI of normal lung (Hounsfield units between -1024 and -200). The closing tool was used to include individual nodules and other pulmonary disease. The ROI was transferred to the co-registered PET scan and manually edited to ensure all pulmonary disease was included. All extrapulmonary structures and disease, including mediastinal lymph nodes, were excluded. Voxels outside the ROI were set to zero and voxels within the ROI with an SUV higher than normal lung (SUV ≥ 2.3) were isolated. These ROIs (capturing all SUV ≥ 2.3 within the lung) were exported into a spreadsheet using the OsiriX “Export ROI” plugin. Finally, the sum from the pixels of each slice (from the exported ROI) was calculated to represent the measurement of “Total Lung FDG Avidity”.
Granuloma specific changes observed before and during TNF neutralization were performed. “Stable” lesions were defined as lesions that maintained the same size and FDG avidity before and after TNF neutralization whereas “dynamic” lesions were those that that increased in size by at least 1mm or FDG avidity by at least 5 SUV. “New” lesions were not present baseline but appeared during the course of TNF neutralization.
Following necropsy, tissue sample homogenates were stored in PBS at -80°C. At processing, M. tuberculosis genomes were extracted with phenol-chloroform as previously described [8]. In brief, blinded samples were re-suspended in 1 mL Tris-EDTA buffer, pH 8.0, with 300 μl of 70°C UltraPure phenol:chloroform:isoamyl alcohol (25:24:1) (Invitrogen) and 250 μl of 0.1 mm zirconia-silica beads (BioSpec Products, Inc.). Tubes were mixed by inversion, incubated for 10 min, and twice vortexed for 4 min with a 1 min break at highest speed using a 24-tube vortex adaptor (MO BIO Laboratories, Inc.). Following a 10 min centrifugation at 14,000 RPM at 4°C, the aqueous layer of each sample was extracted and placed in a fresh tube with 50 μl of 5M sodium chloride. The phenol:chloroform:isoamyl alcohol extraction was repeated with 250 μl and a 30 min incubation at room temperature. After the incubation, samples were centrifuged once more at 14,000 RPM for 30 min at 4°C to separate off the aqueous phase. One volume of isopropanol and 1/10 volume of 3M sodium acetate (pH 5.2) was then added to each extraction to precipitate genomic DNA with an overnight incubation at -20°C. Each DNA pellet was washed with 70% ethanol and centrifuged for 30 min as before. Each pellet was then left to air-dry to remove excess ethanol and subsequently re-suspended in sterile nuclease free water (Ambion). DNA purity and concentration was measured using the Spectramax 190 spectrophotometer (Molecular Devices).
Quantification of chromosomal equivalents (CEQ) of M. tuberculosis was performed using real-time PCR of a single copy gene, Mtb sigF, with a previously described primer-probe combination [8]. The primer and probes for this target were purchased together in a pre-mixed PrimeTime qPCR assay (Integrated DNA Technologies). The sequences are as follows, 5’3’: probe–FAM-TCG GAC TTC GTC TCC TTC-Iowa Black, sigF Fwd–GCG GGT CGG GCT GGT CAA C, and sigF Rvs–CCT CGC CCA TGA TGG TAG GAA C. Real-time PCR was preformed in duplicate on the iQ5 Multicolor Real-Time PCR Detection System (Bio-Rad Laboratories, Inc.) and the 384well-capable 7900HT Fast Real Time PCR System (Applied Biosystems) with TaqMan Universal Master Mix II (Life Technologies). Precise determination of CEQ was derived from a standard curve of serially diluted M. tuberculosis genomic DNA prepared from liquid culture for each qPCR run. Real time PCR efficiency for each run was maintained between 90% and 110%.
While the quantification of both live Mtb and chromosomal equivalents are estimates, the ratios reflect a relative estimate of bacterial killing as published [8]. Accuracy of these estimates both in the detection of live bacteria and chromosomal equivalents may be limited by potential clumping of the bacteria resulting in underestimates of CFUs and minor variations in PCR amplification based on sample-specific differences in protein contamination or PCR inhibitors from blood.
Intracellular cytokine analyses were performed on individual granulomas harvested at necropsy and on PBMC at predetermine time points (6 months post infection). As previously described [9] single cell suspension of individual granulomas or PBMC were stimulated ex vivo with peptide pools of Mtb specific antigens ESAT-6 and CFP-10 (10 μg/ml of every peptide) or controls in the presence of Brefeldin A (Golgiplug: BD biosciences) for 3.5 hours (for granulomas) or 6 hours (for PBMCs) at 37°C with 5% CO2. For PBMC, Brefeldin A was added after 1 hour stimulation with Mtb antigens or controls. Positive control included stimulation with phorbol dibutyrate (PDBu) and ionomycin and negative controls included a media only control and an isotype controls (only for intracellular cytokine markers) for all PBMC samples and for granulomas whenever additional cells were available. For flow cytometry, cells from granulomas were initially stained for viability marker (Invitrogen) followed by cell surface marker CD3 (clone SP34-2; BD Pharmingen) for T cells. Cell surface markers for PBMC T cells included CD4 (clone L200, BD Horizon) and CD8 (clone SK1: BD biosciences) and markers for T cell memory subsets included CD45RA (clone 5H9, BD biosciences) and CD27 (clone O323, eBioscience). Intracellular cytokine staining panel for both granulomas and PBMC included Th1 pro-inflammatory cytokines: IFN-γ (clone B27), IL-2 (clone: MQ1-17H12), TNF (clone: MAB11); Th17 cytokine: IL-17 (clone eBio64CAP17) and regulatory cytokine: IL-10 (clone JES3-9D7) as previously described [9]. Data acquisition was performed using an LSR II (BD) and analyzed using FlowJo Software v.9.7 (Treestar Inc, Ashland, OR). For all PBMC T cells, the non-specific T cell response from the negative control (media) was subtracted from the Mtb specific antigen stimulated responses. MIATA guidelines were followed for sample collection and staining procedure for PBMC samples. The gating strategies used for the evaluation of granulomas and PBMC are described in S8 and S9 Figs. Intracellular cytokine data from granulomas in latent control animals (Fig 5) was previously published [9] but not analyzed based on risk of reactivation. PBMC T cell cytokine data (S10 and S11 Figs) from a subset of the animals was used for computational modeling and has been published [28].
Single cell suspension of each harvested site (i.e., lung granulomas, complex pathologies, grossly normal lung, MLN, liver, spleen) was plated on 7H11 plates (minimum detection level of M. tuberculosis burden was estimated at <10 Colony Forming Units per granuloma) as previously described [2, 8]. Specific bacterial burden of each site (granuloma, complex pathologies, or MLN) was calculated as the product of the bacterial burden on a per gram basis and the total mass of the tissue site.
Total thoracic burden was calculated as the sum of all M. tuberculosis growth from the lung (includes all individual granulomas and more complex pathologies such as consolidations, TB pneumonia, coalescing granulomas and clusters) and mediastinal lymph nodes. Total lung burden was calculated as the sum of all M. tuberculosis growth from lung lesions without mediastinal lymph nodes. The bacterial burden data were then transformed by adding 1 and reported as CFU.
Normal distribution of the data was assessed for each continuous variable using the D’Agostino-Pearson Omnibus Test. For statistical comparison, pair-wise analysis of continuous data was performed by Student's T test for normally distributed data and Mann-Whitney test for nonparametric data. For analyses in which more than two groups were compared, Kruskall-Wallis test was performed with Dunn’s multiple comparisons as a post-hoc test of non-normally distributed data. Pair-wise analysis for matched data was performed using the Wilcoxon rank-sum test. P-values below 0.10 were reported specifically in figures and text.
To assess which PET CT variables could distinguish reactivators from non-reactivators before TNF neutralization, we ran several different simple logistic regression models (and a contingency analysis) to narrow down the best predictors using the following variables: total lung FDG activity, number of “hot” lymph nodes, total number of lymph nodes, total SUV of lymph nodes, number of lesions, and the presence of extrapulmonary lesions. We chose total lung FDG activity and the presence of extrapulmonary lesions as best predictors based on goodness of fit and then used recursive partitioning (a decision tree) to evaluate a cut-off for total lung FDG activity. A receiver operating characteristic (ROC) curve was plotted in order to graphically represent the sensitivity and specificity of the combination of these two predictors. Statistical analysis was performed using Prism 6.0 (Graphpad Software, Inc). The ROC curve was plotted in JMP Pro 10.2 (SAS Institute Inc.).
Given the variability in the number of granulomas per animals, methods were developed to minimize the potential for bias among animals that had a greater number of lesions (i.e., granulomas and MLN) for analysis. We derived a number of representative samples per monkey by first calculating the median number of samples (for which we had bacterial burden) per monkey per group so that one monkey could not contribute more than the median number of samples per monkey.
|
10.1371/journal.pntd.0006485 | Characterization of a Trichinella spiralis putative serine protease. Study of its potential as sero-diagnostic tool | Trichinellosis is a serious zoonositc parasitosis worldwide. Because its clinical manifestations aren’t specific, the diagnosis of trichinellosis is not easy to be made. Trichinella spiralis muscle larva (ML) excretory–secretory (ES) antigens are the most widely applied diagnostic antigens for human trichinellosis, but the major drawback of the ES antigens for assaying anti-Trichinella antibodies is the false negative in the early Trichinella infection period. The aim of this study was to characterize the T. spiralis putative serine protease (TsSP) and to investigate its potential use for diagnosis of trichinellosis.
The full-length TsSP sequence was cloned and expressed, and recombinant TsSP (rTsSP) was purified by Ni-NTA-Sefinose Column. On Western blotting analysis the rTsSP was recognized by T. spiralis-infected mouse serum, and the natural TsSP was identified in T. spiralis ML crude and ES antigens by using anti-rTsSP serum. Expression of TsSP was detected at various T. spiralis developmental stages (newborn larvae, muscle larvae, intestinal infective larvae and adult worms). Immunolocalization identified the TsSP principally in cuticles and stichosomes of the nematode. The sensitivity of rTsSP-ELISA and ES-ELISA was 98.11% (52/53) and 88.68% (47/53) respectively (P > 0.05) when the sera from trichinellosis patients were examined. However, while twenty-one serum samples of trichinellosis patients’ sera at 19 days post-infection (dpi) were tested, the sensitivity (95.24%) of rTsSP-ELISA was distinctly higher than 71.43% of ES-ELISA (P < 0.05). The specificity (99.53%) of rTsSP-ELISA was remarkably higher than 91.98% of ES-ELISA (P < 0.01). Only one out of 20 serum samples of cysticercosis patients cross-reacted with the rTsSP. Specific anti-Trichinella IgG in infected mice was first detected by rTsSP-ELISA as soon as 7 dpi and antibody positive rate reached 100% on 10 dpi, whereas the ES-ELISA did not permit detection of 100% of infected mice before 16 dpi.
The rTsSP is a potential early diagnostic antigen for human trichinellosis.
| Trichinellosis is an important parasitic zoonosis, and has a public health hazard and an economic impact on the safety of animal food. The diagnosis of trichinellosis is difficult and it is often misdiagnosed. There is an evident 2–3 week window stage between clinical manifestations and the anti-Trichinella IgG positive. Serine protease is a superfamily of proteolytic enzymes and exerts a major role in tissue invasion, larval development and survival of the parasites. A T. spiralis putative serine protease (TsSP) was characterized in ES proteins of T. spiralis intestinal infective larvae and adult worms by the immunoproteomics with early infection serum. In this study, the TsSP was expressed and purified. The results revealed that the TsSP was expressed at various T. spiralis stages (newborn larvae, muscle larvae, intestinal infective larvae and adult worms) and it was principally located in cuticle and stichosome of the nematode. The rTsSP was sensitive and specific for detection of anti-Trichinella IgG, and could be regarded as an early diagnostic marker of trichinellosis.
| Trichinellosis is an important food-borne parasitic disease worldwide. Trichinella infection occurs by ingesting raw or undercooked meat containing Trichinella muscle larvae [1]. T. spiralis is the main etiological agent of trichinellosis [2]. Outbreak of human trichinellosis was recorded in 55 countries around the world, and there were 65,818 cases and 42 deaths from trichinellosis reported from 41 countries during 1986–2009 [3]. Fifteen outbreaks of trichinellosis were documented in mainland China during 2004–2009 and pork is the dominating infection source [4,5]. A survey showed that the prevalence of porcine Trichinella infection in small pig farms in central China varied from 0.61% to 3.79% during 2010–2015, although the larval burdens in infected pigs was less than 2 larvae per gram of muscles [6,7]. Hence, trichinellosis has a public health hazard and an economic impact in meat food safety [8].
Since the symptoms and signs of trichinellosis aren’t specific, the diagnosis of trichinellosis isn’t easy to be established according to the clinical manifestations of this disease [9]. At present, the serological test widely applied for diagnosis of human trichinellosis is to detect anti-Trichinella IgG by ELISA and Western blotting with T. spiralis muscle larvae (ML) excretory/secretory (ES) antigens [10], but the principal drawback is the false negative in the early phase of this infection [11]. The occurrence of a 2–3 week window period of anti-Trichinella antibody negative is probable duo to the fact that the major ML ES antigen epitopes are the phase-specific for ML and not recognized by anti-Trichinella antibodies triggered by intestinal infective larvae (IIL) at 6 hours post infection (hpi) and adult worm (AW) at 3 dpi of the nematode in the early stage of Trichinella infection [12]. The ES antigens generated by the IIL and AW might firstly be exposed to host’s immune system and induced the generation of specific antibodies against the nematode. The recent investigation indicated that AW crude antigen positively reacted with swine and mouse infection sera at 7–8 dpi [13,14]. On Western blot analysis, the recombinant T. spiralis cystatin-like protein (rTsCLP) of IIL stage was probed by porcine infection sera at 15–20 dpi [15]. Anti-Trichinella IgG in serum samples of T. spiralis-infected mice was detected by ELISA using ES antigens of AW or IIL as soon as 8 dpi [16,17]. Therefore, it is likely that the diagnostic markers for early Trichinella infection will be exploited from the enteral worms of T. spiralis [18].
In our previous studies, immunoproteomics was used to investigate the early antigens for serodiagnosis of trichinellosis, and a putative serine protease was identified in the ES proteins from T. spiralis IIL and AW by mouse infection sera at 8–10 dpi and early trichinellosis patients’ sera at 19 dpi [12,19]. Additionally, the T. spiralis putative serine protease (TsSP) (GenBank accession no. ABY60762) was highly expressed in surface proteins of IIL stage compared with those of ML stage [20]. The aim of this study was to characteriz the TsSP and investigate the prospective diagnostic values of recombinant TsSP (rTsSP) for early trichinellosis.
The present study was performed in the light of National Guidelines for Experimental Animal Welfare (MOST of People’s Republic of China, 2006). All animal care and use in our research were reviewed and approved by the Life Science Ethics Committee of Zhengzhou University (No. SCXK 2015–0005). All the human serum samples were collected from adults, and the written informed consent was acquired from the adults before samples were used.
T. spiralis isolate (ISS534) utilized in our study was acquired from a naturally infected domestic swine in Henan Province of central China. This isolate was passaged in BALB/c mice in our department. Six-week-old female BALB/c mice were provided by the Experimental Animal Center of Zhengzhou University (Zhengzhou, China). Mice were kept with specific pathogen-free conditions under suitable temperature and humidity.
Fifty-three serum samples from trichinellosis patients were obtained from two outbreaks of human trichinellosis in southwestern China [17]. The sera from patients with paragonimiasis (n = 20), schistosomiasis (n = 34), clonorchiasis (n = 7), cysticercosis (n = 20), echinococcosis (n = 20) and sparganosis (n = 7) were conserved in our laboratory. The diagnosis of these patients was established by fecal parasitological examination or serum specific antibody detection [11,21]. The sera from 104 presumably healthy persons, who came from non-endemic areas of trichinellosis and assayed negative for the before-mentioned helminthiases, were also examined in our study.
In order to observe the dynamics of anti-Trichinella IgG, nine mice were infected orally with 300 T. spiralis ML. About 100 μl of tail blood was collected from infected mice on alternate days during 2–30 dpi and serums were isolated. Serum samples from normal mice were obtained and utilized as the negative control.
T. spiralis ML were obtained from experimentally infected mice at 35 dpi by using the artificial digestion method as described [22,23]. The IIL were recovered from intestines of the infected mice at 6 hpi [24], and the adult worms (AW) were separated from mouse duodenum and jejunum at 3 and 6 dpi, respectively [17]. The newborn larvae (NBL) were obtained from the adult females cultured in vitro in RPMI-1640 with 10% fetal bovine serum (FBS; Gibco) at 37°C in 5% CO2 for 24 h [25]. The crude soluble antigens of AW, NBL, ML and IIL, and the ML ES antigens were produced as described [26,27].
The complete TsSP cDNA sequence was acquired from the GenBank database with accession no. ABY60762. The Pepstats software was applied to predict molecular weight (MW), isoelectric point (pI) and transmembrane helices of the TsSP protein [28,29]. The putative N-glycosylation site was verified with the NetNGly1.0 server (http://www.cbs.dtu.dk/services/NetNGlyc/). The potential B and T cell epitope of the TsSP was calculated with the DNAStar software and the online server of BepiPred (http://www.cbs.dtu.dk/services/BepiPred/), respectively [30]. The tertiary structure of the TsSP protein was predicted on the Expasy website (http://web.expasy.org/). The identification of protein motifs and catalytic triad of the TsSP was from aligning the multiple protein sequences [31].
The total ML RNA was extracted with Trizol reagent (Invitrogen, USA). The full-length TsSP sequences were amplified via PCR with specific primers carrying enzyme BamHI and PstI sites (bold and italicized) (5'-GGGATCCATGATCCTTTTCAAGTGCTTATTTCT-3' and 5'-GCGCTGCAGTCAGCAAACTCAATTTATTTAGAT-3'). The TsSP gene coding regions without a 18 amino acid signal peptide were produced by PCR with oligonucleotide primers carrying enzyme BamHI and PstI sites (bold and italicized) (5'-TTCGGATCCAATTATGAA TGTGGCACCTTAC-3' and 5'-CCGCTGCAGTTAACGGAAAAAAGTGAATGAT-3'). PCR amplification reaction included 25μl premix (DNA polymerase, dNTPs and PCR buffer), 0.5 μl cDNA, 0.4μl DNA polymerase, 1.0 μl 10 μM of each primer, 22 μl ddH2O. The cycling procedure was as follows: 98°C for 5 min; 30 cycles of at 94°C for 3min, 94°C for 45 s,60°C for 45 s, 72°C for 90 s, and finally 5 min at 72°C. The final purified PCR product was digested and cloned into the pGEM-T vector (Promega, USA), then sub-cloned into the pQE-80L carrying the N-terminus His-tag (Novagen, USA). The recombinant pQE-80L/TsSP was transformed into Escherichia coli BL21 (DE3) (Novagen). The rTsSP expression was induced by using 0.5 mM IPTG for 4 h at 30°C. The rTsSP were purified with a Ni-NTA His-tag affinity kit (Novagen). The rTsSP protein were identified on SDS–PAGE analysis [32]. The concentration of the rTsSP protein was assayed as described [33].
The sequences of serine protease homologues from other organisms were aligned using the default settings in the program Clustal X [34]. The phylogenetic relationship among TsSP and other homologues was assayed by using a phylogenetic tree constructed in the MEGA 5.0 under the maximum parsimony algorithm with 1 000 bootstrap replications [35].
Thirteen BALB/c mice were immunized with rTsSP. Each mouse was injected abdomen subcutaneously with 20 μg of rTsSP emulsified in Freund’s complete adjuvant, then the mice were boosted twice with the same amount of rTsSP emulsified with Freund’s incomplete adjuvant at an intervals of 10 days [36]. About 50 μl of blood sample from immunized mice were obtained at 10 days after final immunization and serum anti-rTsSP antibody titer was assayed by ELISA with 2 μg/ml rTsSP as coating antigen [37].
Samples consisted of 5μg rTsSP, 15μg ML crude antigens and ML ES antigens per lane. The protein was separated on SDS-PAGE with 12% separation gel, subsequently transferred onto the membranes (Millipore, USA) at 18 V for 35 min in a semi-dry transfer cell (Bio-Rad, USA) [26]. The membrane was blocked by 5% skim milk in Tris–buffered saline with 0.05% Tween-20 (TBST) at room temperature for 2 h, and incubated with 1:100 dilutions of different sera (anti-rTsSP serum, serum of T. spiralis-infected mice collected at 42 dpi, immune serum from mice immunized with ML ES and crude antigens, and uninfected normal mouse serum) at 4°C overnight. Following being washed, the membrane was incubated with 1:10 000 dilutions of HRP-conjugated goat anti-mouse IgG at 37°C for 1 h. The membrane was colored by use of 3,3’-diaminobenzidine tetrahydrochloride (DAB; Sigma), and terminated by washing the membrane with deionized water.
To detect the relative TsSP expression level in T. spiralis different stages, 15 μg/lane of soluble proteins of ML, IIL, AW and NBL was separated with SDS-PAGE and identified by Western blotting with 1:100 dilutions of anti-rTsSP serum [38]. Rabbit anti-β-actin antibody diluted at 1:400 was utilized as a quantitative protein control to detect β-actin expression. After it was washed three times with TBST, the color development was performed by the enhanced chemiluminescence (ECL) kit (CWBIO, Beijing, China) [39]. The relative expression level of the TsSP protein at various T. spiralis phases was determined with Image J software.
Total RNA was extracted respectively from diverse T. spiralis phases (ML, IIL, AW, and NBL) with Trizol reagent (Invitrogen). The RT-PCR was carried out according to the previous report [38]. By using as an internal control, T. spiralis glyceraldehyde-3-phosphate dehydrogenase (GAPDH, GenBank accession No. AF452239) was amplified as a housekeeping gene in our study. PBS was used as a negative control template in all PCR assays.
The crude proteins from different T. spiralis phases (NBL, ML, IIL and AW) and ES proteins from AW, ML and IIL were prepared as described [27,40]. The above-mentioned antigens and rTsSP were diluted to a final concentration of 1.5 μg/ml. The ELISA procedure was performed as described previously [11]. Briefly, the microtiter plate was coated with the antigens at 4°C overnight. Following being washed with PBST, it was blocked with 5% skimmed milk in PBST at 37°C for 2 h. After washing again, the plate was incubated at 37°C for 1 h with 1:200 dilutions of trichinellosis patients’ serum or 1:100 dilutions of mouse serum, subsequently incubated with HRP-conjugated anti-human/mouse antibody IgG (1:10 000) at 37°C for 1h. After the last washing, the coloration was developed by incubation with o-phenylenediamine dihydrochloride (OPD; Sigma) plus 30% H2O2 for 30 min. The reaction was ceased by 2M H2SO4. The absorbance (optical density, OD) was measured at 490 nm, and all serum samples were assayed in duplicate. The ratio < 2.1 of assayed serum/negative serum OD values was taken as negative and the ratio ≥2.1 as positive [41]. The cut-off value of rTsSP-ELISA and ES-ELISA for detection of the patient’s serum was 0.35 and 0.45, respectively. The cut-off value of the above two ELISA for detection of experimentally infected mice was 0.20 and 0.21, respectively.
To confirm whether the TsSP expressed on the surface of T. spiralis diverse stages, the whole worms were used in IFT [42]. Additionally, the tissue sections with 3 μm thickness of female adults at 3 dpi, ML and IIL were separately cut by a microtome. The intact nematodes and their sections were blocked in 5% normal goat serum diluted with PBS, and incubated using a 1:10 dilution of mouse immune serum, infection serum or negative control serum. FITC-labeled goat anti-mouse IgG diluted at 1:100 (Santa Cruz, USA) was utilized as the second antibody. After they were washed with PBST, the intact nematode and sections were examined under a fluorescent microscopy (Olympus, Japan).
The statistical analysis of data was carried out by using SPSS 17.0 software. All the data were shown as arithmetic means ± standard deviation (SD). The comparison of the TsSP expression level in T. spiralis various stages was performed with one-way ANOVA. Chi-square test was used to determine the difference between groups. The statistical test was regarded significant at P < 0.05.
Bioinformatics analysis revealed that the full-length cDNA sequence of the TsSP gene was 1372 bp (CDS: 2–1290 bp). The predicted MW and pI of TsSP were 47.55 kDa and 8.73, respectively. The signal peptides were located at 1–18 aa (MILFKCLFLLAYTTLAFA). The mature serine protease consisted of 411 amino acid residues of 45.2 kDa, and no transmembrane helix was detected, indicating that the TsSP is a secretory protein. Only one N-glycosylation site 78–81 (NGSQ) of the TsSP was identified. Secondary structures of the TsSP had 18 potential B cell epitopes. The SMART analysis results demonstrated that the TsSP had a domain (at 37-277aa) of trypsin-like serine protease carrying an active site of classic catalytic triad. In three-dimensional model, the motif of catalytic triad (Serine–Histidine–Aspartate) constituted a functional domain carrying substrate binding sites (Fig 1).
A homology comparison of TsSP and other serine protease orthologues in the genus Trichinella was determined (Fig 2), among these sequences, the highest homology was between T. spiralis and T. nativa (with 90% identity). As shown in the phylogenetic tree generated with TsSP and its orthologues (Fig 3), the Trichinella genus was displayed as a monophyletic group with bootstrap value of 87. Within the Trichinella, the close relationships among T. spiralis, T. nativa and T. britovi were supported with a high bootstrap value (95).
The complete TsSP cDNA sequences without signal peptide were 1236 bp. The open reading frame (ORF) of TsSP encoded a 45.2 kDa protein of 411 amino acids. The TsSP coding sequences were cloned into the pQE-80L. Following induction, SDS-PAGE analysis showed that the recombinant bacteria harboring pQE-80L/TsSP expressed a protein band with 45.2 kDa. After being purified, the rTsSP had a single distinct protein band (Fig 4). The molecular weight (45.2 kDa) of the rTsSP was consistent with its expected size.
To determine humoral immune responses to rTsSP in immunized mice, serum specific anti-rTsSP IgG titers at days 10 after the final immunization were measured by ELISA. As shown in Fig 5, anti-rTsSP antibodies could be triggered by the immunization with rTsSP. The titer of serum anti-rTsSP IgG was 1:105 following the last immunization, indicating that the rTsSP has a high immunogenicity.
The results of SDS-PAGE analysis showed that the ML crude antigens had 44 bands with MW of 14.7–97.2 kDa, ML ES antigens had 29 bands with 14.4–96.3 kDa, and the rTsSP had only one band with 45.2 kDa (Fig 6A). On Western blot analysis the rTsSP was probed with anti-rTsSP serum and infection serum. The native TsSP proteins with 25–47 kDa in T. spiralis ML crude and ES proteins were recognized with anti-rTsSP serum (Fig 6B). Furthermore, the rTsSP was also recognized by immune serum from mice immunized with ML ES or crude antigens (Fig 6C). The results indicated that TsSP is one protein component from somatic and ES products of T. spiralis ML.
The TsSP transcription at different T. spiralis stages was assayed by RT-PCR assay and the transcription of GAPDH gene was used as an internal control. The TsSP mRNA transcript (1236 bp) was observed at all T. spiralis lifecycle stages (NBL, ML, IIL and AW). Moreover, the primers for the housekeeping gene (GAPDH) also produced the expected band (570 bp) in different developmental stages (Fig 7).
The results of ELISA revealed that the rTsSP and the native TsSP in crude and ES products of different stages (NBL, ML, IIL and AW) were identified by using anti-rTsSP serum (Fig 8). The results of Western blot analysis demonstrated that the native TsSP of 45.2 kDa in crude antigens of various stages were also probed with anti-rTsSP serum (Fig 9). These results further indicated that the TsSP was expressed at various developmental phases, and existed in both the somatic and ES proteins of the nematode. The TsSP expression level in IIL and NBL were obviously higher than those in the other three stages (ML, AW at 3 and 6 dpi) (P < 0.05).
The IFT using intact parasite revealed that the immunostaining was found on cuticles of different stages (AW, NBL, ML and IIL) by using anti-rTsSP serum (Fig 10). While tissue sections of the nematode were incubated by anti-rTsSP serum, the staining was detected in cuticles and stichosomes of ML, IIL, AW and the embryos within uterus of female adult at 3 dpi.
The sensitivity of rTsSP-ELISA and ES-ELISA for detection of anti-Trichinella IgG in serum samples from trichinellosis patients was 98.11% (52/53) and 88.68% (47/53), respectively (χ2 = 2.910, P = 0.088). As the patients’ serum samples at 35 dpi were tested, the sensitivity of two antigens reached 100% (32/32). Nevertheless, while the patients’ samples at 19 dpi were examined, the sensitivity of the rTsSP was 95.24% (20/21), which was obviously higher than 71.43% (15/21) of ES antigens (χ2 = 4.286, P = 0.038) (Table 1). The specificity of the rTsSP and ES antigens was 99.53% (211/212) and 91.98% (195/212) (χ2 = 14.853, P = 0), when they were applied for detecting anti-Trichinella IgG in sera of patients with other parasitosis and healthy individuals. The cross-reaction of rTsSP with sera of patients with other parasitic diseases was not observed except for one serum sample from patients with cysticercosis (Fig 11).
Serum anti-Trichinella IgG levels in infected mice at different time intervals post infection were measured by rTsSP-ELISA and ES-ELISA, respectively. Specific anti-Trichinella IgG was first detected by rTsSP-ELISA on 7 dpi and antibody positive rate reached 100% on 10 dpi (Fig 12A); when ES-ELISA was used, the specific antibody was first detected on 10 dpi and antibody detection reached 100% on 16 dpi (Fig 12B).
Previous studies showed there is an evident 2–3 week window of anti-Trichinella IgG negative after Trichinella infection, the antibody detection rate could not attain 100% till 1–3 months following Trichinella infection in humans [43,44]. The conventional ELISA with ML ES antigens lacks perfect sensitivity at the beginning of Trichinella infection, so improvements of diagnostic antigens would be of clinical value. In theory, detection of circulating antigens or DNA from T. spiralis live worms seems an ideal early diagnostic method for trichinellosis. But the levels of Trichinella circulating antigens in serum samples are usually lower and its detection rate in patients with clinical trichinellosis was usually only 30–50% [45,46]. Moreover, the persistence of Trichinella DNA is transient in blood circulation and the feces of infected hosts [23,47]. Therefore, determination of Trichinella circulating antigens or DNA has not been used for diagnosis of human trichinellosis. Up to now, determination of anti-Trichinella IgG is the most widely applied diagnostic method of trichinellosis, which is recommended by WHO and the International Commission on Trichinellosis (ICT) [10,48]. Therefore, it would be beneficial to identify antigens better able to diagnose recent Trichinella infection.
Serine protease (or serine proteinase) is a superfamily of widespread proteolytic enzymes in parasites, they exert an important part in physiological and pathological proceses during parasite infection [49]. The protease is related with the larval invasion, molting, digestion and fibrinolysis in parasitic nematodes [50,51]. Previous studies indicated that some secreted serine proteases were found in ES products from T. spiralis ML and AW, including serine protease TspSP-1 and trypsin-like 45 kDa antigen [52,53], and the enzymic activity of the native serine proteases in T. spiralis ML and AW ES proteins was also detected by biochemistry assay [54,55]. Our previous studies demonstrated that while the ML were activated into IIL and cocultivated with intestinal epithelial cells (IEC), the serine protease expression level in IIL stage was evidently increased as compared with ML stage [56,57], suggeting that the serine proteases might be involved in the larval invasion of host’s enteral mucosa. These serine proteases might be the target molecular antigens of the early host’s immune response, and they are possiblly used as the new diagnostic antigens for early trichinellosis [58].
The complete TsSP cDNA sequence was cloned and expressed in this study. The TsSP is attributed to the trypsin-like serine protease superfamily and has 90% identity with T. nativa which is another encapsulated Trichinella species [59]. After being purified, the rTsSP was strongly immunogenic and used for generating anti-rTsSP antibodies. Immunization of mice with the rTsSP elicited specific humoral immune response against rTsSP. The ELISA results revealed that the titer of specific anti-rTsSP IgG in immune serum was 1:105. On Western blotting, the rTsSP protein was recognized with anti-rTsSP serum and mouse infection serum. As shown in Fig 6B, by using anti-rTsSP serum several native TsSP proteins was identified in T. spiralis ML crude and ES antigens. The TsSP might have different isoforms, or the protein was possibly processed by means of post-translational modifications/alternative splicing [11,60,61]. The process might be involved in the phosphorylation, methylation or acetylation of the TsSP after being translated, and they are possible important for the biological functions of the TsSP [38,62,63]. Additionally, it is also possible because the TsSP is a member of serine protease family, and they have the same functional domains.
The TsSP mRNA transcription was detected by RT-PCR at all T. spiralis developmental stage (AW, NBL, ML, IIL) (Fig 7). The TsSP expression was found by ELISA at various stage, but as shown in Fig 9, the TsSP expression level in IIL and NBL were obviously higher than those in the other three stages (ML, AW at 3–6 dpi) on Western blot anlysis. The IFT results demonstrated immunostaining was principally located in cuticle and stichosome of the nematode (Fig 10). Our results indicated that the TsSP was expressed at various T. spiralis phases and the TsSP was likely from the worm’s ES products. Previous studies showed another serine proteases (TspSP-1.2) was also expressed in T. spiralis different stages [38]. The results suggested that the TsSP is an essential protein and act a pivotal part in T. spiralis larval invasion and development. The the enzymatic activity and biological funtions of the rTsSP need to be studied in further experiments.
To investigate the potential use of rTsSP for serodiagnosis of human trichinellosis, rTsSP-ELISA method was establised and applied to assay anti-Trichinella IgG in trichinellosis patients’ serum samples, and the sensitivity was compared with those of ES-ELISA. The results revealed that the sensitivity of rTsSP-ELISA and ES-ELISA was 98.11% (52/53) and 88.68% (47/53), respectively (P > 0.05). Nevertheless, while the trichinellosis patients’ serum samples at 19 dpi were examined, the sensitivity (95.24%) of rTsSP was significantly higher than 71.43% (15/21) of ES antigens (P < 0.05), demostrating that the rTsSP protein was useful for the early diagosis of human trichinellosis. The specificity (99.53%) of the rTsSP was also superior to 91.98% of the ES antigens (P < 0.01). The cross-reaction of the rTsSP was seen only with one out of 20 serum samples of cysticercosis patients. The sensitivity and specificity of rTsSP are similar with that of recombinant T. spiralis 31 kDa protein [11]. The sensitivity of rTsSP for diagnosing early trichinellosis is comparative to those of ELISA using IIL or AW ES antigens, but the specificity of rTsSP-ELISA has an evident advantage over those of IIL and AW ES antigens [16,17]. Importantly, the anti-Trichinella IgG in 100% of the mice infected with 300 muscle larvae was detected by rTsSP-ELISA as soon as 10 dpi, but the ES-ELISA did not permit detection of 100% of infected mice before 16 dpi. The results suggested that the TsSP protein might be secreted by the nematode into the host’s peripheral blood circulation at the early infection stage and elicited an early specific anti-Trichinella antibody response continuing to the muscle stage [16]. Furthermore, our previous study has showed that the rTsSP could be recognized by early mouse infection sera at 8–10 dpi on Western blotting analysis [58]. Consequently, the rTsSP could be of value as potential novel antigen for the early diagnosis of T. spiralis infection in humans.
In summary, this study demonstrated that the TsSP was expressed at various T. spiralis developmental stages, it was likely from the worm’s ES products, and mainly located in cuticle and stichosome of this nematode. The rTsSP was strongly immunogenic. Sensitivity and specificity of rTsSP for detecting anti-Trichinella IgG antibodies are superior to the conventional ML ES antigens which are widely used at present. The rTsSP had the potential valuable as a new diagnositic antigen for early trichinellosis. But more serum samples from patients with trichinellosis and other nematode infection (ascariasis, trichuriasis, hookworm infection, filariasis, etc.) should be tested to further evaluate its sensitivity and specificity.
|
10.1371/journal.ppat.1005238 | A New Glycan-Dependent CD4-Binding Site Neutralizing Antibody Exerts Pressure on HIV-1 In Vivo | The CD4 binding site (CD4bs) on the envelope glycoprotein is a major site of vulnerability that is conserved among different HIV-1 isolates. Many broadly neutralizing antibodies (bNAbs) to the CD4bs belong to the VRC01 class, sharing highly restricted origins, recognition mechanisms and viral escape pathways. We sought to isolate new anti-CD4bs bNAbs with different origins and mechanisms of action. Using a gp120 2CC core as bait, we isolated antibodies encoded by IGVH3-21 and IGVL3-1 genes with long CDRH3s that depend on the presence of the N-linked glycan at position-276 for activity. This binding mode is similar to the previously identified antibody HJ16, however the new antibodies identified herein are more potent and broad. The most potent variant, 179NC75, had a geometric mean IC80 value of 0.42 μg/ml against 120 Tier-2 HIV-1 pseudoviruses in the TZM.bl assay. Although this group of CD4bs glycan-dependent antibodies can be broadly and potently neutralizing in vitro, their in vivo activity has not been tested to date. Here, we report that 179NC75 is highly active when administered to HIV-1-infected humanized mice, where it selects for escape variants that lack a glycan site at position-276. The same glycan was absent from the virus isolated from the 179NC75 donor, implying that the antibody also exerts selection pressure in humans.
| CD4bs is a central viral vulnerability site and isolation of new anti-HIV-1 CD4bs broadly neutralizing antibodies (bNAbs) provides information about viral escape mechanisms. Here we describe a new anti-HIV-1 bNAb that was isolated from an HIV-1 infected donor. The antibody, 179NC75, targets the CD4 binding site in a glycan-dependent manner. Although many CD4bs antibodies have been already described, a glycan-dependent mode of recognition is unusual for anti-HIV-1 CD4bs bNAbs. The glycan-dependent CD4bs antibodies have never been tested for their ability to neutralize HIV-1 in vivo. We infected humanized mice with HIV-1YU2 and treated them with 179NC75 three weeks after infection. We observed a drop in viral load immediately after treatment followed by a viral rebound. The viral rebound was associated with specific escape mutations in the plasma virus envelope, resulting in a deletion of N276 glycan, and in some cases a glycan shift from position 276 to position 460. Similar signature mutations were found in the envelope of the autologous virus cloned from patient’s plasma. This defines the escape pathways from 179NC75, and shows that they are the same in humans and in HIV-1YU2 infected humanized mice.
| Although the envelope glycoproteins (Env) of primate immunodeficiency viruses have extremely variable sequences [1], most of them engage CD4 as the primary cellular receptor to initiate the viral life cycle [2]. The consequence is that the CD4 binding site (CD4bs) is a comparatively well-conserved region of Env that serves as a critical neutralization epitope and an appealing vaccine target. The introduction of single cell antibody cloning techniques [3,4] yielded dozens of broad and potent CD4bs antibodies from infected individuals, some of which neutralize ~90% of HIV-1 strains in vitro [5–7]. Some of these antibodies are also effective at reducing viral load when used to treat infected humanized mice (hu-mice) [8], macaques [9–11] and humans [12].
The most potent group of CD4bs antibodies characterized to date is derived from two VH genes, IGVH1-2 [5,7,13] and IGVH1-46 [6,7,14–16]. These antibodies engage many of the same Env residues as CD4. For example, residue Arg71HC in VRC01-like bNAbs interacts with residue Asp368gp120 on Env, and thereby mimics how Arg59CD4 interacts with the same residue when CD4 binds to gp120 [6,7,13,16]. Although the light chains are less restricted in their origin, specific alterations are required for activity, including mutations and deletions [6,13,16]. Overall, the restricted origins and complex development of these bNAbs from their inactive germline precursors may explain why it has been so difficult to elicit them by vaccination.
A second, far more diverse group of CD4bs-directed antibodies is often referred to as ‘CDRH3-dominated class of CD4bs antibodies’. These antibodies use their CDRH3-loop regions to engage Env [15]. These include b12 [17], HJ16 [18], CH103 [19] and the recently described VRC13 and VRC16 [15]. Structural analyses indicate that all CDRH3-dominated antibodies use loop-based recognition mechanisms, with the CDRH3 contributing 50%-70% of the paratope interface [15,19,20]. They are not VH-restricted since their CDRH3s are randomly assembled from IgH variable, diversity and joining segments during V(D)J recombination [21]. In keeping with their diverse origins, CDRH3-dominated antibodies seem to employ different mechanisms of recognition and they also vary in the angles with which they approach the CD4bs [15].
To isolate new CD4bs bNAbs, we sought HIV-1 infected donors whose sera contained potent neutralizing antibodies that appeared to target the CD4bs. One such donor was EB179. By sorting peripheral blood mononuclear cells (PBMCs) from this individual we isolated a new antibody, 179NC75, that is encoded by IGVH3-21 and IGVL3-1 gene segments. In TZM.bl neutralization assays 179NC75 showed an overall IC80 of 0.42 μg/ml against 120 Tier-2 HIV-1.
Binding assays using various Env-based proteins indicated that 179NC75 is glycan-dependent and belongs to the same sub-class of CDRH3-dominated CD4bs antibodies as HJ16. These glycan-dependent CD4bs antibodies have not yet been tested for activity in vivo. To do so we treated humanized mice infected with HIV-1YU2 with 179NC75 and found that it selects for escape variants with mutations in the potential N-linked glycosylation site at gp120 position 276. Similar mutations were also found in the autologous isolate from the 179NC75 donor, suggesting that selection pressure had been exerted in the human host.
For the human studies, The Rockefeller University Institutional Review Board approved all studies involving patient enrollment, sample collection, and clinical follow-up. Donor EB179 was selected from a group of long-term non- progressors that was followed at the Ragon Institute in Boston, and is also referred to as subject 330183. The subject described in this study provided written informed consent prior to participating in this study. For the mouse studies, 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 (IACUC) of The Rockefeller University, and in accordance with established guidelines and policies at The Rockefeller University (protocol number 13618-H).
Purified IgG samples from 394 HIV-1-infected long-term non-progressors were screened for neutralizing activity against a panel of 14 viruses representing 8 different clades or inter-clade recombinants. IgG purified from donor EB179 was exceptional in its neutralization potency and breadth, ranking within the top 2% of the cohort, and neutralized 11 out of the 14 viruses in the panel (S1A Table). A single leukapheresis sample was obtained 4.5 years after initial diagnosis with clade B HIV-1 infection, at age 44. At the time the sample was collected, donor EB179 had 1038 CD4 T cells/mm3 and a viral load of 3180 copies/ml and was not receiving antiretroviral therapy. Molecular HLA typing revealed HLA A*02:01, 68:02; B*07:02, 53:01; Cw04:01, 07:02; DRB11:01, 15:01.
Single-cell sorting of 2CC core+CD19+IgG+ B cells from donor EB179’s PBMCs was conducted as previously described [3]. Briefly, we sorted memory B cells using the gp120 2CC core protein as bait [22]. Rescue primers were used to amplify both heavy chains [7] and Igλ genes [23]. All PCR products were sequenced and analyzed for Ig gene usage, CDR3, and the number of VH/VL somatic hypermutations (IgBLAST, http://www.ncbi.nlm.nih.gov/igblast/ and IMGT, http://www.imgt.org/). Multiple sequence alignments were performed using the MacVector program (v.13.5.5) with the ClustalW analysis function (default parameters), and then used to generate dendrograms by the neighbor-joining method (with best tree mode and outgroup rooting). To specifically isolate members of the 179NC75 clone we used the following forward primers for the heavy chains: 5’-CTGCAACCGGTGTACATTCTGAAATGAGATTGGAAGAAT-3’ and 5’-CTGCAACCGGTGTACATTCTGAGGTCCAGTGTGAAGAA-3’ (in a 1:1 mix); and for the light chains: 5’-ATGGCCTGGATCCCTCTACTTCTC-3’ and 5’- ATGGCATGGATCCCTCTCTTCCTC-3’ (in a 1:1 mix). The reverse primers were the same as described previously for Ig gene amplification [7].
Purified, digested PCR products were cloned into human Igγ1-, IgK or Igλ-expression vectors as previously described [24]. Antibodies were produced by transient transfection of IgH, IgK and IgL expression plasmids into exponentially growing HEK 293T cells (ATCC; CRL-11268) using polyethyleneimine (PEI)-precipitation [24]. IgG antibodies were affinity purified using Protein G Sepharose beads according to the manufacturer’s instructions (GE Healthcare).
High-binding 96-well ELISA plates (Costar) were coated overnight with 5 μg/ml of purified 2CC core, gp120.YU2 (wild type or mutants) or gp140.YU2 foldon trimer in PBS. After washing 6 times with PBS + 0.05% Tween 20, the plates were blocked for 2 h with 2% BSA, 1 μM EDTA and 0.05% Tween-PBS (“blocking buffer”), and then incubated for 1 h with IgGs that were added as seven consecutive 1:4 dilutions in PBS from an initial concentration of 4 μg/ml. After additional washing, the plates were developed by incubation with goat HRP-conjugated anti-human IgG antibodies (Jackson ImmunoResearch) (at 0.8 μg/ml in blocking buffer) for 1 h followed by HRP chromogenic substrate (ABTS solution; Invitrogen). For competition ELISAs, the plates were coated with 5 μg/ml 2CC core, gp120 or gp140 foldon, washed, blocked for 2 h with blocking buffer and then incubated for 1 h with IgGs added as seven consecutive 1:4 dilutions in PBS from an initial concentration of 32 μg/ml, and in the presence of biotinylated 179NC75 antibody at a constant concentration of 4 μg/ml. The plates were then developed using HRP-congugated streptavidin (Jackson ImmunoReseach) (at 1 μg/ml in blocking buffer).
For ELISAs using BG505 SOSIP.664-D7324 trimers, the plates were coated overnight with 5 μg/ml of D7324 antibody as previously described [25], washed and then incubated with 500 ng/ml of the trimer [25,26]. After a further wash, IgGs were added for 1 h as seven consecutive 1:4 dilutions in PBS from initial concentrations of 4 μg/ml. The endpoint was generated by incubation with goat HRP-conjugated anti-human IgG antibodies, as described above. All experiments were performed at least 3 times.
For EndoH ELISA, the plates were coated overnight at 4°C with 5 μg/ml of EndoH-treated or untreated gp120 in 100 mM sodium bicarbonate/carbonate buffer, pH 9.6. They were then washed with TBS + 0.05% Tween 20 and blocked for 1 h in the same buffer supplemented with 3% (w/v) BSA, and washed again before test antibodies were added for 2 h. After a final wash, the endpoint was generated using goat HRP-conjugated anti-human IgG antibodies, again as described above.
HIV-1 neutralization was evaluated using the luciferase-based TZM.bl cell assay as described previously [27]. Briefly, envelope pseudoviruses were incubated with fivefold serial dilutions of single antibodies and applied to TZM.bl cells that carry a luciferase-reporter gene. After 48 h cells were lysed and luminescence was measured. IC50 and IC80 reflect single antibody concentrations that caused a reduction in relative luminescence units (RLU) by 50% and 80%, respectively.
NOD Rag1−/−Il2rgnull (NOD.Cg-Rag1tm1Mom Il2rgtm1Wjl/SzJ) mice were purchased from The Jackson Laboratory and bred and maintained at the Comparative Bioscience Center of The Rockefeller University according to guidelines established by the University’s Institutional Animal Care and Use Committee. All experiments were performed under protocols approved by the same committee. Hu-mice were treated with 1 mg of 179NC75 sub-cutaneously (s.c.) on day 0, followed by 0.5 mg s.c. injections twice-weekly for a period of 5 weeks [8]. The gp120 sequences from escape variant viruses were obtained as previously described [8].
The autologous virus from donor EB179 was isolated as previously described [28]. Briefly, CD19 and CD8-depleted mononuclear cells were cultured at a concentration of 5 × 106 cells/ml in Iscove's modified Dulbecco’s medium (IMDM; Gibco) supplemented with 10% fetal bovine serum (FBS; HyClone, Thermo Scientific), 1% GlutaMAX (Gibco), 1% penicillin/streptomycin (Gibco), and 1 μg/ml phytohaemagglutinin (Life Technologies) at 37°C and in an atmosphere containing 5% CO2. After 2–3 days, 5 × 106 cells were transferred into IMDM supplemented with 10% FBS, 1% penicillin/streptomycin, 5 μg/ml polybrene (Sigma), and 100 IU/ml of IL-2. The medium was replaced weekly and the HIV-1 content of culture supernatants was quantified using the Lenti-X p24 Rapid Titer Kit (Clontech) according to the manufacturer’s instructions. Env genes from the autologous virus were cloned by reverse transcriptase PCR as described elsewhere [29].
Single, double and triple mutations were introduced into wild-type HIV-1YU2 envelope using the QuikChange (multi-) site-directed mutagenesis kit, according to the manufacturer’s specifications (Agilent Technologies).
Polyclonal IgG purified from donor EB179 had exceptional neutralization capacity, with respect of potency and activity against 11 of 14 Tier-2 viruses in a small cross-clade panel (S1A Table). To map the predominant NAb specificities, we tested EB179 IgG against HIV-1YU2 mutants that are resistant to NAbs targeting the trimer apex (N160K), the CD4bs (N280Y) or the base of the V3 loop (N332K) [8,30–32]. Among these mutants, only HIVYU2N280Y was resistant to EB179 IgG (S1B Table). We conclude that at least a proportion of the neutralization activity present in this serum is directed to the CD4bs.
To isolate and characterize the NAbs present in EB179, we used flow cytometry to sort memory B cells that bound to 2CC core, a gp120 antigen that presents the CD4bs in an exposed and stable conformation [22]. Among CD19+IgG+ B cells, ~0.2% bound strongly to 2CC core. Of the 372 cells sorted, 87 produced paired heavy and light chains, 36 of which represented ten clonally related families (Fig 1A). Antibody sequences obtained from the expanded B cell clones contained higher numbers of somatic mutations compared to antibodies obtained from B cells that appeared only once (S1 Fig). The average number of nucleotide mutations in the heavy chain of clonal sequences was 44.76 (± 3.66, N = 36) compared to 20.82 (± 1.39 N = 51) for unique sequences (S1A Fig). A similar trend was observed when the light chain sequences were analyzed (S1B Fig).
Representative variants from each of the clonal families were selected for further analysis (S2 Table). These variants were expressed as IgG1 antibodies that were tested for binding to a HIV-1YU2 gp140 foldon protein [33] or 2CC core [22], and for neutralizing activity. Except for 179NC9055, all the antibodies bound strongly to the HIV-1YU2 gp140 and/or 2CC core proteins (Fig 1B), and members of clones 1, 2, 3, 4, 6, and 7 neutralized the Tier-1 (i.e., neutralization-sensitive) HIV-1BAL virus (Fig 1C). While antibodies from clones 3 and 7 were only weakly active against the other viruses in the panel, one representative of the most expanded clone 1 (179NC75) strongly neutralized four of the five viruses tested (IC50 ≤0.05 μg/ml, Fig 1C).
To isolate additional 179NC75 variants, we amplified cDNA from the 2CC core+CD19+IgG+ single-sorted B cells using specific VH and VL forward primers (see Methods). We obtained a total of 23 heavy chain and 25 light chain variants from the 179NC75 clonal family. The heavy and light chain sequences carried 34% and 29% amino acid mutations on average, respectively, compared to their germline gene segments IGVH3-21 and IGVL3-1. The various sequences of the 179NC75 clone were similar by up to 73% from clonal members (Fig 2A and 2B). The CDRH3 and CDRL3 regions were 24 and 10 residues long, respectively (Fig 2A and 2B, S2 Table). There were no insertions or deletions.
Variants 179NC 54, 60, 65, 75, 21 and 1055 (indicated in Fig 2A and 2B) were tested for activity against a panel of nine Tier-2 viruses, including three from clade B, one from clade C, two from clade A, two clade A/G recombinants and one clade A/E recombinant. 179NC75 and two closely related variants, 179NC54 and 179NC60, potently neutralized 6 of these 9 viruses, whereas the other antibodies had lesser or no neutralization activity (Fig 2C). Accordingly, we selected 179NC75 for additional analyses.
When tested against an extended cross-clade panel of 120 Tier-2 viruses, 179NC75 neutralized viruses from clades B particularly strongly (S3 Table); its geometric mean IC50 and IC80 values were 0.113 μg/ml and 0.291 μg/ml, respectively (S4 Table). When compared to other CD4bs bNAbs against a panel of 22 Tier-2 clade, B viruses, 179NC75 was more potent than b12 against 13 viruses, than HJ16 against 15 viruses, than VRC01 against 8 viruses, and than CH103 against 6 viruses (Fig 2D). Its overall breadth of activity across the clade B virus panel was 70% (S4 Table).
To map the epitope targeted by 179NC75 and its clonal variants, we performed a series of ELISAs. All members of the 179NC75 clonal family bound to HIV-1YU2 gp120, gp140 foldon [34] and 2CC core [22] proteins (S2 Fig). In a competition ELISA, soluble CD4 (sCD4) and most CD4bs antibodies competed with 179NC75 for binding to gp120YU2, whereas PGT121, PGT128 and 10–1074 did not (Fig 3A upper and lower panels). The 8ANC195 bNAb, which binds an epitope adjacent to the CD4bs [7,35], inhibited 179NC75 binding by ~ 50% (Fig 3A, lower panel). We conclude that the 179NC75 epitope is proximal to the CD4bs.
We next tested how different mutations in the CD4bs affected 179NC75 binding. The D368R single mutation was not sufficient to affect the gp120-binding of 179NC75 family members, but the D368R and N280Y double mutation substantially impaired their binding. In contrast, VRC01 is sensitive to the single D368R substitution (Fig 3B, right upper and lower panels).
The Asn276 glycan site is important for the binding of two different bNAbs: the CD4bs antibody HJ16 [36] and the gp120-gp41 specific antibody 8ANC195 [7,35]. The 8ANC195 epitope lies outside the CD4bs and this antibody binds Env in the presence of CD4 [7,35]. Since HJ16 strongly inhibited 179NC75 binding (Fig 3A, upper panel) and 8ANC195 did so weakly (Fig 3A, lower panel), we assessed whether the binding of 179NC75 family members was affected by the N276D substitution and found that it had a profound impact (Fig 3B, lower left panel). In contrast, the N276D change had no effect on VRC01 binding, as previously reported [37] (Fig 3B). When monomeric gp120s from both YU2 and the clade A/E virus 93TH057 [38] were treated with EndoH, a glycosidase that removes N-linked oligomannose glycans, the binding of 179NC75 and its clonal variants was completely abolished (Fig 3C). To further probe the nature of the glycan-dependency of 179NC75, we tested binding of the Fab to BG505 SOSIP.664 trimers, (fully glycosylated, cleaved, native-like, soluble trimers [25]) produced in HEK293-6E cells in the presence and in the absence of the mannosidase I inhibitor kifunensine. HEK293-6E cells fully process glycans resulting in a mixture of complex-type and high-mannose N-glycans, while HEK293-6E cells treated with kifunensine, produce protein containing only high-mannose N-glycans. We observed that 179NC75 binds to BG505 SOSIP.664 trimer with processed glycans with a KD of ~90 nM, (S3 Fig) but cannot bind to trimers containing only high mannose glycans (S3 Fig, S6 Table). Hence, we conclude that 179NC75 is a glycan-dependent antibody that binds to the CD4bs in a way that involves the Asn276 residue and depends on the presence of complex glycans. In these respects, its epitope is similar to that of the HJ16 CD4bs bNAb.
We compared the neutralization potencies of 179NC75 to the ones of HJ16 [18]. For the 53 Tier-2 viruses that were tested against both HJ16 and 179NC75, 179NC75 neutralized more viruses than HJ16 (26 compared to 19), and was 20-fold more potent (IC50 of 0.118 μg/ml compared to 2.326 μg/ml) (Fig 3D).
Previous reports show that neutralizing antibodies bind BG505 SOSIP.664 trimers with higher affinity as opposed to non-neutralizing antibodies [25]. Therefore, as expected, the more potent variants of the 179NC75 clone, 179NC75 and 179NC1055, bound strongly to BG505 SOSIP.664-D7324 trimers in capture ELISA, while 179NC65 and 179NC21 bound weakly or not at all, respectively (S4 Fig).
Most predicted germline versions of CD4bs antibodies are unable to bind Env antigens [7]. To test whether the germline version of 179NC75 could bind the BG505 SOSIP.664-D7324 trimers, and assess the role of CDRH3 in trimer binding, we generated a germline version of 179NC75 (179NC75gl). The predicted germline version of the antibody was made as previously described by reverting the V and J segments of the heavy and light chains to their predicted germline sequences, while retaining the CDRH3 sequence as found in the mutated antibody [7,39,40]. For comparison, we used the previously published predicted germline versions of VRC01 [39,40], 3BNC60 [7], 1NC9 [7], CH103 [19] and HJ16 (constructed in the course of this study). Although all of the above mature CD4bs bNAbs bound the BG505 SOSIP.664-D7324 trimers, the only predicted germline antibody able to do so was 179NC75gl (Fig 4). An implication is that 179NC75 binding principally involves contacts made by the CDR3s, particularly the exceptionally long (24-residue) heavy chain CDR3.
The loop binding, glycan-dependent CD4bs bNAbs have not been tested for their activity in vivo. To address this issue, we treated six HIV-1YU2–infected hu-mice with 179NC75 for 5 weeks [8,29]. Monotherapy with 179NC75 resembled monotherapy with other bNAbs, in that there was a transient decrease in viral load in most of the treated animals followed by a rapid rebound [8,29,41] (Fig 5A and 5B). Viral env genes were cloned and sequenced from the day-28 plasma of 179NC75-treated mice, a time point where viremia had universally rebounded to levels similar to the day-0 value. Two types of mutations were consistently observed, both proximal to the CD4bs: the first eliminated the glycan-site at position N276; the second involved residues G459 or K460 (Fig 5C). In total, 13 sequences had only a mutation affecting the N276 glycan site, whereas 8 contained mutations in the region near position 460 and 7 sequences contained mutations in both regions (Fig 5C and 5D). In all mice the rebounding viruses carried at least one of these mutations. In mouse 1107 mutations in both areas were observed, resulting in the loss of the N276 glycan but the introduction of a potential N-linked glycosylation site (PNGS) at position 460 (Fig 5C). To confirm that the most commonly observed mutations did confer resistance, HIV-1YU2 Env-pseudoviruses containing one or both of the N276D and K460N substitutions were tested for their sensitivity to 179NC75. All three of the virus mutants were found to be 179NC75-resistant (Fig 5E). We also tested the HIVYU2 N280Y, N332K, N160K and G459D virus mutants. As expected, and consistent with the ELISA data, the N280Y substitution conferred complete resistance to 179NC75, while the N332K and N160K changes had no effect. The G459D mutant was also 179NC75-sensitive (Fig 5E). We conclude, that 179NC75 is a potent neutralizing antibody that exerts selection pressure on HIV-1YU2 in vivo and drives the emergence of resistant viruses with sequence changes proximal to the CD4bs.
To test whether 179NC75 exerted selective pressure on the autologous virus found in subject EB179, we cloned env genes from the donor’s T cells obtained at the time of the leukapheresis. All nine gp120 sequences obtained contained Asn at position 460, introducing a PNGS at that position in eight of the nine sequences (Fig 6A–6C). Five sequences contained an Asn-Gly-Thr insertion immediately N-terminal to position N460, resulting in the sequence NGTNET, and therefore adding another PNGS to the one that was already at position 460. Five other sequences contained the N276S mutation, eliminating the PNGS at position 276. One of the nine sequences included both the Asn-Gly-Thr insertion at position 460 and the N276S change (Fig 6C). Of note is that this pattern of sequence changes is highly similar to the escape mutations seen in the env genes of the 179NC75-treated, HIV-1YU2-infected hu-mice (Fig 5). To test whether the autologous virus from patient EB179 is resistant to 179NC75, we cultured the donor’s CD4+ T cells from the same leukapheresis sample that was used for the antibody isolation. Outgrown virus was then tested for neutralization in the TZM.bl assay for neutralization by the EB179 polyclonal IgG (from the same time point), as well as by 179NC75 and other known bNAbs including the CD4bs antibody 3BNC117 [7], the V3-stem binding antibody 10–1074 [42] and the V1/V2 apex-binding antibody PG16 [43] (Fig 6D). As expected, the EB179 polyclonal IgG failed to neutralize the autologous virus. Amongst the two CD4bs antibodies, the autologous virus was fourfold more resistant to 179NC75 than to 3BNC117, suggesting that the EB179 antibody repertoire has CD4bs antibodies that differ from 3BNC117 and VRC01-class bNAbs. Interestingly, the autologous virus was also resistant to PG16 and 10–1074, indicating that the patient may have additional neutralizing antibodies bearing similar specificities in his antibody repertoire. Taken together, the data imply that loop-based, glycan-dependent CD4bs bNAbs of the 179NC75 family exert selective pressure on HIV-1 in vivo.
The CD4bs is a highly conserved epitope on the HIV-1 Env and an important potential target for neutralizing antibodies. Although this site evolved to avoid antibody accessibility, two major groups of CD4bs bNAbs have been discovered [15]. The first group, exemplified by VRC01, is VH-restricted, IGVH1-2 or IGVH1-46, with the heavy chains positioned in a CD4-like orientation and CDRH2 making significant contacts with gp120 [6,7,15]. The CDRL3 [7,21,44], and in some cases also CDRL1 [6], of the corresponding light chains have to be short and compact to minimize potential interference and clashes with the glycans that surround the CD4bs. The emergence of these antibodies involves many somatic hypermutations, some of which are in the framework regions [45]. The second group of CD4bs bNAbs, which includes b12 and HJ16, is far more heterogeneous. These antibodies bind to gp120 via a CDRH3-dominated, loop based mechanism [15]. As might be expected, members of this group of CD4bs bNAbs arise from different VH segments and carry fewer somatic mutations [17–19]. The new antibody described in this study, 179NC75, is a loop binder that is closely related to HJ16. Similarly to HJ16, its Env-binding and virus-neutralizing activities are dependent on the N276 glycan [36]. Consistent with the CDRH3 loop-based mechanism of recognition that was described for antibodies that are not VH-restricted [15], when we generated the predicted germline version of 179NC75, where all mutations were reverted but the CDRH3 was retained, the antibody bound to BG505 SOSIP.664 trimers. This could indicate that any residual mutations present in the CDR3s of the reverted antibody might allow binding. Interestingly, the germline version of HJ16 also had some binding to BG505 SOSIP.664 trimers (Fig 4), however this binding was lower that the one of 179NC75, which could be attributed to a shorter CDRH3 (19 versus 24 residues).
Serum antibodies that are CD4bs-specific and N276-dependent have been described in HIV-1-infected individuals in two separate studies [32,46]. In the first study, an HJ16-type of CD4bs antibody response was found to be part of the second wave of serum neutralization in the CAP257 patient [46]. Viruses cloned from CAP257 after the emergence of these CD4bs antibodies carried an N276D or T278A mutation that were considered to be responses to antibody selection pressure [46]. In a second study, serum from individual VC1004 contained CD4bs-targeted NAbs that were sensitive to the N276D substitution but not D368R [47]. However, as the antibodies responsible for the serum activity were not cloned in either study much of what we know about the in vivo activity of these N276-dependent class of CD4bs antibodies is inferential. Our 179NC75 therapy experiments in HIV-1–infected hu-mice demonstrate that escape variants contain very similar, and sometimes identical, mutations to ones present in the autologous virus isolated from the infected human from whom the 179NC75 antibody was also derived. We conclude that the CDRH3-dominted N276-dependent CD4bs antibodies are effective at suppressing viremia in vivo and thence driving the emergence of escape variants.
|
10.1371/journal.pgen.1005261 | Dissecting the Function and Assembly of Acentriolar Microtubule Organizing Centers in Drosophila Cells In Vivo | Acentriolar microtubule organizing centers (aMTOCs) are formed during meiosis and mitosis in several cell types, but their function and assembly mechanism is unclear. Importantly, aMTOCs can be overactive in cancer cells, enhancing multipolar spindle formation, merotelic kinetochore attachment and aneuploidy. Here we show that aMTOCs can form in acentriolar Drosophila somatic cells in vivo via an assembly pathway that depends on Asl, Cnn and, to a lesser extent, Spd-2—the same proteins that appear to drive mitotic centrosome assembly in flies. This finding enabled us to ablate aMTOC formation in acentriolar cells, and so perform a detailed genetic analysis of the contribution of aMTOCs to acentriolar mitotic spindle formation. Here we show that although aMTOCs can nucleate microtubules, they do not detectably increase the efficiency of acentriolar spindle assembly in somatic fly cells. We find that they are required, however, for robust microtubule array assembly in cells without centrioles that also lack microtubule nucleation from around the chromatin. Importantly, aMTOCs are also essential for dynein-dependent acentriolar spindle pole focusing and for robust cell proliferation in the absence of centrioles and HSET/Ncd (a kinesin essential for acentriolar spindle pole focusing in many systems). We propose an updated model for acentriolar spindle pole coalescence by the molecular motors Ncd/HSET and dynein in conjunction with aMTOCs.
| During cell division, chromosomes are divided into two daughter cells by the mitotic spindle, a complex structure made from microtubules. The correct formation of the mitotic spindle is essential, as missegregation of chromosomes can lead to cell death or cancer. Therefore several mechanisms cooperate in nucleating the microtubules needed for the mitotic spindle and focusing them into a bipolar structure. One of these mechanisms, which has only recently been identified, is microtubule nucleation by acentriolar microtubule organizing centers (aMTOCs). These structures have been observed in several cell types, notably also in cancer cells, but is not known how they are formed and which function they might have in mitotic spindle assembly. We identified the pathway of aMTOC formation in Drosophila, which enabled us to perturb their formation in order to study their role during spindle formation. We show that aMTOCs are a source of microtubule nucleation, but their contribution to spindle formation is normally masked by other, more dominant, pathways of microtubule nucleation. Furthermore, we have identified a role for aMTOCs in focusing of mitotic spindle poles by the molecular motor dynein in cells in which centrioles are missing.
| Centrosomes are the major microtubule organizing centers (MTOCs) in many cells, and they consist of a pair of centrioles surrounded by a cloud of pericentriolar material (PCM), which contains proteins that nucleate and stabilize microtubules (MTs). For many years centrosomes were thought to be the sole drivers of mitotic spindle assembly in animal somatic cells by nucleating MTs, which then randomly search and capture kinetochores in the cytoplasm to form a bipolar spindle [1]. More recently it has become clear that centrosomal MT nucleation cooperates with at least two other pathways of MT nucleation—the chromatin and augmin-mediated pathways—to drive efficient mitotic spindle assembly [2–4]. In current models, MT assembly is induced around the chromatin and the plus ends of these MTs are then captured by the kinetochores and continue to grow from there in a MT bundle [5–8]. This leads to the formation of kinetochore fibers (K fibers) with minus ends that are pushed away from the kinetochores. These K fibers then coalesce into a bipolar spindle together with astral MTs emanating from the centrosomes in a mechanism involving three distinct steps: First, K fibers become crosslinked by the kinesin-14 Ncd/HSET. Second, K fibers are captured by the tips of growing centrosomal MTs, which is also mediated by Ncd/HSET. Third, K fibers are then transported towards centrosomes along astral microtubules by dynein motors, leading to their integration into the bipolar spindle [6,9]. Furthermore, the augmin complex is required for spindle MT amplification by nucleating MTs that branch off the sides of existing MTs, which increases the density of MTs within the mitotic spindle [10–15].
While all three MT nucleation pathways normally make important contributions to spindle assembly, they appear to be partially redundant. Many cell types that usually have centrosomes can assemble bipolar mitotic spindles in their absence, solely relying on chromatin- and augmin-mediated MT nucleation—although spindle assembly without centrosomes takes longer and is usually less accurate [16–20]. Molecular motors are sufficient for focusing MTs into a bipolar spindle without the need for centrosomes to dictate the organization of the two spindle poles (we define spindle poles as the focused collection of MT minus ends of the spindle [21]). In this case the kinesin-14 Ncd/HSET becomes crucial as it crosslinks MTs in a centrosome-independent way; loss of kinesin-14 proteins has been shown to cause severe pole focussing defects in the absence of centrosomes in female meiosis in Drosophila and mouse oocytes, mitosis in Arabidopsis, and in Xenopus egg extracts [22–25]. The dynein complex also plays a crucial role in acentriolar spindle pole focusing in some systems such as acentriolar spindles reconstituted from Xenopus egg extract or from cell free extracts prepared from HeLa cells [25–28]. The exact mechanism by which dynein contributes to acentriolar pole focusing however is unclear, as its normal function in pole focusing relies on the transport of K fibers towards centrosomes [9], which are not present in this case.
While the chromatin-mediated and augmin-dependent MT nucleation pathways are well studied, our knowledge of other acentriolar mechanisms of MT nucleation during mitosis is limited. One such mechanism has been described in centrosome-free mouse oocytes and early mouse embryos where centrosome function is replaced by multiple acentriolar MTOCs (aMTOCs) to which the centrosomal proteins γ-tubulin and Pericentrin localise [29–31]. These aMTOCs form de novo in prophase in the cytoplasm and around the nuclear envelope, and a bipolar spindle is formed in later stages of meiosis through the progressive clustering of multiple aMTOCs into just two poles [30]. In contrast, much less is known about the nature and function of aMTOCs in somatic cells. The presence of γ-tubulin enriched aMTOCs that mediate the de novo formation of MTs has been described in acentriolar Drosophila cultured cells [32,33]. In acentriolar DT40 chicken cells, aMTOCs containing the pericentriolar proteins CDK5RAP2 and γ-tubulin that nucleate MTs have been described [19], while in monkey cells in which the centrosome has been removed by microsurgery aMTOCs containing Pericentrin could be observed integrating into the mitotic spindle [34]. Furthermore, imaging of spindle formation in pig kidney cells showed that even in the presence of centrosomes peripheral, non-centrosomal MT clusters form and are utilized in spindle formation [35]. Interestingly, the ability to form aMTOCs appears to be upregulated in several cancer cell lines that still contain centrosomes; these aMTOCs lead to the formation of multiple spindle poles that need to be clustered into a bipolar spindle [36].
It is unclear, however, how aMTOCs are formed in somatic cells in the absence of centrioles. Moreover, although aMTOCs appear to contribute to spindle assembly in at least some systems [4,35,36] the significance of aMTOC mediated MT generation in spindle formation in somatic cells is still largely uncharacterized. In order to shed light on these open questions, we decided to study aMTOC formation and function in somatic Drosophila cells in vivo. We first set out to elucidate the pathway of aMTOC assembly. We found that aMTOCs consistently form in ~50–60% of mitotic fly somatic brain cells that lack centrioles, and that aMTOC assembly depends on the same proteins that are required to drive mitotic PCM assembly around the centrioles: Asl, Spd-2 and Cnn [37–45]. By identifying the proteins essential for aMTOC formation we then had means to specifically ablate formation of aMTOC formation in the absence of centrioles. Using these tools we performed a detailed genetic analysis to dissect the contribution of aMTOCs to mitotic spindle assembly in somatic Drosophila cells by comparing different acentriolar fly mutants, which either lack centrioles and aMTOCs or lack centrioles but form aMTOCs. Surprisingly, we find that aMTOCs do not detectably contribute to spindle assembly in the absence of centrosomes. In the absence of both centrosomes and MT nucleation from around the chromatin, however, aMTOCs significantly promote the assembly of monopolar spindles. Most importantly, we show for the first time that aMTOCs are essential for dynein-mediated acentriolar spindle pole focusing. On the basis of these observations we propose a revised model for acentriolar spindle organization by the molecular motors dynein and Ncd.
It had previously been shown that γ-tubulin-containing aMTOCs can be found on spindle poles of acentriolar Drosophila cultured cell lines [33] and in Drosophila cultured cells in which centrioles have been depleted by knocking down the levels of the centriole duplication protein Sas-4 by RNAi [32]. In cell lines derived from Sas-4 mutants, however, aMTOCs could not be observed [46]. To test whether PCM proteins were detectable at acentriolar spindle poles in vivo we examined Drosophila Sas-4 mutant larval brain cells. Electron microscopy studies have shown that these mutant cells lack detectable centrioles [16]. Staining for the centriolar protein Ana1 confirmed the lack of centrioles in Sas-4 mutants (Fig 1A). Despite the absence of centrioles, however, we noticed localisation of the PCM protein Cnn at one or both spindle poles in ~56% of Sas-4 mutant brain cells (Fig 1A and 1B). The staining of these Cnn foci was on average ~60% fainter than Cnn staining on centrosomes in WT cells (Fig 1C), and the acentriolar Cnn foci were on average ~60% smaller in diameter than centrosomes (Fig 1D). These observations support the previous conclusion that Sas-4 mutant brain cells lack centrosomes [16], but suggest that about half of these cells have some detectable Cnn at least one spindle pole.
To analyse the mechanism of recruitment of Cnn to acentriolar spindle poles we analysed GFP-Cnn behaviour, together with Jupiter-mCherry as a MT marker, in living Sas-4 mutant cells. As cells prepared to enter mitosis (judged by nuclear envelope breakdown—NEBD), a small number of GFP-Cnn dots began to appear in the cytoplasm of some cells, from which MTs then appeared to emanate (Fig 2A, arrowheads, S1 Movie). As cells entered mitosis, more Cnn dots of variable size and brightness became visible; these often associated with MTs and they tended to become clustered at the spindle poles as mitosis proceeded (Fig 2A, arrows). Importantly, a similar pattern was observed when we followed the behaviour of the centrosomal proteins Spd-2-GFP, γ-tubulin-GFP and Asl-GFP (although the levels of Asl-GFP in these dots was very low and no dots were detectable using anti-Asl antibodies in fixed cells—S1 Fig) (Fig 2B–2D). Based on the existing literature, we hereafter refer to these PCM foci as aMTOCs.
The PCM proteins PLP, Grip71WD, Msps, Aurora A and TACC all co-localised to some extent with the Cnn-containing aMTOCs in fixed Sas-4 mutant cells, strongly suggesting that these structures contain many of the proteins normally concentrated at centrosomes during mitosis (S1 Fig). We conclude that, as in S2 cells depleted of Sas-4 [32], aMTOCs can form during mitosis in Drosophila cells lacking centrioles in vivo. These aMTOCs appear to contain several known PCM proteins, and are formed prior to mitosis in the cytoplasm where they organize MTs that gradually become localized to the mitotic spindle poles.
Next, we wanted to investigate the molecular pathway of aMTOC assembly. It has recently been shown that mitotic PCM recruitment in flies is largely dependent on just three proteins: the centriole duplication protein Asl, which is recruited to the outer region of the mother centriole, and the PCM scaffolding proteins Spd-2 and Cnn [38]. These are recruited to mother centrioles in an Asl-dependent manner, and assemble into a scaffold-structure that spreads outwards from the centrioles and recruits most other mitotic PCM components [37,38,47]. The centriole duplication protein Sas-4 has also been implicated in PCM recruitment [48,49], so we first tested which, if any, of the core centriole duplication proteins Sas-6, Ana2, Sas-4 or Asl are required for aMTOC assembly.
We quantified Cnn staining on mitotic spindle poles in Sas-6, ana2, Sas-4 and asl mutants, which all lack centrioles in third instar larval stages [16,50–52]. Interestingly, robust aMTOCs (identified by Cnn staining on at least one spindle pole) were present in ~50–60% of mitotic cells from Sas-6, ana2 and Sas-4 mutants, but were essentially undetectable in asl mutants (Fig 3A and 3B), and the same was true when we used other PCM proteins as aMTOC markers (S2 Fig). Moreover, Spd-2-GFP, GFP-Cnn and γ-tubulin-GFP did not detectably form aMTOCs in living asl mutant brain cells (Fig 3C). We conclude that Asl is essential for aMTOC formation in these cells.
We wondered whether Asl, by analogy to the centrosomal mitotic PCM recruitment pathway [38], was able to initiate PCM formation in the cytoplasm even in the absence of centrioles. This might happen by recruitment of Spd-2 and Cnn to cytoplasmic Asl, which could then provide a scaffold for PCM recruitment. We therefore examined PCM localisation on mitotic spindle poles in cnn; Sas-4 and Spd-2 Sas-4 double mutant strains, using the PCM protein γ-tubulin to assess the presence of aMTOCs. While ~51% of mitotic Sas-4 mutant cells had detectable γ-tubulin foci on spindle poles, this was reduced to ~27% in Spd-2 Sas-4 mutant cells and to only ~1% in cnn; Sas-4 mutant cells (Fig 4A and 4B). Thus, Cnn appears to be essential for aMTOC formation, while Spd-2 has an important, but more minor role. Indeed, Spd-2 may function through its effect on Cnn recruitment [38] as Cnn was less efficiently localised at acentriolar spindle poles in the absence of Spd-2 (Fig 4C). Taken together these data suggest that aMTOC formation during mitosis is dependent on the same set of proteins that are essential for mitotic centrosome assembly.
These observations provided us with a strategy to genetically investigate the potential function of aMTOCs in acentriolar spindle assembly. Sas-4 mutant cells lack centrioles, but have aMTOCs, while asl mutants lack both structures. Thus, by comparing the behaviour of Sas-4 and asl mutant cells we can infer the contribution of aMTOCs to potentially any acentriolar biological process.
We first assessed the contribution of aMTOCs to acentriolar mitotic spindle assembly using time-lapse microscopy in mutant neuroblasts expressing the MT marker Jupiter-mCherry and the centriole marker GFP-PACT [53]. The latter protein also concentrates in nuclei during interphase, allowing us to precisely determine the time of NEBD [54] (Fig 5A and S2 Movie). As previously observed both Sas-4 and asl mutant cells were able to form acentriolar spindles [16,40]. On average, both asl and Sas-4 mutant neuroblasts were significantly delayed in spindle assembly compared to WT, confirming previous data in Drosophila and DT40 chicken cells lacking centrioles [16,19]. Surprisingly, however, there was no detectable difference in the timing of NEBD to anaphase onset between asl and Sas-4 mutants (Fig 5B), suggesting that aMTOCs do not detectably contribute to the efficiency of spindle assembly in these cells.
In the absence of centrosomes, the bulk of spindle MTs are nucleated by the chromatin-mediated pathway [55]. We wondered, therefore, whether the contribution to MT nucleation by aMTOCs could be masked by MT nucleation from around the chromatin. In flies, a mutation in misato (mst) inhibits regrowth of MTs from around the chromatin; misato (mst) mutants are larval lethal, and mutant mitotic cells fail in bipolar spindle assembly and usually only organize monopolar spindles of low MT density [56]. Cells that lack both Mst and centrioles can grow acentriolar MT arrays that appear to be nucleated from cytoplasmic foci, but the nature of these foci is unclear [56].
To test whether aMTOCs are formed in the absence of centrioles and the chromatin-mediated spindle assembly pathway we stained mst; Sas-4 and mst; asl double mutant cells with antibodies against α-tubulin and Cnn. Cnn stained aMTOCs on the poles of monopolar spindles in mst; Sas-4 double mutant cells, but not in mst; asl double mutant cells (Fig 6A). To test whether aMTOCs aid in mitotic array formation in the absence of both centrioles and the chromosome-mediated spindle assembly pathway we quantified mitotic cells with monopolar spindles. We found that approximately 60% of mitotic mst; Sas-4 cells contained monopolar spindles, while only ~25% of mst; asl double mutant cells contained monopolar spindles of MTs (Fig 6B). We conclude that aMTOCs help to establish and/or maintain the monopolar spindles formed in cells lacking both the centrosomal and chromatin pathways of spindle assembly.
To test whether aMTOCs nucleate these monopolar MT arrays, or whether they are simply recruited to, and then help to stabilize them, we performed a MT-regrowth assay. In mitotic mst; Sas-4 cells the MTs often grew back from one or several foci in the cytoplasm [56] and we found that these foci were often marked with Cnn (S3A Fig). After ~10 minutes of re-growth, the MTs had all usually coalesced into one acentriolar monopolar array that often had one Cnn-marked aMTOC at the pole (S3A Fig). Quantification of the formation of MT arrays over time showed that in mitotic mst; Sas-4 cells the number of cells with monopolar MT arrays grew steadily over time until, after 30 minutes, ~50% of cells contained monopolar MT arrays again (Fig 6C). Remarkably, in mst; asl cells almost no regrowth of MT arrays was observed, and less than 5% of cells exhibited any MT arrays even after 30 minutes (Figs 6C and S3B). Thus, aMTOCs can nucleate MTs in mitotic cells, and they are an important source of MT generation in mitotic cells lacking centrosomal and chromatin MT nucleation.
In addition to providing a site for MT nucleation during mitosis, centrosomes are also involved in spindle pole focusing. In many systems spindle pole focusing in the absence of centrosomes has been shown to rely on the MT crosslinking kinesin Ncd [22–25] and the dynein complex [25–28]. The mechanism by which dynein contributes to acentriolar pole focusing is unclear, as its primary function in pole focusing in unperturbed cells is to transport K fibers towards centrosomes [9].
In Drosophila Ncd is nonessential, but it is crucial for acentriolar meiotic spindle assembly in oocytes [22]. To test if Ncd is essential for acentriolar spindle formation in Drosophila somatic cells we generated Sas-4 ncd double mutants. Surprisingly, these cells were often still capable of forming bipolar mitotic spindles, although many cells exhibited defective spindle phenotypes, ranging from poorly focused and misformed to highly abnormal spindles (Fig 7A, left panels). Quantification of spindle assembly showed that these cells had a reduced ability to form bipolar spindles compared to Sas-4 or ncd mutants alone, but were still able to form a spindle in ~26% of all mitotic cells (Fig 7B). As we often observed aMTOCs at the poles of the Sas-4 ncd cells that formed spindles (Fig 7A, arrow), we wondered whether these structures contributed to spindle formation in the absence of Ncd and centrioles. We therefore generated asl ncd double mutants to assess spindle formation in the absence of Ncd, centrioles and aMTOCs (Fig 7A, right panels). Quantification of the different spindle phenotypes showed that, strikingly, only 4% of mitotic asl ncd cells had formed a bipolar mitotic spindle, suggesting that aMTOCs indeed can ameliorate the severe spindle focusing defects observed in the absence of centrioles and Ncd. To confirm that it was the presence of aMTOCs, and not another potential intrinsic difference between Sas-4 and asl mutants, that enhanced spindle focusing in the absence of Ncd we assessed spindle formation in cnn; Sas-4 ncd triple mutants. In addition to lacking centrioles these cells also cannot form aMTOCs, as Cnn is required for aMTOC formation (Fig 4A and 4B). We could not observe any bipolar mitotic spindles in cnn; Sas-4 ncd triple mutant cells, confirming our hypothesis that aMTOCs likely aid in spindle pole focusing in the absence of centrioles and Ncd (Fig 7B).
How do aMTOCs aid spindle formation in the absence of centrioles and Ncd? We hypothesised that dynein might transport K fibers along the MTs nucleated by aMTOCs to help focus spindle poles. If true, then cells which lack centrioles and Ncd but have aMTOCs would not form bipolar spindles if dynein was also lacking. To test this possibility we generated Sas-4 ncd dhc (dynein heavy chain) triple mutants and assessed spindle formation in these cells. Intriguingly, almost none of these cells showed bipolar mitotic spindles (~1%, Fig 7B) and the few spindles we observed invariably exhibited unfocussed poles (Fig 7C). To rule out that this phenotype could originate from another, aMTOC independent, role of dynein we assessed spindle formation in dhc single mutants and Sas-4 dhc double mutants. Both of these had relatively normal spindle morphology, suggesting that dynein in conjunction with aMTOCs is required to ameliorate spindle formation in the absence of Ncd (Fig 7B).
These data suggest that dynein function in spindle focusing might be to transport kinetochore MTs towards MTOCs—either centriolar or acentriolar. If so, cells which have aMTOCs but lack Ncd and dynein should have a very similar phenotype to cells which have centrosomes but lack Ncd and dynein: in both cases K fibers could not be transported towards (either acentriolar or centriolar) MTOCs on the spindle poles. To test this hypothesis, we generated ncd dhc double mutants (which have centrosomes but lack Ncd and dynein). Only ~3% of these mitotic cells formed a bipolar spindle, which was very similar to the situation in Sas-4 ncd dhc triple mutants (which have aMTOCs but lack Ncd and dynein) (Fig 7B). We conclude that dynein is able to focus acentriolar mitotic spindles in an analogous fashion to spindle focusing in centrosomal cells, namely by transporting K fibers towards aMTOCs. In support of this hypothesis, dynein clearly localised to centrosomes (as previously observed [57]) and also aMTOCs on mitotic spindle poles in fixed (Fig 7D) and live larval brain cells (Fig 7E and S3 Movie), but failed to localise in asl mutants (S4 Fig).
Finally, we wanted to assess whether the ability of dynein and aMTOCs to promote spindle formation in acentriolar cells might be sufficiently robust to allow these acentriolar cells to proliferate even in the absence of Ncd. We compared acentriolar Sas-4 ncd tissue (which lacks Ncd, but has aMTOCs) with asl ncd tissue (which lacks both Ncd and aMTOCs). We noticed that while Sas-4 ncd brains were slightly smaller than WT brains they still looked relatively normal and had imaginal discs attached to them (Fig 8A and 8B). In comparison asl ncd brains and imaginal discs were much smaller than normal (Fig 8A and 8B), indicating a much stronger defect in cell proliferation [58]. This suggested that the severe proliferation defect in asl ncd larvae was rescued by aMTOC and dynein-mediated spindle focusing in Sas-4 ncd larvae. In support of this interpretation, Sas-4 ncd dhc larvae (which have aMTOCs but lack dynein) also had very small brains and discs (Fig 8A and 8B). We conclude that aMTOCs and dynein can cooperate to focus acentriolar spindle poles in cells lacking Ncd, and that this pathway substantially increases the efficiency of spindle assembly and thereby the ability of these cells to proliferate.
Here we confirm the presence of aMTOCs in acentriolar somatic cells and show that the pathway of aMTOC assembly in fly somatic brain cells is genetically very similar to the pathway of mitotic centrosome assembly. At centrosomes, Asl is present exclusively around the mother centriole [59–61], where it helps recruit Spd-2 and Cnn, which then assemble into a scaffold that spreads outwards around the mother centriole and which ultimately recruits the other mitotic PCM components [37–45]. Our data suggest that in ~50–60% of mitotic brain cells that lack centrioles, Asl can still recruit some Spd-2 and Cnn to form cytoplasmic scaffold structures that can then recruit other PCM components to form aMTOCs. Thus, while mother centrioles greatly increase the efficiency of mitotic PCM assembly, these three proteins can self-assemble into a mitotic scaffold structure even in their absence. The number and size of the aMTOCs formed, however, was very variable, suggesting that centrioles not only make mitotic PCM assembly more efficient, but they also serve to regulate the amount of PCM assembled and to ensure that only two MTOCs are normally formed during mitosis.
Structures similar to the aMTOCs described here have also been observed in other Drosophila tissues. In oocytes bundles of MTs were observed in the cytoplasm during acentriolar meiotic spindle formation [22]. No centrosomal markers such as γ-tubulin and CP60 have been observed clustering on meiotic spindle poles [22,62,63] and therefore it is unlikely that these MT bundles represent aMTOCs. In Drosophila embryos, injection of an antibody raised against Spd-2 lead to displacement of PCM from the centrosomes and thereby removed centrosomally nucleated MTs. In these embryos MT asters could be observed to form in the cytoplasm, which were then organised into mitotic spindles [4]. As centrioles are still present in these embryos it is unclear, whether these asters were nucleated from self-assembled aMTOCs as described here, or whether they are fragments of PCM originally nucleated around the mother centriole and then displaced from the centrosomes either through a lack of Cnn or by Spd-2 antibody injection [4]. Finally, a study has described the biochemical purification of cytoplasmic complexes containing Sas-4, Cnn, Asl and Plp (S-CAP complexes; [64]) from Drosophila embryonic extracts. Although S-CAP complexes have a composition similar to the aMTOCs described here, they require Sas-4 for their assembly [64]. The aMTOCs we describe here form in the absence of Sas-4, suggesting that they are unrelated.
Our identification of the factors required for aMTOC assembly in Drosophila allowed us to perform a detailed genetic analysis of the contribution of aMTOCs to different aspects of mitotic spindle assembly in acentriolar fly cells. Surprisingly, we find that aMTOCs do not detectably increase the rate of spindle assembly in somatic brain cells that lack centrioles. This suggests that, in contrast to meiotic spindle assembly in mouse oocytes [30], aMTOCs do not play a major role in generating spindle MTs in the absence of centrosomes, at least in this cell type.
Our data show, however, that aMTOCs can function as a major source of MT regrowth in cells that lack both centrosomal- and chromatin-mediated MT nucleation after the MT cytoskeleton has been depolymerised by cold treatment. Before cold treatment mst; asl mitotic cells (that lack the chromatin-, centrosome- and aMTOC-pathway of MT assembly) still had mitotic MT arrays, but these were significantly fewer than mst; Sas-4 cells (that lack the chromatin- and centrosome-pathways, but still have aMTOCs). We speculate that the MT arrays in mst; asl mitotic cells arise from augmin-mediated MT nucleation from pre-existing MTs, which can no longer happen once all MTs have been removed by cold-treatment.
Interestingly, we find that aMTOCs do play an important part in the mechanism of acentriolar spindle focusing by dynein. Dynein usually transports MTs emanating from kinetochores along astral MTs to the centrosome [6,9]. Our observations suggest that dynein can also transport MTs towards aMTOCs. Based on our results we propose an updated model for acentriolar spindle pole focusing by the minus end directed motors Ncd and dynein (Fig 9). In Drosophila cells with centrosomes, K fibers become crosslinked by Ncd, while dynein transports K fibers along centrosomal MTs towards the centrosome [9] (Fig 9A). When centrosomes and aMTOCs are lost, Ncd becomes essential for focusing acentriolar spindle poles as it can crosslink K fibers independently of centrosomes (Fig 9B). Therefore loss of Ncd leads to severely unfocused poles in acentriolar cells, as dynein, in contrast to Ncd, does not have the ability to statically crosslink MT minus ends (Fig 9C). If aMTOCs are present, however, then dynein can partially compensate for the loss of Ncd by focusing spindle poles through the transport of K fibers towards aMTOCs (Fig 9D).
This last point is important, as previous studies that reported a role for dynein in acentriolar pole focusing did not describe the molecular mechanism by which dynein is able to do this [25–28]. We suspect that structures similar to aMTOCs must have been present in these earlier studies for dynein to fulfil its role. These studies used Xenopus egg or cell extracts for their analyses. Interestingly in Verde et al. 1991 [28] the observation of a MT crosslinking material that accumulated at MT minus ends is described, and electron microscopy showed that this material had electron-dense properties similar to PCM [28]. Our finding that dynein needs to cooperate with aMTOCs to focus acentriolar spindles is an important addition to our understanding of the mechanism of spindle pole coalescence by molecular motors.
The following mutant alleles were used in this study: Sas-4s2214 [16], Sas-6c02901 [50,51], aslB46 (this study), Spd-2G20143 [43], cnnf04547 [45], cnnHK21 [44,65], ana2169 [66], mstLB20 [56], dhc6-10 [67] and ncd1 [68]. The following transgenic insertion lines were used: Asl::Asl-GFP [69], Ubq-GFP-Cnn [37], Ubq-Spd-2-GFP [43], ncd:: γ-tubulin37C-GFP [70], Ubq-Dlic-GFP (this study—a full length Dlic [dynein light intermediate chain—CG1938] cDNA was cloned into the Ubq-GFPNT Gateway vector [71]; Nina Peel, personal communication) and Jupiter-mCherry [72]. w67 was used as wild type control. An asl null allele was generated by inducing imprecise excision of the P-Element P{EPg}HP37249 located near the 5’UTR about 470bp upstream of the asl start codon. A 2.1kb deletion removing nearly 500bp upstream of the start codon and over half of the coding region was recovered (S4 Fig); the allele was named aslB46. In contrast to previously published alleles of asl [69] this new allele aslB46 produces no detectable N- or C-terminal Asl protein in western blotting or immunofluorescence experiments (S4 Fig). We used this new aslB46 mutant as it represents a true null mutation with no residual part of Asl being expressed, however we observed that aMTOC formation is also ablated in previously published aslmecD allele [69] (Cnn staining on spindle poles can be detected in 0.55% ± 0.55 of aslmecD cells).
Fixation and stainings of third instar larval brains was performed as previously described [53]. The MT regrowth assay in larval brains was performed as described in [56]. The following primary antibodies were used at a 1:500 dilution: Rabbit anti-Dlic (this study—raised against amino acids 1–293 of the Drosophila Dlic (CG1938) coding sequence; Nina Peel, personal communication), rat anti-Asl [73], rabbit anti-Asl (N-terminal and C-terminal) [37], guinea pig anti-Cnn [45], rabbit anti-Spd-2 [43], mouse anti- γ-tubulin (GTU88, Sigma), mouse anti-actin (Sigma-Aldrich), rabbit anti-PLP [53], rabbit anti-TACC [74], rabbit anti-Grip71WD [75], rabbit anti-Msps [76], rabbit anti-Aurora A [77], mouse monoclonal anti-α-tubulin (DM1α, Sigma-Aldrich) and rabbit anti-Histone H3 Phospho S10 (Upstate Biotechnology). Alexa488 anti-guinea pig and anti-mouse and Alexa568 anti-rabbit as secondary antibodies were used at a 1:1000 dilution (Molecular Probes, Life Technologies). Fixed preparations were examined using either a Zeiss LSM780 confocal microscope using a 63x/1.40 NA objective, or on a Zeiss Axioskop 2 microscope (Carl Zeiss, Ltd) with a CoolSNAP HQ camera (Photometrics), using a 63x/1.25 NA objective (Carl Zeiss, Ltd). Images were processed with Fiji [78] and adjusted to use the full range of pixel intensities.
Larval brains were stained with antibodies against α-tubulin to visualise spindles, Cnn to visualise aMTOCs and anti-Phospho-Histone H3 and were imaged on a Zeiss Axioskop 2 microscope. Only mitotic cells identified by Phospho-Histone H3 staining were scored, and all phenotypes were always quantified blindly, without knowing which genotype was being counted.
Brains were dissected from third instar larvae in PBS and the attached imaginal discs were removed. The brain was transferred to a drop of PBS on a clean coverslip and covered with a slide, which semi-squashed the brain. The coverslip was sealed with a drop of Voltalef 10S oil (VWR) to stop evaporation of the PBS and brains were analysed using a 63x/1.4 NA lens on a Perkin Elmer ERS Spinning Disk confocal system mounted in an inverted microscope (Axiovert 200M; Carl Zeiss Ltd) with a charge-coupled device camera (Orca ER; Hamamatsu) with Ultraview ERS software (Perkin Elmer), or using a 60x/1.4 NA lens on an Andor Revolution XD Spinning Disk confocal system mounted on an inverted Nikon (TE-2000E) with an Electron Multiplying charge-coupled device camera (iXon, Andor) and IQ2 software (Andor).
Third instar larval brains were dissected and the brain lobe volume was measured as previously described [79].
|
10.1371/journal.pcbi.1002915 | Collective States, Multistability and Transitional Behavior in Schooling Fish | The spontaneous emergence of pattern formation is ubiquitous in nature, often arising as a collective phenomenon from interactions among a large number of individual constituents or sub-systems. Understanding, and controlling, collective behavior is dependent on determining the low-level dynamical principles from which spatial and temporal patterns emerge; a key question is whether different group-level patterns result from all components of a system responding to the same external factor, individual components changing behavior but in a distributed self-organized way, or whether multiple collective states co-exist for the same individual behaviors. Using schooling fish (golden shiners, in groups of 30 to 300 fish) as a model system, we demonstrate that collective motion can be effectively mapped onto a set of order parameters describing the macroscopic group structure, revealing the existence of at least three dynamically-stable collective states; swarm, milling and polarized groups. Swarms are characterized by slow individual motion and a relatively dense, disordered structure. Increasing swim speed is associated with a transition to one of two locally-ordered states, milling or highly-mobile polarized groups. The stability of the discrete collective behaviors exhibited by a group depends on the number of group members. Transitions between states are influenced by both external (boundary-driven) and internal (changing motion of group members) factors. Whereas transitions between locally-disordered and locally-ordered group states are speed dependent, analysis of local and global properties of groups suggests that, congruent with theory, milling and polarized states co-exist in a bistable regime with transitions largely driven by perturbations. Our study allows us to relate theoretical and empirical understanding of animal group behavior and emphasizes dynamic changes in the structure of such groups.
| The patterns exhibited by moving animal groups like flocks of birds and schools of fish are typical of self-organizing systems in which global structural and dynamical properties arise from local interactions between individuals. Despite their apparent complexity, such systems can often be described, and understood, in terms of these emergent properties, rather than the detailed low-level description needed for depicting the individual dynamics. Here we show that schooling fish (in groups of 30 to 300 golden shiners) can be described in terms of the degree of alignment and degree of rotation among group members. We demonstrate that shiner schools exhibit three distinct behaviors: a swarm state with low speeds and little order; a strongly aligned state where the fish move with higher speeds; and a milling state, where each fish moves around the center of the group. Simulations have previously predicted this type of behavior, and we relate our findings to a well-known model of collective motion to help highlight similarities and differences between models and animal groups. Our results give insight into the regulation of group structure among animals, and also inform us generally about how global structures arise naturally from interactions among components of dynamical systems.
| Many animal groups display coordinated motion in which individuals exhibit attraction towards others and also a tendency to align their direction of travel with near-neighbors [1]. The functional complexity of such aggregates are thought to result from relatively local, self-organizing interactions among individuals that endow many groups, such as flocking birds and schooling fish, with the capacity to move, respond to threats and make decisions collectively [2]. Thus the way in which collective dynamics emerge from inter-individual social interactions likely has profound consequences on the selection pressures experienced by group living organisms.
Theoretical considerations of the self-structuring properties of groups have suggested that certain features of interaction among individuals may give rise to a relatively small number of specific collective states [3]–[7]. For example, models representing local repulsion, directional alignment and longer-range attraction among individuals [3], [5] predict that groups in which individuals follow these behavioral rules exhibit only three collective states; a swarm state in which individuals aggregate but are locally and globally disordered, a milling state in which individuals have a high degree of alignment with local neighbors but overall the group exhibits a rotating milling formation (or torus, in three-dimensional space) and a polarized state in which individuals tend to be aligned with each other over a long range and consequently the group experiences net movement. While these patterns have all been observed in nature among different organisms [8] and can evolve in simulations of simple predator and prey behaviour [9], experimental support for the existence of these states, and for any one system transitioning between them, has not previously been determined.
In addition to displaying commonly observed group structures these models also emphasize an important unifying feature—that collectively moving animal groups can be considered as a dynamical system in which multiple group states, or dynamically-stable states, exist, and that collective properties, such as the spatio-temporal configurations exhibited, may be robust to exactly how behavioral tendencies such as repulsion, alignment and attraction are mediated. Furthermore, under this scenario, animal groups are also predicted to exhibit ‘multistability’ whereby more than one collective state coexist for identical individual behavior, with groups transitioning relatively quickly between the alternate structural configurations. In the model of Couzin et al. [3], for example, the milling and polarized state co-exist over a region of parameter space (see Paley et al. [10] for a formal analysis of this bistable regime). The question of which collective behavior is adopted therefore depends on the initial configuration of the system, in this case the positions and orientations of individuals: if groups start in a relatively disordered state they tend to form a milling formation; but tend to remain in the polarized configuration if they begin sufficiently aligned with one another. Perturbations, such as disruption induced by predator attacks [1], or more generally, sources of intrinsic [11], [12] or extrinsic noise [13], can cause the system to leave its existing dynamically-stable state and enter an unstable transitional regime. Depending on the degree and type of perturbation the group may find itself either drawn back towards the previous state, or if perturbed sufficiently far, to be drawn into the alternative dynamically-stable state.
Thus even though the group behavior results from a large number of relatively local interactions, the group-level dynamics can be described using relatively few and simple lower-dimensional ‘order parameters’ that portray the collective dynamics, such as global polarization and the degree of collective rotation. This approach is familiar to us in physical systems where a multitude of different inter-molecular interactions result in only four fundamental states of matter; solid, liquid, gas and plasma. When properties such as density and energy are altered, a physical system can undergo ‘phase transitions’ between these states. Similarly, at a certain level of description, we can view animal groups as having the potential to exhibit abrupt changes in spatial or temporal patterns, and thus phase transition-like behavior (see Buhl et al. [14] for an experimental example of density-driven transitions in locust swarms).
However, biological systems are not in equilibrium. There are no conserved thermodynamic quantities, such as momentum and energy, and individual motion typically results not from thermal fluctuations or external forcing, but from individual self-propulsion and decision-making. Nevertheless, the concept that local interactions reduce to common non-equilibrium collective states is a key insight provided by computational modeling of collective animal behavior [3], [4], [15] which relates more generally to phase transition theory in non-equilibrium systems [16]. Additionally, despite individual motion likely being governed by a complex stochastic decision-making process based on the positions, movement and size of neighbors (and even hidden features like individual's state), two recent studies of schooling fish—Katz et al. [17] and Herbert-Read et al. [18]—demonstrate that interactions can be effectively reduced to local tendencies to be repelled from, or attracted towards, neighbors. Similar evidence exists for aggregating ducks [19] and swarming locusts [20].
Despite these advances, to date, no experimental study has quantified the dynamical states of collective motion exhibited by any species of group-living animal, nor determined whether the general predictions of existing models of collective behavior hold—notably, that groups will exhibit, and transition among, relatively few (dynamically-stable) states. Here we investigate the emergence of macroscopic collective states under highly controlled laboratory conditions using schooling fish (golden shiner, Notemigonus crysoleucas). This is a convenient model system for investigating collective behavior since individuals are relatively small (average length approximately 5 cm in our study), and naturally form highly cohesive schools in very shallow and still water [21]. Digital tracking of fish in a range of group sizes (from 30 to 300 fish) allows us to obtain detailed data regarding the individual positions and velocities of schooling fish over long periods of time. We use these data to analyze how group size and perturbations (driven both by inevitable contact with the boundary of the tank, but also by changes in motion by individuals within the group in the absence of boundary influence) affect group behavior and function to transition the group between alternative dynamical states.
Seven replicate experiments were conducted for group sizes 30, 70 and 150 fish, and three replicates were conducted for 300 fish (due to limitations in our capacity to house very large numbers of fish). Each replicate consisted of filming fish swimming in a large shallow tank (2.1 m×1.2 m, water depth 5 cm) for 56 minutes (at 30 frames per second). A similar approach has been taken previously [17], [18], [22], [23]. Individual fish were tracked following the methodology of Katz et al. [17] to obtain time series of the positions and velocities. These time series constitute the raw data from which we base our analyses.
To describe the collective structure of the fish shoals we use two order parameters, identical to previous categorization in simulation models [3], [11]. First, the polarization order parameter Op, which provides a measure of how aligned the individuals in a group are. It is defined as the absolute value of the mean individual heading,where is the unit direction of fish number i. Op takes values between 0 (no alignment on average) and 1 (all fish are aligned). Second, the rotation order parameter Or, which describes a group's degree of rotation about its center of mass. To define this measure we introduce the unit vector pointing from the shoal's center of mass towards fish i. The rotation order parameter Or is then defined by the mean (normalized) angular momentumwhich, by construction also takes values between 0 (no rotation) and 1 (strong rotation).
Time series of these two order parameters give us valuable information about the global structure of a group and how the structure changes during an experiment. However, important pieces of information are not captured, like the group density, the average individual speed, and how close a group swims to the tank boundary. For a fuller picture of the collective dynamics, we will also use these order parameters. Other group properties could also be measured.
Throughout our entire period of filming the fish were cohesive, and for all group sizes three dynamically-stable collective behaviors were observed [3]; the swarm (S), polarized (P) and milling (M) group state. Snapshots of these distinct patterns are shown in Fig. 1A for a group of 150 fish (Videos S1, S2, S3, S4 contain video extracts of all group sizes). As in [3], these separate modes of motion can be categorized by only two structural properties (order parameters) of the group—its polarization Op and its degree of collective rotation Or. Groups repeatedly transitioned between these collective states, as is evident in representative time series of the order parameters shown in Fig. 1B.
To demonstrate more clearly these dynamically-stable states, in Fig. 2 we consider the proportion of time groups spend in different regions of the two-dimensional phase space spanned by the order parameters Op and Or (red representing more time spent in a given region and blue the least time; black areas signify regions of the phase space not visited by groups in our experiments). While all three states of motion were manifest in all groups, there are also visible differences relating to increasing group size. For the smallest group size of 30 fish we see that the polarized group state predominates (high Op and low Or). Only rarely did groups of this size exhibit swarm behavior (low Op and low Or), and even less frequently did they adopt the rotating group state (low Op and high Or). The fluctuations in the order parameters are also most frequent for this group size (Fig. 1B). For a group size of 70 fish the frequency of transitions decreases and the collective states corresponding to the three dynamically-stable states become clearly distinguishable as ‘hotspots’: the polarized state is no longer dominant, with milling and swarm behavior also being common. As group size is increased further, to 150 and then 300 fish, groups spend most of their time milling, displaying fewer transitions into (and among) the polar state and swarm state. For all group sizes the milling state has an equal probability of rotating clockwise and counter-clockwise, i.e. groups did not exhibit a handedness (see Fig. S1).
A natural consequence of increasing the number of fish in the tank is that the mean density within the experimental arena becomes higher, and hence the effects of the tank boundary become more pronounced. To reveal whether the higher density of fish per tank area in larger groups could cause the increased stability of the milling state we performed an experiment (4 replicates) with 30 fish in a smaller tank (0.66×0.38 m), which corresponds to the mean density of 300 fish in the larger tank. The density plot of the order parameters is shown as an inset in the 30 fish density plot in Fig. 2, and reveals that confinement by the boundaries and higher mean density do not lead to increased time spent milling. However, the time spent in the polarized state was reduced; contact with the smaller tank caused this group size to exhibit more frequent transitions to the swarm state than when it was in the larger tank. At least for the 30 fish, and possibly as a general result, the presence of the boundary does not increase the stability of the milling state per se. Rather, the stability of milling is largely determined by the size of the group. Although the functional reason for milling is not yet known, it does, however, allow individuals to be locally polarized, which could be important for information transfer, while allowing the group to remain in a specific area. Swarm behavior allows the group to remain in an area but is locally disordered and this may be more susceptible to predation [1].
To gain further understanding of the relationship between group size and the stability of the different group structures we employed the canonical model of grouping of Couzin et al. [3], in which there are no boundary interactions. Exploring the collective behavior of simulated individuals (see Methods for simulation set up and details) we find that the model produces qualitatively similar results across the range of group size in our experiments (30, 70, 150 and 300 agents), with the polarized states being dominant for the smallest group size, and an increasing proportion of the group's time is spent in the milling state as group size increases (see Fig. 3).
Another aspect of group size is the self-regulation of density. Theoretically, when more members are added to a group of self-propelled particles, the density can either remain approximately constant, in which case the system is H-stable, or the density can increase, and the system is catastrophic [4]. In our experiments, the individuals within the group regulate their spacing such that density tends to remain stable regardless of group size. The mean area occupied by the fish grows approximately linearly with group size and the packing fraction (and density) remains nearly constant (See Fig. S2). This regulatory behavior places the fish in the category of H-stable systems.
To quantify the relation between group size and collective state we need to explicitly define the different states. Given the relatively clear demarcation of states revealed by our data in Fig. 2 we employed a simple approach in which we discretize the phase space in terms of the order parameters. The specific range of values were motivated by the high-density regions observed in Fig. 2. We thus define that the school is in: the polar state (P) when Op>0.65 and Or<0.35; the milling state (M) when Op<0.35 and Or>0.65; and the swarm state (S) when Op<0.35 and Or<0.35. Outside these ranges we define the system to be in a transitional regime (T). On average, therefore, each region is characterized by the dominant dynamical state of the fish school within that particular region. The regions defining the dynamical states are overlaid the density plots in Fig. 2 (the qualitative nature of our results does not depend on the precise nature of how these regions are defined, see Fig. S3).
As shown in Fig. 1B, groups frequently transitioned between the three collective states. A transition is considered completed if the group moves from one of the three states to another. By this definition, a school can move from one of the dynamical states, into the transition region, and then back to its previous state, without having undergone a transition. We quantify, statistically, the transitions between states and investigate how the transition patterns depend on the group size.
As we previously saw in Fig. 2, the proportion of time spent in the polarized state decreases strongly with group size from 0.57 to 0.26, 0.18 and 0.05 (30 to 300 fish shoals, respectively, light blue columns in Fig. 4A). Likewise, the proportion of time in the milling state increases with group size from 0.03 to 0.18, 0.36 and 0.45 (yellow columns Fig. 4A). The group size has however little effect on the proportion of time spent in the swarm state or in the transition region, which varies between 0.09 and 0.11 (dark blue columns Fig. 4A), and 0.31 and 0.44 (brown columns in Fig. 4A), respectively. The fraction of transitions between states, as illustrated by Fig. 4B, exhibit little variation between the group sizes. The only visible trend is a small increase in the number of transitions between the milling state and the swarm state (see also Fig. S4 for alternative graphics).
To complement this picture it is important to note two (inter-related) features that do change with group size: Firstly the rate at which groups exhibit transitions decreases as a function of group size (from 2.0 transitions/min for 30 fish, to 1.4, 1.2, and 0.8 transitions/min for 70, 150, and 300 fish respectively); Secondly, the stability of the milling state increases as a function of group size (the longest time a group of 30 spent milling was 17 s, this increased to 110 s, 708 s, and 1245 s for group sizes 70, 150 and 300, respectively. See Fig. 4C for rank plots of time spent in a state before transitioning). This means that the large proportion of time the 30 fish spent in the polarized state is an accumulated effect of many visits into the state, while the proportion of time the 300 fish spent in the milling state is greatly affected by the milling state being more stable for larger groups.
A further feature of the transitional behavior of groups is that transitions from one dynamical state to another only accounted for 47% (n = 1943) of the total number of visits into the transition zone (n = 4119), counting only visits lasting longer than 1 s. This demonstrates that the schools experience frequent perturbations, of which some result in transitions to another state (Fig. 4B), and the others back to the preceding state.
From observing the schooling behavior it appears that perturbations to the group act as triggers to transitions between collective states (see Videos S1, S2, S3, S4) and we can identify two main sources for fluctuations that result in transitions; interactions with the tank wall (boundary effects) and fluctuations due to the intrinsically noisy nature of individual motion. We note that these processes are not mutually exclusive.
In order to reveal more clearly the role of boundary effects on state transitions we use our extensive time series to characterize the typical nature of transitions and relate these to whether the group tends to be relatively close to, or far from, the boundary. We present data for 150 fish in Fig. 5 (for other group sizes see Fig. S5). In Fig. 5A the arrows represent the average trajectories that groups take through Op and Or space when transitions occur and, unlike Fig. 2, the density plot now depicts the distance db from the center of mass of the group to the closest point at the tank boundary; red colors represent a relatively large distance and blue colors relative proximity to the boundary. A more detailed view of the transition dynamics is presented in Figs. 5B–D. Here, for each of the transitions from polar to milling, polar to swarm and milling to swarm state, the average trajectories are plotted as a velocity field in the Op and Or phase space overlaid on the density plot showing the distribution of trajectories (for the reverse transitions and other group sizes see Figs. S6 and S7).
The data shown in Fig. 5 verify that transitions happen both close to the wall of the tank, and in the center of the tank. For large parts of the transition region schools are relatively close to the boundary, such as from the polarized state to the swarm state, where most transitions take place (Fig. 5C). Although not as clear, the transitions between the milling state and the swarm state are also characterized by being, on average, closer to the wall (Fig. 5D). The exception is for transitions that occur with high values of the order parameters, that is, between the polarized state and the milling state (Fig. 5B), and vice versa. These tend to occur both close to and also away from the wall. This indicates, as evidenced by the video footage (Videos S1, S2, S3, S4), that both boundary and other triggering mechanisms are important for inducing transitions between collective states.
When in the milling state, interactions with the boundary can result in a local increase in density near the wall, due to the inherently constrained nature of motion when abutting the boundary. This can cause the mill to transition into a polarized state as shown in Fig. 6A. Another way in which the mill can break down is due to the action of individuals at the group edge; if fish turn or move away from the edge of the group this can seed the unraveling of the milling formation into a polarized state. Conversely, when in a polarized state individuals at the front of the group can turn towards the main mass resulting in a perturbation that prompts the group to turn, potentially initiating the mill formation. This last example is evidenced in Fig. 6B and demonstrates that the milling formation can emerge as a group effect from the individual interactions—without direct interaction with the tank boundary.
In the case of polarized groups the boundary also has inevitable consequences on transitions; a polarized group may swim directly towards a wall, or corner, of the tank. Individuals reaching the wall slow down and tend to become disordered (unaligned) and the group can transition into the swarm state. That this mechanism of transition is dominating is demonstrated both by the average transition path from polarized to swarm in Fig. 5A, which crosses a region where db is small, as well as the short average time groups spend in the polarized state before transitioning, as shown in Fig. 4C.
For tractability many previous models of animal grouping (including that of Couzin et al. [3]) have assumed that individuals move at constant speed and social response is represented by adjusting direction of travel in response to the positions and/or orientations of near neighbors. Recently, however, two experimental studies on schooling fish, Katz et al. [17], involving golden shiners (Notemigonus crysoleucas—the species used here), and Herbert-read et al. [18], involving mosquitofish (Gambusia holbrooki), have highlighted the importance of speed regulation to collective behaviors. In the former study it was found that individual social interactions can be approximated qualitatively by pairwise interactions that are functions of the position and speed of each individual. While the spatial nature of these interactions was found to be relatively independent of individual speed, the magnitude, or strength, of response to neighbors decreased greatly as individual speed decreased. Also Viscido et al. [24] found a positive correlation between average group speed and polarity for shoals of 4 and 8 giant danios (Devario aequipinnatus). This suggests that there is an important relationship between individual speed and the degree to which individuals coordinate their motion with neighbors, a relationship that is not captured in many models of collective motion [3], [15], [25], [26].
Examining the relationship between the mean speed of individuals in the group and the ‘packing fraction’ (a measure of the density of individuals within the group) and the order parameters Op and Or, we observe that low speed is associated with the group being relatively dense and both locally- and globally-disordered (the swarm state). The two locally-ordered (milling and polarized) states are characterized by higher mean speed and a decreased packing fraction (see Fig. 7A for group size of 150 fish; this relationship is common among all group sizes as shown in Fig. S8). Consequently the relationship between density and order is the opposite of that predicted by the most studied models of grouping behavior, notably the Vicsek model [15]; although we note that such simple models have been extremely useful in developing understanding of group dynamics for other animal aggregates, such as locusts [14], and other species of schooling fish [23].
From our data we cannot distinguish between two, not mutually exclusive, hypotheses regarding the causal relationship between speed and order; does decreasing speed induce local disorder through weakened social interactions, or does perception of local disorder reduce an individual's speed? Since golden shiners do not appear to explicitly respond to the body orientation of neighbors, and rather respond more-or-less exclusively to individuals' positions in space [17], increasing speed likely increases local order. However, a dense, slow moving and disordered group is also likely to further reduce individual speed (not least through increased risk of collision) - thus both causal relationships likely co-exist.
To demonstrate the plausibility (and indeed, generality) of speed-induced transitions, we return to the model employed above, from [3]. We verify that changing individual speed does result in qualitatively the same transitional behavior seen here; swarm behavior for relatively low speed and bistable milling and parallel group motion as individual speed increases. This result holds up to 150 agents. For groups of 300 agents the milling state is dominant and no instances of the polar state are found. (see Fig. 8 for results from simulations with 150 agents and Fig. S9 for remaining group sizes. Simulation details are found in Methods).
In order to deepen our understanding of the local dynamics we also quantified the relationship between the speed of an individual and the degree of order in its immediate vicinity (within a radius distance of 15.5 cm). In Fig. 7B we show the resulting relationship as a contour plot for groups of 150 fish (see Fig. S10 for corresponding plots of 30, 70 and 300 fish, and for different radial proximity distances). A strong association is evident between individual speed and local order. Assuming an unknown causal direction in the relation between the speed and the local order, as discussed above, there are two ways we can average the contour plots; either over all values of the speed for each value of the local order, or vice versa, for each value of the speed we average over the local order. The results of both procedures, for all group sizes, are overlaid on the contour plot in Fig. 7B. While the two approaches yield disparate curves, they both demonstrate a similar relationship between individual speed and local order. Although the fish have to align at higher speeds to maintain group cohesion, it is unclear why they should become disordered at low speeds. There is also a much greater variance in local order at low speeds, demonstrating a wide degree of flexibility when individual speeds are low. Interestingly, across the group sizes the two sets of curves are close to identical. This suggests that, from the perspective of a focal individual, it may simply adopt the same local rules regardless of group size (consistent with the findings of Katz et al. [17] for groups of 10 and 30 fish).
Contrary to the polarization order parameter, the rotation order parameter has little meaning on a purely local scale. Rather we compute the rotation order parameter separately for a series of shells placed around the center of mass of the group, as illustrated in Fig. S11. This allows us to obtain a well-defined measure that provides insight into the structural organization of the milling state. As can be seen in Figs. 7C and 7D this state is characterized by a center with low speed and low degree of structure that contributes little to the milling state, while as we move towards the edge of the group the speed and impact from each shell on the milling state increases. In Fig. 7D the curves are almost identical for all group sizes suggesting that, again, scaling the size of the group has little effect on the local structural signature of this collective state.
These results demonstrate that the ordered polarized and milling states are locally near identical from the perspective of a focal individual, regardless of group size. These data also support the prediction of a multi-stable locally-ordered regime in which the group can transition back and forth between the polarized and milling state through stochastic and boundary-induced effects.
Despite the multitude of local interactions that result in coordinated group motion we demonstrate that schooling golden shiners predominantly exist in three ‘fundamental’ dynamically-stable states of the underlying dynamics: swarm, milling and polarized motion. We establish that group states, and transitional behavior, can be represented in low-dimensional space, a projection that allows us to see the path taken by groups between the three dynamically-stable states as well as to relate the collective states exhibited to properties such as group size, individual speed and perturbations to the group. We note that it is possible that further collective states may be found within the classified dynamically stable regimes described here, but the present states are highly consistent with the theoretical predictions of three regimes
A key question in the study of collective behavior is whether different group-level patterns result from all individuals responding to the same external factor [27], or individuals changing behavior [1], or whether multiple dynamically-stable collective states co-exist for the same individual behaviors [3]. Our results provide evidence for the importance of the two latter processes in the behavior of schooling fish: transitions from the swarm, to the milling or parallel group states (and vice versa) involve a social feedback whereby individuals adjust behavior—in this case their speed—in response to prevailing local conditions. Low average speeds among group members correspond to them occupying the dense, disordered swarm regime.
Higher speeds correspond to higher local order (alignment among group members) and groups existing in either the milling or polarized state. Transitions between these states occur with negligible, or no, change in local density, order or speed; instead perturbations such as collisions with the boundary, or (seemingly stochastic) fluctuations in motion at the group edge (in the case of milling to polarized state transitions) or front (in the case of polarized to milling transitions) result in the group leaving one dynamically-stable state, and either then returning to that state, or transitioning to the alternative locally-ordered regime. Thus the milling and polarized states appear to be bistable; the state exhibited by the group effectively being dependent on starting conditions and/or the nature of perturbations, as well as the group size. Theoretically [3] and experimentally (analysis of shoals of 30 fish in small tank), milling states are seen to be less stable for small groups, when controlling for boundary condition effects. It is likely that the relationship we found between speed and local order is a generic feature of mobile groups with local interactions. Furthermore, qualitatively similar features have been observed in small groups (4 and 8 fish) of the giant danios [24].
A key challenge for animal behavior in this, and future, decades is to understand how the microscopic mechanisms of interactions among molecules, physiological systems and neural circuits result in behavior at higher levels of organization. Whereas we focused on collective behavior resulting from interactions among individual organisms, the general approach adopted shares commonalities with approaches that have successfully been used to characterize the dynamical properties of gene interaction networks [28], neuronal circuits [29], how locomotion is coordinated among limbs, each of which has many degrees of freedom [30], and how the behavior of individual organisms (such as Caenorhabditis elegans), despite apparent complexity, can be deconstructed into a discrete number of low-dimensional behavioral dynamically-stable states (see Stephens et al. [31]). We suggest that development, and adoption of, such techniques in the behavioral sciences could facilitate the advent of increasingly integrative and quantitative insights.
Our work demonstrates that such an approach to data collection and analysis can reveal underlying simplicity in the dynamical properties of collective behavior in groups. Collective behavioral states appear to result from both behavioral feedback processes whereby individuals both adopt, and influence, the behavior of near neighbors and also as multi-stable regimes in which individual behavior does not change, but rather perturbations induce relatively abrupt transitions between alternate and co-existing dynamically-stable behavioral states. Prey groups have been observed to switch states upon detecting a predator [32] and risk can be dependent on these states [9], [33]. Whether the mechanisms for switching between states as identified here are somehow themselves adaptive would be an interesting question to address in future work.
The experimental setup, the automated tracking procedure and the methods for constructing detailed trajectory data were the same as described in Yael et al. [17] and details are found there. The data used for the analyses in this work are time series of positions and velocities of the individual fish. The tracking accuracy varied with the structure of the groups. As we show in Fig. S12, ordered groups were more precisely tracked, while dense and disordered groups were more prone to tracking errors. On average the percentage of frames with tracking accuracy above 80% were 88% for 30 fish, 91% for 70 fish, 80% for 150 fish and 71% for 300 fish. Since our focus in this paper does not rely on us maintaining identities for long periods of time there is ample data from which to calculate global patterns and distributions of properties such as speed.
We calculated the polarization order parameter Op and the rotation order parameter Or (see definitions in Results and Discussion) for each frame and smoothed the resulting time series using a moving average with a span of 30 frames (corresponding to 1 second). In Fig. 7B we used a definition of the polarization order parameter that is restricted to the local neighborhood around a focal individual. Similarly, we used a radial version of the rotation order parameter in Fig. 7D, in which only fish inside a shell of given radius surrounding the full shoal's center of mass were included.
All Or-Op density plots were made by dividing up the phase space into 30 times 30 bins and counting the number of values falling into the respective bins (Figs. 2, 3 and 5B–C) or calculating the average value in the bins (Figs. 5A and 7A). Only bins with counts above a certain threshold were included (100 counts in Figs. 2 and 7A and 20 counts in Fig. 5). Before plotting, the bin values were interpolated over a finer mesh of 300 times 300 points.
Deciding exact thresholds for when a shoal is in a certain state or not is hardly possible, even though it is easy to approximately mark out the regions in the Or-Op phase space that corresponds to the swarming, polar and milling states. Since the analyses we did were not critically dependent on whether precise demarcations could be made, we used a heuristic approach and—motivated by the high density regions observed in Fig. 2, as well as visual verification from the videos—defined the dynamic states as follows: polar state (P) when Op>0.65 and Or<0.35; milling state (M) when Op<0.35 and Or>0.65; and swarm state (S) when Op<0.35 and Or<0.35. Outside these ranges we defined the system to be in a transitional state (T).
To calculate the packing fraction of a group we first used an alpha-shape algorithm [34] to measure the area spanned by the group. Dividing the number of fish by the measured area and then multiplying by the average area of a fish body (40 times 5 pixels) produced the packing fraction value.
To calculate the average transition paths in Fig. 5A we first interpolated all transition time series to have the same length. Then we averaged the interpolated transition paths between one state and another. The vector plots in FigS. 5B–D were constructed by first differentiating each of the transition time series to create a velocity vector field, which we then coarse-grained by dividing the Or-Op phase space into 30 times 30 bins and averaging over the vector field in each bin.
Simulations were performed using the constant-speed agent-based model described in Couzin et al. [3] with 30, 70, 150 and 300 individuals. In this model the individuals move with constant speed and interact with each other through three types of interactions: repulsion, alignment of orientation and attraction. Centered on each individual are three spherical non-overlapping behavioral zones; zone of repulsion, zone of orientation and zone of attraction. The distribution of neighbors across these three zones is what decides the desired heading of an individual. For a detailed description of the model algorithm, see [3]. In our simulations we varied the speed from 0.1 to 4.1 unit lengths per unit time in increments of 0.1 and performed 500 replicates for each value of the speed. Each simulation was run for 2500 time steps and the order parameters in the final simulation step recorded. The remaining model parameters remained fixed throughout the simulations and were: zone of repulsion 1; zone of orientation 3; zone of attraction 15; field of perception 270 degrees; turning rate 60 degrees; error 0.2 radians; time step increment 0.1. Note that the simulations are not parametrized to fit the schools of fish. Rather, we use the simulations to display a generic quality of self-propelled particle models that aligns with experimental observations.
The frequency of the milling state rotating in a clockwise or counter-clockwise direction was analysed using a quasi-poisson distributed Generalised Linear Mixed Model (GLMM). Direction (a within-subject fixed factor) and group size (between-subject covariate) were the explanatory variables along with their interaction term, and shoal identity the random variable. The time spent in each state as a proportion of the total trial time was calculated for each group of fish, and then was analyzed as a function of group size using quasi-binomial Generalised Linear Models (GLM). Each state was analyzed separately, as well as the time spent in transition between states. Quasi-negative binomial GLMMs were used to analyse the frequency of transitions from each state (swarm, milling and polarised) to a different state (the ‘to’ state, a within-subject fixed factor), again with group size as an additional explanatory variable and shoal identity the random variable. These models were run separately for each ‘from’ state and included the ‘to’ state×group size interaction. The persistence of each visit to a state before transitioning into a different state (i.e. its stability) was analysed using quasi-negative binomial GLMMs. As a visit to a state within one shoal was not independent from the duration of other visits within that shoal, the analysis was carried out separately for each state. Group size was used as an explanatory variable and shoal identity as the random variable. The parameter theta for the quasi-negative binomial GLMMs was estimated from running negative-binomial GLMs without the random variable first, which gives an estimate for theta. Non-significant interaction terms were removed before testing main effects, but with the statistics given for main effects including any other main effects regardless of their significance. All statistical tests were carried out using R 2.14.2.
|
10.1371/journal.pcbi.1005460 | Highly accessible AU-rich regions in 3’ untranslated regions are hotspots for binding of regulatory factors | Post-transcriptional regulation is regarded as one of the major processes involved in the regulation of gene expression. It is mainly performed by RNA binding proteins and microRNAs, which target RNAs and typically affect their stability. Recent efforts from the scientific community have aimed at understanding post-transcriptional regulation at a global scale by using high-throughput sequencing techniques such as cross-linking and immunoprecipitation (CLIP), which facilitates identification of binding sites of these regulatory factors. However, the diversity in the experimental procedures and bioinformatics analyses has hindered the integration of multiple datasets and thus limited the development of an integrated view of post-transcriptional regulation. In this work, we have performed a comprehensive analysis of 107 CLIP datasets from 49 different RBPs in HEK293 cells to shed light on the complex interactions that govern post-transcriptional regulation. By developing a more stringent CLIP analysis pipeline we have discovered the existence of conserved regulatory AU-rich regions in the 3’UTRs where miRNAs and RBPs that regulate several processes such as polyadenylation or mRNA stability bind. Analogous to promoters, many factors have binding sites overlapping or in close proximity in these hotspots and hence the regulation of the mRNA may depend on their relative concentrations. This hypothesis is supported by RBP knockdown experiments that alter the relative concentration of RBPs in the cell. Upon AGO2 knockdown (KD), transcripts containing “free” target sites show increased expression levels compared to those containing target sites in hotspots, which suggests that target sites within hotspots are less available for miRNAs to bind. Interestingly, these hotspots appear enriched in genes with regulatory functions such as DNA binding and RNA binding. Taken together, our results suggest that hotspots are functional regulatory elements that define an extra layer of regulation of post-transcriptional regulatory networks.
| All the cells in a given organism contain the same genome, yet their phenotype can be very diverse. The vast majority of this diversity arises from the differences in the expression of genes and proteins in them. One of the main mechanisms involved in controlling the protein and mRNA repertoire in cells is post-transcriptional regulation. The recent development of high-throughput sequencing techniques gives us now an unprecedented opportunity to investigate how post-transcriptional regulation works and which are the elements involved in defining the final set of mRNAs and proteins inside cells. In this work, we have performed a comprehensive computational analysis of several post-transcriptional regulators in a commonly used human cell line in order to understand which factors are involved in post-transcriptional regulation and how they coordinate their function. The results of our analysis show that this process is orchestrated around small regions in the mRNAs where many regulators bind and may compete with each other to regulate the mRNAs. The investigation and characterization of these regions gives us insight into the underlying combinatorial control that causes gene expression to differ across cell types and in diseases.
| Post-transcriptional regulation is the set of mechanisms and processes that control gene expression at the RNA level and affect the properties and the amount of RNA transcribed and translated into proteins. This regulation is performed mainly by RNA binding proteins (RBPs) and miRNAs, which primarily target the 3’ untranslated region (3’UTR) of transcripts. In animals, miRNAs usually function by promoting translational inhibition and decay of mRNAs [1]. In contrast, RBPs have a wider range of functions and are often involved in multiple post-transcriptional processes.
Although some miRNAs are predicted to target thousands of mRNAs [2], not all of their predicted targets are down-regulated upon miRNA transfection [3], and many seem to be regulated only in certain cellular contexts or under stress conditions [4]. In some cases, miRNAs have even been found to promote translational activation [5] or increase mRNA levels. All these different complex functions suggest that miRNAs and RBPs take part in combinatorial regulation, where the combination of factors that bind to an RNA determines its fate.
In humans, more than 1000 RBPs [6,7] and ~2500 miRNAs [8] are involved in this complex regulation. MiRNAs are known to act cooperatively to down-regulate mRNAs when bound close in space [3,9,10]. RBPs can compete for binding to AU-rich elements (AREs) [11] or cooperate in mRNA regulation [12,13]. Moreover, they can compete and collaborate with miRNAs, or even perform opposite functions in different contexts [14]. For instance, AUF1 has been found to both compete with AGO2 for binding to the mRNA and cooperate with it [15]. Similarly, HuR has been found to compete with miRNAs for binding [16,17] but also to cooperate with miRNAs both to stabilize and degrade target mRNAs [18,19].
Previous studies that aimed to decipher the interactions between miRNAs and RBPs have been focusing either on single genes [20–24] or on the interactions between a single RBP and miRNAs using cross-linking immunoprecipitation coupled to high-throughput sequencing (CLIP-seq) data [13,16,17], and only recently these interactions have been explored at a transcriptome-wide scale [25,26]. However, the differences across CLIP protocols [27] and existing computational tools used to analyze the resulting data [28,29] complicates their integration and comparison, which is required for obtaining a picture of their interactions on a global scale. Additionally, CLIP methods typically include a significant amount of noise that can lead to the identification of artifactual binding sites if stringent filtering criteria are not used [30,31].
In this work, we reanalyze a comprehensive collection of CLIP experiments from HEK293 cells to shed light on the complex interactions between RBPs and miRNAs. By using a stringent processing pipeline and integrating multiple datasets, we show that post-transcriptional regulators bind preferentially in specific regions of 3’UTRs, i.e. hotspots, where we find enrichment of both RBP and miRNA target sites. These hotspots, rather than being experimental artifacts as previously thought [26,30], share characteristics typical of other regulatory regions: high sequence conservation, accessibility, and enrichment in AU-rich elements (AREs). Additionally, our results suggest that they might function by favoring competition among regulators. Upon AGO2 knockdown (KD), we observe that the changes in expression level of transcripts that harbor miRNA target sites within hotspots are significantly different from those of transcripts that contain miRNA target sites outside them, which suggests that RBPs binding in hotspots prevents RISC association. Similar changes are observed when an RBP is knocked down, highlighting that competition for binding may occur not only between miRNAs and RBPs but also among RBPs. Interestingly, these hotspots are enriched in genes with roles in transcriptional and post-transcriptional regulation.
Taken together, these results suggest that post-transcriptional regulation is focused in hotspots within 3’UTRs where several regulators can bind in close proximity and modulate the functions of the regulatory network both at transcriptional and post-transcriptional level.
To investigate the complex interactions between RBPs and miRNAs on a transcriptome-wide scale, we reanalyzed previously published CLIP data for 49 RBPs in HEK293 cells, which correspond to a total of 107 experiments. The list of all the RBPs analyzed along with a brief description from STRING database [32] can be found in S1 Table. An analysis of the functions of these proteins using GO terms revealed that many are involved in similar processes, especially in post-transcriptional regulation of gene expression (Fig 1a; S2 Table).
In order to obtain a set of significant RBP binding sites that can be compared across experiments, all the datasets were reanalyzed using the same pipeline. This pipeline, which has been used previously to identify binding sites from PAR-CLIP and iCLIP datasets [33,34], is focused on improving read preprocessing using custom scripts and mapping using BWA-PSSM [35], which allows to model the high mutation rate observed in CLIP data [36,37], improving both sensitivity and specificity. Moreover, duplicate removal, which usually is performed to discard reads that map to the same location, was done on the sequence of the reads after low quality and adapter removal at the 3’end. Given the high mutation and indel rate observed in CLIP and PAR-CLIP data [36,37], this approach was used in order to keep as many reads that represent RBP binding sites as possible. As expected, a high percentage of the reads confidently mapped from the CLIP datasets contained one or more mutations, especially T to C conversions in PAR-CLIP datasets, which are indicative of crosslinking events (S1 Fig). After mapping, significant clusters were identified using Pyicos [38] (see Materials and Methods and S1 Appendix). For many of the RBPs analyzed, most of the clusters were located on 3’UTRs (Fig 1b), which is consistent with the observation that many of these RBPs are involved in post-transcriptional regulation (S2 Table). On average, each mRNA expressed in HEK293 is bound by 14 different RBPs in its 3’UTR, and 20% of them are bound by more than half of the RBPs analyzed. A detailed analysis of the distribution of RBP binding sites along 3’UTRs showed that most of them bind preferentially towards the 3’UTR edges, where there is also a higher density of miRNA target sites (Fig 1c).To understand if the observed positional bias indicates that the proteins bind in the same regions and therefore interact or compete with each other, we calculated spatial correlations between their cluster enrichments, calculated as the log2 ratio of CLIP over RNA-seq reads (see Materials and Methods). For each pair of proteins, we calculated the Pearson correlation between CLIP enrichment values for each distance between -200 and 200 nt in each 3’UTR. The correlations calculated for these values were averaged over all 3’UTRs, giving us an average spatial correlation profile for the two RBPs (S2 Fig). These correlations were then compared to those obtained from shuffling the clusters to calculate their z-scores at each position (Fig 2a). In each row of Fig 2a, the z-scores of these spatial correlations are shown for each pair of RBPs. The higher the z-score, the more significant the correlation between two RBPs in a particular position is. In 746 out of 1084 pairwise combinations of RBPs (excluding pairing of an RBP with itself), we observed that the highest positional correlation z-score was in the +/- 9 nt interval. This result reflects that for 69% of all RBP pairs, the most significant positional correlation was observed when the clusters of the two RBPs overlap in the same 3’UTR (Fig 2a). If we consider only the positional correlations around AGO2 binding sites, this percentage increases to 89%. As seen in Fig 2b, the strongest positional correlations around AGO2 binding sites were found for other AGO proteins and TNRC6 proteins, which are also part of the RISC complex. Other RBPs display weaker positional correlations with AGO2 although they are highly significant compared to the random expectation.
The previous analyses demonstrated that RBPs exhibit distinctive binding preferences around AGO2 binding sites. To further evaluate the correlations between RBPs and AGO2, we analyzed the enrichment distribution of their binding sites around predicted target sites of miRNAs expressed in HEK293 cells. 75% of all the target sites from expressed miRNAs in HEK293 were bound by one or more RBPs (S3 Table). As expected, the binding of AGO2 and the other AGO proteins peaked on top of predicted miRNA target sites, especially on target sites of miRNAs highly expressed in HEK293 cells (hisites) (Fig 3a). To identify proteins that were significantly enriched around miRNA target sites, we computed an empiric p-value by comparing their enrichment around target sites with that around random target sites. Interestingly, a total of 26 out of 47 RBPs analyzed showed a significant enrichment around hisites (empiric p-value < 0.05; Fig 3b and S3 Fig). From these, 13 are known or predicted to interact with RISC according to STRING database [32] (S4 Fig). In some cases, the enrichment profile of the RBPs peaked on miRNA target sites. In other cases, the enrichment increased across the miRNA target site, e.g. WDR33, or was less dependent on miRNA expression, such as in the case of HuR and EWSR1 (Fig 3a and S3 Fig). Notably, the enrichment distribution around hisites (Fig 3b) in many cases resembles the positional correlation with AGO2 described before (Fig 2b).
To gain insight into why RBPs have a strong positional bias around hisites, we calculated the correlation between RBP enrichment on target sites and miRNA expression level, i.e. the sum of expressions of all the miRNA targeting them. Our results show a significant correlation between RBP enrichment on target sites and miRNA expression not only for AGO proteins (AGO1-4), but also for the proteins from the polyadenylation complex CF-Im68 and CPSF-73 and FUS (Fig 3c and S5 Fig).
For each RBP we also calculated the percentage of its clusters that overlap hisites. This number ranges from less than 10% to more than 40% (Fig 3d, bars), and usually is less than 5% (Fig 3d, dots). Even though the overlap was small, in most cases the association between clusters and hisites was significantly higher than expected by chance (permutation test p-value < 0.01; colored bars and dots). Interestingly, the combined set of all RBPs overlapped more than 77% of hisites (75% excluding AGO and TNRC6 proteins), which suggests that the cumulative effects of all RBPs could have a crucial role modulating miRNA function.
Several components of the polyadenylation complex, including the cleavage factors CF-Im25, CF-Im59 and CF-Im68, showed a strong enrichment on hisites (Fig 3b) and an overlap with hisites comparable to that of AGO proteins (Fig 3d). We analyzed their enrichment on hisites according to their location on the 3’UTRs. Each 3’UTR was divided in three equally sized regions and hisites in these regions were classified according to their location as proximal, medial and distal. Our hypothesis was that if this binding was specific, RBPs from the polyadenylation complex should be more enriched on target sites at the end of the 3’UTR. As expected, we observed a strong enrichment on distal hisites for most of these proteins (Fig 4). Two distinct types of profiles were predominant. The cleavage factors CPSF-160 and Fip1 peaked exactly on the target site region in a similar way to AGO2 (Fig 3a), although their enrichment on hisites was much lower. Alternatively, CstF-64 and WDR33 displayed a change from low enrichment before the target site to high enrichment after the target site. These two different types of binding profiles around hisites may reflect the different kinds of interactions of these RBPs with miRNAs or the miRNA machinery.
To evaluate the impact of all RBPs together, we looked at their binding in non-overlapping 50 nt windows across 3’UTRs. In order to discard weak binding sites, we considered only significant clusters with a positive enrichment relative to RNA-seq data. The list of all the windows containing RBP binding sites mapped on them can be found in the S2 Appendix. We observed that the windows with more than 3 RBPs binding were more frequent than expected if RBPs would bind independently (S6 Fig). We also observed that windows containing more RBPs displayed a stronger positional bias towards 3’UTR edges, similar to that of miRNA target sites (S7 Fig).
One of the main characteristics of hotspots is that they are more accessible than windows with less RBPs binding on them (Fig 5a top; spearman correlation coefficient rho 0.14, 0.14 and 0.15 for windows overlapping expressed miRNAs, non-expressed miRNAs and not overlapping miRNAs respectively; p-value < 2.2e-16 in all cases). However, if RBP hotspots are functional regulatory elements in 3’UTRs, we expect them to have some features common to other known functional elements such as a higher conservation relative to its surrounding area. For each of the 50nt windows we measured their average conservation using phyloP scores [39] and their SNP density. As seen in Fig 5a, we observed a significant correlation between the amount of RBPs binding in a window and its average conservation using phyloP scores (rho 0.19, 0.20 and 0.20 for windows overlapping expressed miRNAs, non-expressed miRNAs and not overlapping miRNAs respectively; p-value < 2.2e-16 in all cases). As expected, the conservation on average was higher for hotspots overlapping miRNA target sites, which are often highly conserved across species. Additionally, we observed that the number of RBPs binding in a window had a modest but significant negative correlation with the sum of minor allele frequencies (rho = -0.04, -0.05 and -0.03 for windows overlapping expressed miRNAs, non-expressed miRNAs and not overlapping miRNAs respectively; p-value < 2.2e-16 in all cases). This reflects both a lower frequency of SNPs and that the SNPs in the window are less frequent in the population. Together, these results indicate that RBP hotspots are functional regulatory elements under negative purifying selection.
Considering the dependencies among RBP binding sites, we decided to define hotspots in 3’UTRs as windows containing at least 5 RBPs. Some examples of RBP binding in 3’UTR hotspots can be seen in S8 Fig and S9 Fig. Using this definition, approximately 4% of all windows are classified as hotspots, whereas 56% of them are not bound by any RBPs (Fig 5b). We noticed that the number of RBPs binding in a window is positively correlated with U-content (r = 0.21, p-value < 2.2e-16) and negatively correlated with G-content (r = -0.2; p-value < 2.2e-16) (S10 Fig). Additionally, windows targeted by several RBPs have much higher sequence accessibility, measured as the probability that at least 20 consecutive nucleotides are unpaired (Fig 5a). This result is consistent with the fact that hotspots are more accessible regions, which favors the binding of multiple RBPs.
We used cWords [40] to identify motifs enriched in hotspots. We identified several AREs, including UAUUUAU, among the top 20 ranked words enriched both in hotspots and in conserved regions (Fig 5c). The core ARE element AUUUA was enriched in hotspots as well, although its frequency does not increase linearly with hotspot size (S10 Fig). We also noticed that the words enriched in hotspots overlapping miRNA target sites are very similar to those found in all hotspots (Fig 5d, S4 Table). Notably, we found an almost complete G-depletion in the top 100 words enriched in hotspots, which is consisted with the observation that hotspots have higher accessibility.
A previous study concluded that many regions found to be targeted by several different RBPs in CLIP-seq experiments are artifacts caused by biases in the experimental technique [30]. To understand if the hotspots are a result of CLIP background, we investigated the overlap between the hotspot regions identified here and a set of binding sites obtained from a GFP PAR-CLIP experiment [30]. This experiment contained 3 datasets belonging to protein-RNA complexes with different molecular weight, which we analyzed using the same pipeline described before and pooled together. We identified 11323 significant GFP binding sites, i.e. background clusters. It has to be noted that our mapping pipeline discarded most of the data as insignificant, which is consistent with the expectation that there should be very few genuine binding sites (S5 Table).
The center of the significant background clusters with positive enrichment over RNA-seq were extracted and overlapped with the previously identified windows in 3’UTRs. 7% of background clusters (837 sites) overlapped with previously identified hotspots containing 5 or more RBPs. Thus, only 5% of hotspots overlapped background clusters. To validate that these background clusters were not biasing the results, we discarded all the windows containing GFP sites and repeated the analysis in Fig 5. This filtered dataset recapitulated the results obtained previously (S11 Fig). Together, these results support the idea that RBP hotspots are regulatory 3’UTR elements and not CLIP artifacts.
We have observed that hotspots are more conserved, more accessible, and enriched by miRNA target sites and AREs, including UAUUUAU, which has been associated with stronger miRNA effect and mRNA stabilizing effect [3,41]. Hence, we set out to assess the effect of RBP hotspots on hisites using previously published AGO2 KD microarray data [42]. For each transcript, we defined a new set of 50 nt windows centered on hisites and measured the effect of the presence of a hotspot (excluding AGO2 when defining the hotspots) overlapping 1, or 2 or more hisites in a transcript upon AGO2 KD. As a control, we used two sets of transcripts, one where all the hisites were in windows containing 2 or less RBPs and another one in which transcripts contained no hisites at all. By calculating the cumulative fractions of fold expression changes of transcripts upon AGO2 KD, we found that the presence of a hotspot overlapping hisites in a transcript prevents its upregulation upon AGO2 KD (two tailed KS test p-value = 0.0017 and 6.9e-08 for 1 and 2 or more target sites blocked compared to genes without hotspots on hisites respectively) (Fig 6a). This result cannot be explained by differences in 3’UTR length, number of hisites in 3’UTRs, or expression biases across categories (S12 Fig). Thus, it suggests that RBP hotspots can prevent the binding of RISC on miRNA target sites.
Additionally, we explored the function of hotspots overlapping the binding sites of other RBPs using published KD data. Similarly to the previous analysis, we defined hotspots centered on significant HuR clusters identified with CLIP data and measured the effect of the presence of a hotspot (without considering HuR) overlapping 1, or 2 or more HuR binding sites upon HuR KD [16]. As a control, we used transcripts where all the HuR binding sites were located in windows containing 2 or less RBPs (not counting HuR), or not bound by HuR. As expected, upon HuR KD, transcripts containing HuR binding sites show lower expression compared to transcripts not bound by HuR (Fig 6b, two-tailed KS test p-value < 0.001 for transcripts with 1 and 2 or more HuR sites blocked compared to genes without hotspots on HuR sites). We also observed a small but significant difference between transcripts containing hotspots overlapping HuR binding sites and those that do not have them (KS test p-value < 0.001 in both cases). In this case, transcripts containing HuR sites overlapping hotspots show a decreased expression upon HuR KD than transcripts containing HuR sites outside hotspots. This result suggests that upon HuR KD other RBPs with a negative effect on mRNA stability bind in those locations and thus promote mRNA downregulation. Accordingly, we found that 55% of the hotspots overlapping HuR sites contained AGO proteins, TNRC6 proteins, AUF1 or TTP, which are all known to be involved in promoting mRNA decay. Similar results were observed when analyzing the effect of hotspots overlapping AUF1 and TTP binding sites (S13 Fig). Taken together, these results show that changing RBP concentration within cells may affect their binding on mRNAs and modify their post-transcriptional regulation.
In order to understand the biological function of RBP hotspots transcriptome-wide, we sought to characterize the transcripts containing hotspots. Transcripts with hotspots possess some features that suggest that they are under strong post-transcriptional regulation, as they have longer 3’UTRs (spearman correlation coefficient rho = 0.17; p-value = 1.2 e-60) while keeping approximately the same density of miRNA target sites (S14 Fig). Furthermore, we also noticed that they are significantly higher expressed than transcripts without hotspots (rho = 0.3; p-value = 1e-191 S14 Fig).
We used PANTHER [43] to characterize their functions using GO-terms. The most significant molecular function terms identified were polyA RNA binding, RNA binding and nucleic acid binding (p-value < 0.001; S6 Table). Among the genes that contain these terms RNA binding proteins, splicing factors and transcription factors stand out (S15 Fig), which suggests that hotspots could be central in the regulation of both transcriptional and post-transcriptional processes.
In this work, we have reanalyzed a comprehensive collection of high-throughput CLIP experiments in HEK293 cells in order to better understand the complex interactions between RBPs and miRNAs in post-transcriptional regulation. Our results show that RBPs and miRNAs often bind in the same regions within 3’UTRs, which suggest that they function as regulatory hotspots that facilitate competition between the regulators. We show that these hotspots function in an analogous manner to promoter regions in accessible chromatin regions, and the RNA fate depends on which of the regulators bind to the mRNA. In turn, this regulation would also depend on external cues or post-translational modifications that modulate the relative concentration of these factors or their affinity to mRNA. Interestingly, RBP hotspots are enriched in transcripts involved in transcriptional and post-transcriptional regulation, such as RNA binding proteins, splicing factors, transcription factors and translation factors (S15 Fig), thus suggesting that these regulatory hotspots play a role in auto-regulatory networks of regulators previously reviewed [44]. This result is also in agreement with recent findings that show that RNPs tend to regulate the mRNAs of other RNPs and themselves thus creating auto-regulatory networks in Drosophila [45].
We have used positional correlations to assess the interactions between RBPs assuming that RBPs that bind in the same location may interact. Surprisingly, we found that most RBPs (69% of all RBP pairs analyzed) tend to have overlapping binding sites (Fig 2a and S2 Fig). Using this approach we confirmed some known positional correlations, such as those among polyA complex proteins [46], the IMP and the AGO proteins [36], the FET family proteins (TAF15, FUS, EWSR1) [47], and the snoRNA processing proteins FBP, NOP56 and NOP58 [48]. Moreover, we found correlations that were previously unknown. Some of these can be explained by the similarity in binding motifs, such as those among HuR, TTP and AUF1 [11,13,16]. However, the consistent correlation of all the RBPs analyzed with AGO2 had not been previously described in literature. Additionally, the finding that many of these RBPs are also enriched on hisites further supports the positional correlations. It has to be noted that these miRNA target site predictions are independent of the CLIP data [49], which speaks against these overlaps being an artifact of the CLIP protocol.
One intriguing question is why the RBPs bind on miRNA target sites. If RISC directly interacts at miRNA target sites with a particular RBP, it would be expected that CLIP enrichment covaries with the expression of the miRNA that targets it. Yet, in most cases we did not find a clear correlation (S5 Fig) and thereby direct interaction is probably not the general mechanism to explain RBP enrichment at miRNA target sites. Nevertheless, we found a positive correlation for several proteins from the polyadenylation complex although it is only significant for CPSF-73 and CF-Im68 (Fig 3c). We found that many proteins from the polyA complex bind at hisites. As expected, we observed a stronger enrichment of the RBPs from the polyA complex on distal hisites compared to proximal or medial (Fig 4). Interestingly, we observed two clearly different types of binding profiles around hisites: one for the cleavage factors, Fip1 and CPSF-160, which show a strong enrichment specifically on distal hisites; and another one for WDR33 and CstF-64, which show an increase in the enrichment specifically after hisites. Our data merely shows that miRNAs and the RBPs from the polyA complex bind in the same location, which is likely the result of the presence of both sets of regulators in hotspot regions, which are more accessible, and thus preferred for both purposes. One may speculate that the strong enrichment of polyA complex RBPs on hisites could indicate some kind of functional interaction between these two pathways. However, the nature of these interactions cannot be explored using computational methods and additional experiments would be required to investigate it.
We have also shown that RBP binding sites cluster in regulatory hotspots in 3’UTRs. These hotspots are more frequent than expected if RBPs would bind independently (S6 Fig) and are significantly enriched on predicted miRNA target sites (Fig 5b). Furthermore, they are AU-rich (S10 Fig) and contain AREs, which are both more conserved and overrepresented (Fig 5c). AREs and AU-rich context of miRNA targets have previously been associated with effective miRNA target sites [3,41,50] and are known to be targeted by many RBPs both with stabilizing and destabilizing functions. These results, together with the high accessibility of the RBP hotspots, could explain the large number of RBPs binding in the regions and their role as regulatory elements in 3’UTRs. It has to be noted that several of the RBPs analyzed in this paper are known to bind A-, U- or AU-rich motifs. Some of these RBPs, such as HuR, TTP and AUF1, can bind AREs. However, many others bind completely different motifs, such as the RBPs from the FET family, which bind AU-rich stem loops, and the proteins from the polyadenylation complex among others [11,13,16,36,46,47,51]. Besides, CLIP crosslinking introduces a U-bias, either by the use of UV-C radiation in HITS-CLIP and iCLIP [52] or by the incorporation of 4-thiouridine in the RNA to cross-link the RBPs in PAR-CLIP [36]. These two factors may explain a part of the U-richness observed in the hotspots, as shown by a sequence composition analysis of the significant clusters in 3’UTRs of the individual RBPs (S16 Fig).
Finally, we have shown that RBP hotspots regulate miRNA target site accessibility and could favor the competition between miRNAs and RBPs in 3’UTRs. Upon AGO2 KD, transcripts containing miRNA target sites in hotspots do not show significant increased expression levels, which suggests that these target sites were protected by RBPs binding in the same hotspots (Fig 6a). Interestingly, the opposite effect was found upon KD of HuR, AUF1 and TTP. Upon KD of these RBPs, transcripts containing their binding sites in hotspots show a stronger reduction in expression levels than those that have their binding sites isolated. A high fraction of those hotspots contain AGO2 or other down regulatory RBPs, which suggests that by removing the RBPs, other ones bind and affect mRNA stability (Fig 6b and S13 Fig). These results are in agreement with recently reported findings that show that the presence of RBP binding sites of overlapping PUM1/2 or HuR binding sites reduce their impact on mRNA stability [26].
A previous study concluded that many regions found to be targeted by several different RBPs in CLIP-seq experiments are artifacts caused by biases in the experimental technique [30]. As a result, these regions have been excluded from previous works analyzing the combined effect of RBPs and miRNAs in post-transcriptional regulation [26]. In our analysis, we observed that the number of RBPs that target a hotspot weakly correlates with mRNA expression (S14 Fig) regardless of the use of mFDR and RNA-seq normalization of CLIP data. This bias probably hinders the identification of hotspots in lowly expressed genes, but it is unlikely that this is the reason why we observe hotspots in 3’UTRs. To validate that these regions are truly regulatory elements, we have analyzed the 3 GFP PAR-CLIP experiments used to identify CLIP artifacts [30] and studied their overlap with hotspots. Using our pipeline, only 2% and 9% of the original GFP libraries can be confidently mapped to the genome (S5 Table). These results confirm the stringency of our pipeline and give extra evidence that the processing pipeline discards most of the spurious CLIP binding artifacts. Besides, additional analyses of this dataset further support that the identified regulatory hotspots are not the result of background CLIP binding sites. Firstly, it was described that background reads, i.e. the reads that appear in multiple datasets derived from a CLIP experiment of a protein that does not bind RNA [30], are G-rich. In contrast, our RBP hotspots are characterized by a general G-depletion and are AU-rich (S10 Fig). Secondly, the CLIP enrichment of RBPs is in most cases not different inside or outside hotpot regions. If CLIP binding in hotspot regions would be spurious, we would expect a consistently lower CLIP enrichment of all RBPs in these regions. However, for many of the RBPs analyzed the number of RBPs in a window does not affect the distribution of their enrichment values (S17 Fig). Thirdly, the regulatory hotspots that we identify are experiencing increased selective pressure, as shown by the higher PhyloP scores and lower SNP frequencies (Fig 5a). The correlation between conservation and number of RBPs in a window is observed for all windows regardless of their overlap with miRNA target sites, which indicates that the increased selective pressure happens due to the preservation of binding sites for more RBPs. Moreover, it supports the conclusion that these regions are indeed functional regulatory elements. Fourthly, we show that hotspots more often coincide with miRNA target sites, which are independent of CLIP-seq data. In conjunction, we see a significant functional effect of hotspots in the regulation of sequence accessibility both using KD data from AGO2 and other RBPs (Fig 6 and S13 Fig). Taken together, we believe that our stringent pipeline for the processing of the datasets, which includes duplicate removal, quality score aware mapping of reads, peak calling of clusters in transcripts, and normalization by gene expression, removes or diminishes the importance of most of the reads that were shown to result in background when a less stringent data pipeline was used [30] and thus allow us to identify truly regulatory regions targeted by several RBPs. Accordingly, only 5% of our regulatory hotspots, i.e. windows containing 5 or more different RBPs, overlap background sites identified by analyzing the GFP CLIP sites identified with our pipeline. Removal of these windows from our dataset did not alter the characteristics of the identified hotspots (S11 Fig), which confirms our observation that RBP hotspots are not CLIP artifacts.
Many studies have previously investigated the interaction between miRNAs and RBPs using both experimental and computational methods [14,25,26,53,54]. Both competition and collaboration between miRNAs and RBPs have been described, but these interactions have been often portrayed as isolated events rather than a general mechanism in post-transcriptional regulation. In this work, we have shown that the overlap between miRNA target sites and RBPs is very extensive, with more than 75% of all hisites targeted by one or more of the RBPs analyzed (excluding AGO and TNRC6 proteins), thus suggesting that RBP hotspots play a major role in miRNA regulation and post-transcriptional regulation.
Taken together, our analyses suggest that post-transcriptional regulation often happens in hotspots where several trans-acting factors bind and may compete and cooperate for regulating mRNAs. This organization thus facilitates fast changes on mRNA expression induced as a response to external cues and facilitate cell adaptation to environment changes.
110 CLIP (including CLIP-seq and PAR-CLIP) and 3 RNA-seq datasets were downloaded from GEO database [55]. The Sequence Read Archive (SRA) accession numbers of all the datasets analyzed can be found in S5 Table.
Reads from all the experiments were preprocessed using custom python scripts. First, reads were trimmed to remove low quality scores and 3’ adapter sequences (only CLIP datasets). Next, we removed duplicates by collapsing all identical reads. This step was performed instead of collapsing reads that map to identical locations to keep fragments that contain different mutations as a result of cross-linking and that represent true crosslink events that otherwise would be discarded. After these steps, all reads longer than 19 nucleotides were further analyzed. This minimum length was set to minimize the amount of incorrectly mapped reads that could come from contamination [56]. Reads were mapped to the human genome (hg19) using BWA-PSSM with parameters -n 0.04 -l 1024 -m 400 -P 0.5 [35]. Then, all unmapped reads were then mapped to an exon-junction index containing all annotated unique exon-junctions from human Ensembl70 transcripts [57]. Only reads mapped at any of the steps with a posterior probability > 0.99 were considered for further analysis. For PAR-CLIP datasets, we used a custom matrix for scoring T to C mismatches assuming a 12.5% T to C conversion rate.
Datasets for the same proteins were joined into a single dataset and analyzed together. Additionally, we also pooled the datasets of CstF-64 and CstF-64τ and FXR1 and FXR2. Reads were clustered according to their genomic positions, requiring that at least 1 nucleotide overlap. Significant clusters were calculated using Pyicos [38], using the exons from the longest protein coding transcript for calculating the randomizations. Only clusters with a false discovery rate (FDR) < 0.01 were considered for further analysis. The RNA-seq datasets were also joined and used together in further experiments. The statistics summarizing the preprocessing steps, mapping, clustering and peak calling can be found in S5 Table. Additional analysis performed to validate the processing pipeline can be found in the S1 Appendix. The pipeline used for preprocessing and mapping of the data is publicly available on GitHub under an Open Source license (https://github.com/simras/CLAP).
We obtained the significantly overrepresented biological process GO-terms associated with the RBPs included in the analysis using the gene enrichment analysis method performed by PANTHER [58]. The clustering and visualization of enriched GO-terms was done using REVIGO (http://revigo.irb.hr/) [59].
For each 3’UTRs we calculated mk, the average number of RNA-seq base calls per nucleotide and then we normalized to M, the total amount of mapped RNA-seq reads in the experiment as follows
mk=∑j=1lkrjM·lk
where lk is the length of the 3’UTR for gene k, and rj is the count of RNA-seq reads in position j.
For each transcript, we built a single-nucleotide resolution profile of the RBP binding sites, i.e. significant clusters with an FDR < 0.01 after peak calling, normalized to RNA-seq. The enrichment e of CLIP in a position i of a particular 3’UTR k is calculated as
ei,k=cimk·N
where ci is the count of clip reads in position i, N is the total amount of uniquely mapped CLIP reads, and mk is the average gene expression as defined above.
Good miRNA target site predictions for conserved and non-conserved miRNAs were downloaded from microRNA.org (http://www.microRNA.org; August 2010 release) [49]. From this set, we selected only target sites containing at least a 6-mer seed site, and selected targets that belonged to miRNAs expressed in HEK293.
We used small RNA-seq data (GSM1279922) [60] to estimate the expression levels of each miRNA. First, we selected from the dataset reads that were 15-27nt long, which corresponds to the length range of mature miRNAs. Next, we mapped the RNA-seq to a set of non-redundant human miRNA sequences downloaded from miRBase [8] using BWA-PSSM [35]. The expression of each miRNA was defined as the number of reads mapping to its mature miRNA sequence. We defined as expressed miRNAs only the top 20% of the mature miRNAs (155 miRNAs; minimum amount of mapped reads mapped 367). All the other miRNAs not included in this set are regarded as non-expressed miRNAs in HEK293 cells.
For each of the target sites, we selected the set of targets that overlapped Ensembl70 [57] transcripts expressed in HEK293 cells and defined a set of non-overlapping target sites. To define which target sites to keep, we overlapped the seed sites of their target sites and kept the one targeted by the most highly expressed miRNA. If several miRNAs shared the target site, we added their expression. Finally, we kept only target sites that contained a 6-mer seed site in the selected transcript. The number of target sites kept at each step of the processing is summarized in S7 Table.
For some of the analyses we divided target sites according to their total expression, i.e. the sum of expressions of miRNAs targeting the same site, in three equally sized groups: highly expressed (hisites), moderately expressed and lowly expressed.
To measure the significance of our results, we created 100 random sets of miRNA target sites containing as many target sites as the original set preserving their distribution along 3’UTRs. We divided the set of expressed genes with predicted miRNA target sites into 30 equal size groups with similar 3’UTR lengths. Then, for each target site in a particular 3’UTR, we assigned it to another of the 3’UTRs in the set. In case that the length of the new 3’UTR was different from that of the original 3’UTR, the relative coordinates of the target site were calculated so that it would have the same relative position within the 3’UTR in relation to its length. This procedure preserved the characteristic positional distribution of miRNA target sites along 3’UTRs.
The significant CLIP clusters for each of the RBPs were overlapped with the genes from Ensembl70 [57] annotation using fjoin [61]. Only the longest protein-coding transcript for a gene was considered. If a cluster would overlap the CDS and a UTR region, the UTR annotation was assigned.
We analyzed the positional distribution of data across 3’UTRs of expressed genes (RNA-seq coverage > = 50%) and around hisites. Each 3’UTR was divided in 50 equally sized bins. For each bin, the mean value per nucleotide was calculated and then averaged across all expressed genes. In the case of CLIP data, the position of significant CLIP clusters (FDR < 0.01) was used to draw the profiles. In the case of hotspots, the position of the 50nt windows containing n (n = 1,2,…31) RBPs mapped on them was used. For miRNA target sites, the position of the target seeds in 3’UTRs was used.
To find the positional correlation between the binding of two different proteins, we calculated the Pearson correlation between the enrichment values along a 3’UTR. If the value at position i is called xi for one RBP and yi+d for the other RBP binding a distance d from the first, the Pearson correlation was calculated with fixed d over all positions i, in the interval from 1 to l − d, where l is the length of the 3’UTR (for negative d, the interval is from 1 − d to l). This was done for all values of d from -200 to 200. For each d, the correlation values were averaged over 3’UTRs. UTRs shorter than 400 nt were discarded.
The fluctuations of the correlation coefficients are heavily dependent on the number of CLIP sites. To estimate the background distribution, we shuffled the CLIP data in a way that preserved the clustering of tags. Clusters were defined as contiguous regions in which the enrichment value was above 10−6. The clusters identified in a sequence were moved to a random location in the sequence while ensuring at least one position in between clusters. After shuffling all sequences, positional correlations were calculated as above. This was repeated 100 times and for each d, and the mean and standard deviation of the 100 values obtained in the shufflings were calculated. Using these estimates, the z-score was calculated for the unshuffled data. In Fig 2a, the distribution of all z-scores calculated was considered and divided in 1000 quantiles. Each quantile was assigned a color from the scale, ranging from dark blue to red as shown. In Fig 2b, z-scores were row-normalized and assigned a color using the same procedure as described above.
To identify hotspots we divided the 3’UTRs of expressed genes (at least 50% RNA-seq coverage in the 3’UTR of the longest protein-coding transcript) in non-overlapping windows of 50 nt. We overlapped the center of the RBP CLIP significant clusters with them and assigned each cluster to a single window. Only clusters with a positive enrichment over RNA-seq were considered. We also uniquely assigned each miRNA target site of expressed miRNAs in HEK293 cells to a window if the overlap between the seed site and the window was bigger than 5. Otherwise, the miRNA target sites were discarded.
We simulated 10000 times the distribution of hotspot sizes by randomly sampling the binding location of the proteins assuming a uniform distribution of the RBPs in them. We considered the total amount of windows in which we observe significant clusters of each RBP and the total amount of windows in 3’UTRs (S7 Table). The size distribution of hotspots from simulated and real data can be seen in S6 Fig.
PhyloP scores [39] calculated from 100 vertebrate genome alignments (including hg19 human genome assembly) were downloaded from UCSC genome browser. For each of the 50nt non-overlapping windows, we calculated the mean phyloP score across the window, discarding regions that were not present in any of the other species.
Word enrichment analyses were done using cWords [40]. The input data sets were made using the 3’UTR window data described above. For each window, we extracted its sequence and associated it to the number of RBPs binding in it. Using this method we defined two datasets: one containing all windows and another one containing only those overlapping target sites for expressed miRNAs.
In the first analysis, windows were ranked using the amount of RBPs binding in them. Thus, the resulting words were differentially enriched in windows according to the number of RBPs binding in them. In the second analysis, we ranked the windows using their mean phyloP score.
We used RNAplFold [62] to calculated the sequence accessibility of the 3’UTRs. Specifically, we predicted the probability that 20 contiguous nucleotides in the sequence are unpaired using the parameters -u 20 -L 40 -W 120. The obtained accessibility values were then mapped to the 3’UTR windows and averaged across windows with the same number of RBPs binding and across windows with the same number of RBPs that overlap miRNA target sites.
The complete data set of the 1000 genomes project containing all variants mapped to hg19 assembly [63] was downloaded. Of all the variants, we only used mutations regardless of their size and required them to be present in at least two individuals in a population of 5008. We calculated the mean of the sum of all minor alleles as 1 - major allele frequency regardless of which was the reference allele across windows as described above.
We downloaded the microarray data containing the expression values for AGO2 KD (GSM95818, GSM96819, GSM96816 and GSM96817) [42] and HuR KD (GSM738179, GSM738180, GSM738181, GSM738182, GSM738183) [16] from GEO database. We calculated differential expression upon AGO2 KD using the limma package [64] in R. We also downloaded processed data from KD experiments in AUF1 [13] and TTP [11].
We defined 50 nt windows around hisites (35 nt upstream of the target site 3’ end, 14 nt downstream of the target site 3’end). If the windows extended beyond transcript boundaries, we shrank them so that they would fit inside the transcript. In each of these windows, we checked the presence or absence of each of the RBPs.
For each transcript we measured the amount of hisites that would be free, i.e. 2 or less RBPs (excluding AGO2) would bind in the window around the hisite, and the amount of hisites that would be blocked, i.e. 5 or more RBPs (excluding AGO2) would bind in the window around them. We used these measurements to divide the genes according to the amount of free or blocked hisites they contained in 3 groups: 0, where all target sites are free; 1, where only 1 target site was blocked; 2 or more, where 2 or more target sites were blocked. As an additional control, we added the rest of genes containing no hisites.
For cumulative fraction plots centered on RBP binding sites, we defined 50 nt windows centered around the binding sites of the RBP of interest. The groups of transcripts used to evaluate the role of hotspots on RBP binding sites were built in an analogous manner to the one described above.
|
10.1371/journal.pcbi.1002699 | Evaluating the Adequacy of Gravity Models as a Description of Human Mobility for Epidemic Modelling | Gravity models have a long history of use in describing and forecasting the movements of people as well as goods and services, making them a natural basis for disease transmission rates over distance. In agent-based micro-simulations, gravity models can be directly used to represent movement of individuals and hence disease. In this paper, we consider a range of gravity models as fits to movement data from the UK and the US. We examine the ability of synthetic networks generated from fitted models to match those from the data in terms of epidemic behaviour; in particular, times to first infection. For both datasets, best fits are obtained with a two-piece ‘matched’ power law distance distribution. Epidemics on synthetic UK networks match well those on data networks across all but the smallest nodes for a range of aggregation levels. We derive an expression for time to infection between nodes in terms of epidemiological and network parameters which illuminates the influence of network clustering in spread across networks and suggests an approximate relationship between the log-likelihood deviance of model fit and the match times to infection between synthetic and data networks. On synthetic US networks, the match in epidemic behaviour is initially poor and sensitive to the initially infected node. Analysis of times to infection indicates a failure of models to capture infrequent long-range contact between large nodes. An assortative model based on node population size captures this heterogeneity, considerably improving the epidemiological match between synthetic and data networks.
| An accurate representation of disease transmission between spatially-distinct regions is an essential part of modelling epidemic behaviour on a national or international scale. Gravity models, which describe movement fluxes between regions in terms of their populations and distance from each other, have a history of successful use in the geography and economics and are increasingly used in epidemiology. We look at the ability of a range of gravity models to fit human movement data from the UK and the US. In particular, we compare the behaviour of a simple flu-like epidemic model on synthetic networks generated by fitted gravity models and on the original network present in the data, using time to first infection. For UK data, epidemic behaviour on synthetic networks matches that on the original data quite closely. For US movement data, synthetic networks perform much worse. We develop an analytic expression for infection time between two regions which indicates that our gravity models fail to capture long range connections between large populations. A model with assortative mixing based on population size greatly improves the match between synthetic and data networks.
| Gravity models of population movement characterize the distribution of trips between discrete locations, based on the populations of the origin and destination and the distance between them. The use of gravity models to describe movement between population centres can be traced back at least as far as the work of Zipf in the 1940's [1]. In this work, Zipf provides a theoretical motivation for movement between cities 1 and 2 being governed by a P1P2/d relationship, where P is the respective city population and d is separation distance. He also identifies this relationship in passenger and freight movements between US cities [1]. Until recently, the main areas of application for gravity models has been in analyzing and forecasting the demand for goods and services in spatially distributed populations [2]. Consequently, much of the theoretical work has focused on the modeling of journey costs and discrete choice models [3].
More recently, gravity models have been adapted to describe the spread of a range of biological agents, such as invasive species [4] and plant pathogens [5]. Of particular interest in the context of this paper is their use as a description of infectious disease transmission between regions. As a description of movement between spatially distinct populations, gravity models offer a simple model of disease transmission strength between meta-populations. Xia et al. model the dynamics of pre-vaccination measles as city meta populations connected by a gravity-based movement of infectious individuals [6]. The model succeeds in capturing most of the spatiotemporal properties of epidemics, including case rates, periodicity and fade-out behaviour. A similarly structured model was applied to seasonal influenza data in the US by Viboud et al. [7]. In this case, it was found that fitting the underlying gravity model to commuting data successfully captured the observed synchrony in epidemics as a function of distance, population size and transmission. Gravity models are now increasingly used in both metapopulation- and individual-based epidemic micro-simulations [8]–[11].
The aim of this paper is to look at the ability of gravity models to represent commuter movement data in the UK and US to the fidelity needed to capture expected patterns of epidemic spread. We examine which models best capture the statistical properties of the data and what aspects are poorly represented. We address how the choice of level of spatial aggregation influences the fit of models. Most importantly, we examine the behaviour of a simple SIR epidemic on synthetic networks reconstructed from fitted gravity models and ask how epidemic behaviour depends on the underlying model and on the level of aggregation.
The UK population and commuting data is taken from the 1991 census. The data set combines information for England and Wales collected by the Office for National Statistics and Scottish data from the General Register Office for Scotland [12], [13]. The smallest region of aggregation of data available is the Census Area Statistics Ward, corresponding to the electoral wards which define the basic political and administrative geographical units in the UK. We look at the fit of models at three different levels of aggregation, featuring wards, district and counties as the basic regions. We concentrate on the district level, as districts correspond well with individual cities and are comparable in many ways to counties in the US data-set.
Information on commuting patterns comes from the associated Special Workplace Statistics for both Scotland and England and Wales. The data sets comprise the responses from a randomly chosen 10% of the surveyed population who were asked for the location of their place of work and their means of transport for commuting. From this can be constructed the flow of commuters between any two pairs of wards (in this paper we aggregate across modes of transport to recover the total flow).
The models we fit to the data are functions of the destination work population, so we exclude the small fraction of wards (1%) that have no commuter population. For the remaining wards, our data set consists of a population of wards with resident population and location (as northing and easting) and a set of commuting sub-populations between pairs of wards. This data set can easily be spatially aggregated to the district of county level.
We have chosen to use the 1991 census data as opposed to the more recent 2001 set, which entails two potential drawbacks. Firstly the data is a decade older, but more importantly only 10% of commuting data was obtained in 1991. Use of the 2001 data is problematical, however, due to disclosure control methods introduced in that year. The so-called ‘small cell adjustment method’ (SCAM) involves the adjustment of small cell counts (1 or 2) in the data to either 0 or 3 (precise details of the algorithm have not been released). Stillwell and Duke-Williams discuss some of the consequences arising from SCAM [14]. Among those relevant to the current work are that small commuter counts are more likely at longer distances and these are particularly important for fitting the power-law distributions used in our models and also that small numbers of individuals may still trigger disproportionately large local epidemics. A further issue is that SCAM is not applied to movements with origins in Scotland and is hence spatially inhomogeneous.
US population and commuting data is taken from the US Census in 2000. The data is available from the US Census Bureau web-site [15]. With regard to working habits, the census asks participants to identify the location of the place they worked at most frequently in the previous week. Hence we might expect the resulting data to contain more irregular and long-distance journeys than the UK data. Commuting data for 2000 at the county level was retrieved from the site along with populations and geographical centroid location for each county. We consider only counties from the 48 contiguous states and exclude all movements to US territories. The mean population of the included counties is approximately 90,000 residents.
We look at movement models that calculate the probability, pij, of a journey from node i to node j(1)where Pi is the resident population, Ci is the population who work in node i (regardless of where they live) and θ is a vector of parameters. The function f(d) is a distance kernel which encapsulates the effect of separation on the probability of a journey between two locations. We normalize to impose the constraint .
As well as this ‘unconstrained’ model, we consider a ‘constrained’ model which assumes that the probability of a journey emanating from node i (including those within node i), qi, is matched to that in the data. That is,where Tij is the flow from node i to j and Ttot is the total number of journeys in the data. We then express the probability of a particular journey in a conditional form, , where(2)By setting parameters to zero, we can examine a range of sub-models. We look at models of distance interaction of two forms, smooth kernel (SK),and a two-piece ‘matched’ kernel (MK),where π is a point probability mass for commuting within the node of residence (d = 0), α is a distance scale below which the kernel function saturates, and γ, γ2 are power parameters. Here A is a function of the other parameters such that f(d) is continuous at dc.
These models are fitted to the movement data by constructing a likelihood function for the data and maximizing using standard MCMC techniques. We use a multinomial likelihood for the distribution of journeys among nodesDiscarding terms independent of the movement model gives a log-likelihood of(3)In the case of the locally-constrained model, the log-likelihood can be split into two terms(4)If the set of qi are considered as N additional independent parameters, it is easy to show that local constraints correspond to the maximization of the likelihood with respect to qi, allowing a direct comparison of the two model types.
The raw commuting data can be represented as a weighted graph, the nodes of which are the locations represented in the data. Edge weight is then just the number of reported journeys between two nodes. One can then simulate an epidemic occurring on this data-derived network. An equivalent interpretation is that of a metapopulation, where the patches are the locations in the data and the origin-destination flow matrix is used to construct the patch-to-patch coupling matrix. Similarly, we can construct synthetic networks from the fitted movement models and compare the dynamics of epidemics on those networks with the dynamics of epidemics on the data-derived network.
In order to make these comparisons, we employ a simple stochastic SIR (susceptible-infected-recovered) model such as might be used to describe the spread of influenza. A connection from node i to node j consists of a population, Nij, resident in i and working in j. To simplify the model, we assume that all residents of a node are working (in reality, approximately 40% of the resident population travels to work). Of this population, Sij(t) and Iij(t) are the number of individuals who are susceptible and infected at time t, respectively. The susceptible population is subject to forces of infection from the node within which they reside, , and the one at which they work, . Both the working and resident populations are assumed to be well mixed and hence the forces of infection are given byInfected individuals recover into a removed class at a rate σ. Hence in a naïve population, R0 = (βh+βw)/σ, neglecting local saturation effects. Unless otherwise stated, we use the following parameter values: βh = 0.5/d, βw = 0.4/d and σ = 0.5/d, giving a generation time of 2 days and an R0 of 1.8.
Comparison of the behaviour of the epidemic model on different networks is based on the times to first infection for network nodes from a given initial infection site. Under certain simplifying assumptions, an approximation for the mean time to infection between two nodes can be calculated for the above epidemiological model.(5)The parameters Λ1,2 and ς1,2 represent network properties, such as the fraction of journeys between the two nodes. r = βh+βw+σ is the epidemic growth rate in a large node. The equation illustrates how epidemiological effects combine with network properties to determine the speed of infection across a connection. The details of the analysis and the definitions of the parameters can be found in the Supplementary Information Text S1.
Table 1 shows the maximum likelihood values and associated parameter estimators for a range of models at the district level of aggregation in the UK. The biggest influence on log-likelihood is whether local constraints are imposed. For a given kernel type, a locally constrained model has a greater log-likelihood by a margin of approximately 230,000. The locally constrained model effectively has an extra parameter for each node (the number of workers living there). Model comparison statistics (such as the Akaike information Criterion) offset the number of model parameters against the log likelihood, but these 456 additional degrees of freedom are clearly insufficient to account for the increase in likelihood. We have omitted further discussion of models which include an exponent on the destination population, τd. The improvement in likelihood over a model with only commuter population dependence is minimal (∼2500) and not discernible in the statistics of flow distributions or behaviour of the epidemiological model.
A smaller difference arises between models with a smooth offset power law kernel (SK) and those with a matched two-section kernel (MK - see the Models section). For a given constraint type, use the MK model improves maximum likelihood by approximately 20,000 for the addition of two new parameters; a critical distance, dc, beyond which an outer power, γ2, is used. In general, credible intervals for parameters are typically less than 0.5% of their maximum likelihood estimates. Intervals for π and γ2 are slightly wider at around 1.5%, reflecting the smaller amount of data to estimate them Such narrow intervals reflect the large amount of data used in the fitting, rather than sensitivity in any of the observable statistics of the fitted model.
Across different levels of aggregation, the MK, locally constrained model consistently gives the best fit. Parameter values generally change monotonically as the aggregation unit is made smaller. In particular, note the trend in commuter population power, τc, and the point mass, π, which both approach 1 as the size of the aggregation unit approaches 1, a feature we address in the discussion. The wide range of powers and scaling parameters for the inner kernel can be understood in the context of the limiting behaviour of the offset power law. For large γ, the offset power function converges to an exponential distribution.where s = α/γ. Given the magnitude of the estimates of γ for the MK model, the ratio s is perhaps a better estimate than α of the spatial scale of the inner kernel. Its size is comparable to that of a typical node at each level of aggregation.
Figure 1 illustrates the source of the differences in likelihoods. From the point of view of the two part likelihood expressed in equation 4, the globally-constrained model needs to fit the total outflow from each node as well as the relative probabilities of journeys starting from each node. As shown in Figure 1A, the model generally underestimates the number of travelling workers in a node.
Figure 1B shows the distance distribution among connections in the SK and MK constrained models in comparison to that found in the data. The distribution from the movement data has a clear ‘kink’ at approximately 150 km, beyond which journeys are more common than would be expected under a pure power-law distribution. Only 1% of journeys are longer than 225 km, with 90% being less than 42 km. As a result, shorter journeys dominate the likelihood and introduce a strong bias in the longer journeys for the SK model.
Figure 2 compares the behaviour of epidemics on synthetic networks derived from these model fits with epidemic dynamics on the network constructed using the data. Times to first infection for each node are shown, calculated as the mean time (over 100 realisations) to the first infection of a resident of the node. Figure 2A and B show the results for the MK model for epidemics started in Camden, London, at the district and county level of aggregation. The fit is quite accurate across all nodes for both aggregations, with a root mean square error of 1.6 and 1.9 days for the district and county levels respectively. In contrast, the use of the SK model has a pronounced and characteristic effect on the progress of an epidemic (Figure 2C). Times to infection match well up to approximately 15 days, at which point the infection of subsequent nodes is delayed by up to 2 weeks. This is because the SK model underestimates the degree of contact over longer distances. Figure 2C therefore indicates that the later infected nodes are infected across long distances, from some of the initially infected nodes, rather than along longer chains of short range transmissions. The good agreement of times to infection between the data and synthetic networks is generally maintained across different initial nodes. The exception is for initial nodes with very small populations. In these cases, the times to infection for other nodes is uniformly faster for the MK modelled network than the data network (see Figure 2D). The faster transmission from the smallest nodes is matched by the faster transmission to the smallest nodes (points under the line with times>20 days in Figure 2A and B) and appears to be a general feature of these models.
Figure 3A shows times to infection on the data network against log node population. The shape and gradient of the main diagonal band is common to all seeding points, indicating a strong linear relationship between time to infection across the network and the log population of a node. For seeding in London, a second initial band of similar gradient can be seen for times less than 10 days, representing spread from the initial seeding through the Home Counties by London commuters.
The analytical approximation for the time to infection of an epidemic process on a travel network (developed in Supplementary information text, S1) is able to shed greater light onto the influence of epidemiological processes and network structure on the behaviour of the epidemic. In general, times to infection for a particular node will depend not only on the epidemiological parameters and the movement fluxes to and from that node, but also on the structure of the travel network as a whole. Times to infection will depend on the structure in two main ways. Firstly, through how direct a route exists between two nodes and secondly on the degree of clustering. Clustering potentially allows each node to be subject to more than one force of infection from different connected nodes.
In order to investigate the importance of clustering (i.e. multiple competing sources of infection), we can construct a distance matrix between nodes with elements given by the transmission times predicted by equation 5, which implicitly assumes only a single source of infection for each node. A measure of distance between 2 points in the network is then the shortest path between those points given by the distance matrix (shortest distances can be calculated using the Floyd-Warshall algorithm [16]). Figure 3B and C compares times to infection estimated using this shortest path algorithm applied using the data network with those generated by simulating an epidemic on the same network, at the county and district level of aggregation respectively. At the county level, the agreement with theory is fairly good, suggesting that, in general, the force of infection experienced by a given node is dominated by a single infected contact and that clustering plays only a minor role. For district aggregation, agreement is good for the first 15 days, after which the theoretical predictions for many nodes are late by 7–10 days. This suggests clustering of connections is accelerating transmission, almost certainly through large conurbations. We discuss this effect in detail in the final section.
Figure 3D illustrates the sensitivity of times to infection to variation in the epidemiological parameters. As discussed in SI Text S1, theory predicts that times to infection should scale proportionally with changes in the parameter grouping . This effect can be clearly seen in Figure 3D in which times to infection for different values or r are plotted against those for the default value, r0. The gradient of lines in Figure 3D should be well approximated byAnd this is the case over a range of values of βh, βw and σ (R2>90%).
Table 2 shows the best fit parameters for a range of models to the US movement data. The order of goodness of fit is the same as that for the UK data set. Credible intervals are generally smaller than for the UK data set (<0.5% for all parameter values), reflecting the larger dataset from which they are inferred. Type of constraint is again the dominant effect with a difference in log-likelihood that cannot be accounted for by the extra effective degrees of freedom (3109 nodes in this case). There is a clear secondary effect from allowing the distance kernel to be of two sections which suggests that the distance distribution of connections may have a ‘kink’ in it similar to the UK case.
Figure 4A and B illustrate the quality of synthetic networks generated from the best fitting local, matched model. There is strong agreement with data for predicted inflows to nodes (Figure 4A), but a weaker match to the distance distribution of journeys, particularly between 300 and 1200 km.
Figure 5A shows mean times to infection on the data network for all counties in the continental US against the log of their populations. There is a strong linear correlation between node infection time and log population and this relationship is to a large extent independent of the initial point of infection. The effect matches that seen in the UK epidemics, but is more pronounced.
As illustrated by Figure 5B and C, the epidemic dynamics recovered from the best fit US MK model are much poorer than seen for the UK. For a significant fraction of nodes, deviations from the target behaviour are large, going beyond the 95% confidence intervals for the times to infection on the data network. The distribution of deviations is also not consistent across initial seeding points. Seeded in Los Angeles County, times to infection from the MK model are higher than for the data network for low times to infection, but too low for the nodes with longest time to infection. From Clinton County, Iowa (population approx. 50,000), infections times are uniformly too low for the synthetic network.
The significantly poorer fit in the US than in the UK is to be expected given the log likelihood values for the underlying model. In SI Text, S2, we derive an approximate measure of the goodness of fit of the mobility model which can be compared between different datasets. This mean deviance measure is based on the relative log likelihood, Δ, and is defined aswhere is the total number of trips in the relevant dataset. The derivation shows that the quantity is also related to the expected deviation in time to infection across the model network. Figure 5D compares this goodness of fit measure to the corresponding measure of the goodness of fit of simulated epidemics on the synthetic network to those on the data network. This latter goodness of fit between epidemics is quantified by the root-mean-square (RMS) difference in mean times to infection between the data and synthetic networks, averaged across a range of initial nodes. Each point on the figure is based on the best fitting MK model applied to a different underlying dataset or level of spatial aggregation. Values of suggest that the fits of the MK model to the US data at the state level and to the UK data at the district and county level are of comparable quality. RMS differences in time to infection are also similar for these fits at 2–2.5 days. The value of for the county level US data set is much greater, indicating a worse fit, and the RMS difference in time to infection is also much larger.
There is considerable variation in the goodness of fit among epidemics started from different nodes between the MK and data US networks. This suggests that the model is failing to capture accurately some subset of the work flows in the dataset. We can use the expression for the time to infection (Equation 5) to calculate the theoretical mean time to infection for all connections in both the data network and the synthetic MK-based network to try to identify what the essential discrepancies are. The two networks differ not only in the work flows between nodes, but also in which connections are present, so it's necessary to aggregate the time to infection information to allow comparison. In Figure 6, average times to infection between pairs of nodes are shown aggregated into bins by source and destination log population size. Use of log population size is suggested both by the form of equation 5 and the strong correlation it has with infection times.
Figure 6 shows that for connections between large nodes, transmissions on the synthetic network are markedly slower than on the data network (red region). This is balanced by faster times to infection for other connections, particularly from small population source nodes. The majority of connections, represented by the centre of the diagram, fit quite well. These results suggest a distinct mechanism that is missing from the simple MK model that affects movements between highly populated nodes. Since only a small fraction of total movements is between such nodes, the effect is swamped in the likelihood. The best fitting region is for log populations between about 5.3 and 6.2, corresponding to the sources of the bulk of outflows.
To better capture the interactions between large population centres, we adapt the MK model to make it assortative with respect to node population size. The assortative model categorises nodes as large or small according to a critical size, Pc. To allow large sparsely-distributed population centres to make contact with each other, different gravity model parameters are then fitted for large to large interactions and for all other possible contacts.
The best fit parameters for the assortative model are shown in Table 3. The main contrast with the simple MK model is in the outer power parameter, γ2. The previous value of 1.8 is reduced to 1.21 for movements between large nodes, but increases to above 2 for all other types of connection. This encourages longer transmissions among large nodes and shorter transmissions where one of the nodes is small. The threshold population size distinguishing large and small populations (also fitted) was estimated at approximately 158,000.
The gain in likelihood of the assortative model over the simpler version is not large. However, the improvement in the ability to reproduce the epidemic timing seen for the data network is significant (See Figure 5D for the corresponding value). The improvement in the quality of the fit in terms of epidemic behaviour can be seen Figure 7. Figure 7A and B should be compared with Figure 5B and C respectively. The assortative structure has clearly lessened some of the bias towards transmission being too fast to smaller nodes and too slow to larger ones in the epidemic initiated in Los Angeles (Figure 7A). Equally, the epidemic started in Clinton County has ‘slowed down’, converging towards the behaviour seen for the data network (Figure 7B). These improvements in fit are reflected in the RMS time differences shown in Figure 5D.
Understanding what aspects of human movement patterns are important to capture in transmission models is important in improving our ability to predict the spatiotemporal spread of emerging epidemics. As increasing volumes of finely resolved mobility data become available, one option is to incorporate these data directly into epidemic simulations. However, the availability of such directly applicable and comprehensive data sets is confined largely to Western Europe and the United States and concerns primarily human movement. In many parts of the world, such as Africa and East Asia, such data may be available only patchily, if at all, or at an inconvenient scale of aggregation [17], [18]. Hence models are necessary to extrapolate to data poor areas. In this paper, we have looked at how well gravity movement models perform when fully supported by data. How well they perform with limited data is a topic for further work.
Optimally, we would like a mechanistic but parsimonious model of human mobility which captures just sufficient detail to adequately represent the spatiotemporal spread of infection. ‘Adequately’ is clearly a subjective term, but a clear minimum criterion is that any model of mobility produces spatiotemporal dynamics that are not qualitatively different from those produced by raw mobility data itself. A much more rigorous criterion might be that a mobility model reproduces the connectivity of specific individual locations sufficiently accurately that the expected time to infection of every location estimated from mobility data is reproduced by a model to a certain level of precision.
In this paper, we have focussed on the first criterion rather than the second – we are interested in matching the marginal statistical properties of spatial epidemics at the level of the ensemble of included locations, rather than the risk profile of each individual location.
Our study shows that quite simple gravity models are able to capture many features of UK commuter flows at a variety of spatial scales. However, for a good fit to the observed patterns, it proved necessary to constrain the models to exactly reproduce the number of commuters resident in each node (i.e. total outflow, including self-flow). As shown in Figure 1A, the unconstrained model did quite poorly at matching this feature of the data. This seems surprising in light of the fact that the fraction of workers living in a node is well described as a fraction of the total population (approximately 36% in the UK). However, from equation 1, the probability of a worker living in node i isThis expression is clearly dependent on the number and ‘attractiveness’ of other nodes around node i, indicating that the globally constrained model is density dependent. As result, the model favours greater outflows in more population- and node-dense areas. The locally constrained model largely removes this density dependence.
The introduction of a matched two-part kernel greatly improves the ability of gravity models to reproduce the observed distance distribution of journeys. The improvement to the fit is mainly seen in journeys longer than 200 km which are quite rare in the data and hence give only a modest improvement in likelihood.
The UK data set allows for models to be fitted to three nested levels of aggregation; ward, district and county. As the spatial scale of aggregation decreases, most of the parameters change monotonically (τc,λ = α/γ,γ2,π). The theoretical limit of this aggregation process would be at the level of the individual at which point, powers on population sizes are meaningless. Hence the apparent convergence of the parameter τc to 1 suggests that gravity models may continue to be valid down to this level. The apparent convergence of π to 1 also suggests the distance kernel at the individual level might be smooth without a discontinuity at 0. This is encouraging for the use of gravity models in individual-based micro-simulations [10], [19].
Epidemics run on synthetic mobility networks derived from the best-fit UK model match quite closely simulated outbreak behaviour on the data network. The only clear bias, which is replicated at all levels of aggregation, is faster rates of spread both to and from the smallest nodes than are seen on the data network. For the poorer fitting models (locally constrained SK and globally constrained MK), their ability to reproduce epidemic dynamics on the data network is not well predicted by their likelihood values. The globally constrained MK performs comparably to the locally constrained model, but the SK model is unable to reproduce the timing of infection of the nodes infected latest in an epidemic (Figure 2C). This is because infection does not spread to distant nodes in a wave-like manner (i.e. utilising long chains of strong, short distance connections), but rather through weaker, long distance connections from the first few infected nodes. As a result, biases in the models in the strengths of rare long distance connections can have an effect on an epidemic out of all proportion to the effect of those long-range connections' contribution to the likelihood.
The transmission time theory developed in SI Text, S1, proves to be a good predictor of the times to infection generated by simulation, although with a slight tendency to overestimate times to infection. It represents the sensitivity of epidemic dynamics to the epidemiological parameters well, both qualitatively and quantitatively. The approximation gives a fairly accurate prediction for first infection times at the UK county level (Figure 3C), under the assumption that infection travels between nodes via the minimum-time route only. However, at the district level, this approach starts to break down. While infection times of nodes affected early in an epidemic are well described, later infection times are considerably over-estimated. This appears to be because some nodes are subject to significant forces of infection from more than one source node. Close examination of infection times indicate that transmission through large conurbations is accelerated in comparison with the naive single quickest path algorithm. Neighbouring nodes in cities are very well connected and hence generate many sources of force of infection for a susceptible node. A related issue is that the assumption of small proportional flows between nodes underpinning the analysis may be broken between densely populated contiguous nodes. Essentially, nodes within conurbations are often so strongly connected that they cannot be regarded, epidemiologically, to be independent weakly interacting communities. An equally accurate but more parsimonious division of the population geographically could be achieved by amalgamating such urban nodes. The time to infection expression we derived could be used to discriminate which nodes should be amalgamated by defining a minimum transmission time for two nodes to be considered independent. This is an area for future work.
When applied to the US data, the rank order of fit of the different gravity models examined was the same as for the UK analysis, but the quality of fit was considerably worse for the US county-scale data (the finest level of aggregation considered). The mean deviance measure, , for the best fit MK model makes this clear (Figure 5D). Although the predicted node inflows match the data well, reproduction of the distance distribution of journeys is not particularly accurate (Figure 4).
Comparing epidemics run on the best fit synthetic and data networks for the US shows a much poorer fit than obtained in the UK analysis to times to infection, in line with what might be expected from the mean deviance statistic. The much improved fit at the state level compared with the county level suggests that some of the problem in matching the data arises from heterogeneities in movements and population at the county scale. A number of refinements of the basic MK model were attempted, such as: inclusion of a mean income term and population density term alongside the population terms in equation 2; inclusion of a third matched kernel section to better match the distance distribution (see Figure 4B). None of these produced any appreciable improvement in likelihood or epidemic fit. However, the deviances in time to infection of nodes shown in Figure 6 suggest a limited number of flows into and out of the most populated nodes are underestimated by the model. Addressing this with an assortative spatial interaction between large and small populations improves the fit between epidemics. The principle change in the model is an increase in the distance over which movements can occur between large population centres.
A plausible explanation for the excess in long-range journeys between large nodes in the US is domestic air travel. The extreme size of the continental United States and its sparsely populated central region encourages long range flights to connect the two coasts, for example, usually between highly populated nodes. There is much less need for flights over short distances and few closely spaced large population nodes for them to fly between. These kinds of flows, although important for disease transmission, may not be well modelled by power law kernels and gravity models. As they are relatively rare, they carry little statistical ‘weight’ within the likelihood expression, but still play a crucial role in long distance disease transmission.
Comparison between the parameter estimates for the two regions are hard to make, as the levels of aggregation differ. Perhaps the closest match is between US counties and UK districts, both in terms of population size and demographically (large conurbations are made up of several of the units in both cases). As can be seen from tables 1,2 and 3, parameter values are not generally transferable. The exponent of long distance connections is consistent between MK models (approx 1.8), but it is this aspect of the US fit which changes most with the introduction of assortativity and hence is most strongly associated with the poor performance of the MK model in the US. As discussed, the poor fit in the US is probably the result of the demographic heterogeneity and large size of the country. It seems more likely that parameter estimates for the UK may transfer better to other European countries with similar demographies.
It is clear from Figure 6 that although an assortative model considerably improves the fit to the spatio-temporal dynamics of the epidemic, it doesn't have much effect on the underlying biases of the gravity model with regard to the US mobility data. Further work is necessary to identify what features of the data are essential to an accurate model. Preliminary show that dividing the country into regions of similar population density (east and west coasts separated by a central region) leads to better fits within each but requires additional models for movements between them. A recent paper Simini et al. presents a radical new model defining the relative probabilities of trips purely in terms of resident population distributions [20]. Simulations show that this model performs at least as well as the assortative model described in this paper in matching epidemic behaviour. The time to infection theory (SI Text S1) suggests several applications. In meta-population models, it allows a comparison between local epidemic timescale and that of transmission to neighbouring nodes. Hence it is possible to optimise meta-population structure to ensure that individual nodes represent largely independent populations weakly linked to each other. The theory also suggests an alternative approach to clustering on a network in terms of its accelerating effect on the speed of epidemics on a network.
|
10.1371/journal.ppat.1002407 | A Temporal Role Of Type I Interferon Signaling in CD8+ T Cell Maturation during Acute West Nile Virus Infection | A genetic absence of the common IFN- α/β signaling receptor (IFNAR) in mice is associated with enhanced viral replication and altered adaptive immune responses. However, analysis of IFNAR-/- mice is limited for studying the functions of type I IFN at discrete stages of viral infection. To define the temporal functions of type I IFN signaling in the context of infection by West Nile virus (WNV), we treated mice with MAR1-5A3, a neutralizing, non cell-depleting anti-IFNAR antibody. Inhibition of type I IFN signaling at or before day 2 after infection was associated with markedly enhanced viral burden, whereas treatment at day 4 had substantially less effect on WNV dissemination. While antibody treatment prior to infection resulted in massive expansion of virus-specific CD8+ T cells, blockade of type I IFN signaling starting at day 4 induced dysfunctional CD8+ T cells with depressed cytokine responses and expression of phenotypic markers suggesting exhaustion. Thus, only the later maturation phase of anti-WNV CD8+ T cell development requires type I IFN signaling. WNV infection experiments in BATF3-/- mice, which lack CD8-α dendritic cells and have impaired priming due to inefficient antigen cross-presentation, revealed a similar effect of blocking IFN signaling on CD8+ T cell maturation. Collectively, our results suggest that cell non-autonomous type I IFN signaling shapes maturation of antiviral CD8+ T cell response at a stage distinct from the initial priming event.
| Although it is well established that type I IFN responses protect against viral infections by inducing expression of antiviral genes and modulators of adaptive immune responses, its function at different stages of viral infections has remained poorly studied. In this paper, we administered a monoclonal antibody that blocks the common type I IFN signaling receptor to mice at different times after WNV infection to dissect the temporal functions of IFN. Administration of the blocking antibody at day -1 resulted in a massive increase in viral replication and the number of WNV-specific -CD8+ T cells. In contrast, treatment with a single dose of antibody at day 4 had limited effects on viral dissemination, but instead promoted development of dysfunctional CD8+ T cells that produced lower levels of cytokines and expressed proteins implicated in T cell exhaustion. Thus, we show a stage-specific effect of type I IFN in optimal maturation of antiviral CD8+ T cell responses. Our study provides new insight as to how and when innate immune signals affect maturation of antiviral CD8+ T cells after the initial priming event with viral antigen.
| Type I interferons (IFN) comprise a family of cytokines that that were identified originally for their ability to render cells resistant to virus infection [1]. Type I IFN binds to a common IFN-αβ receptor (IFNAR), which initiates a signaling cascade that results in phosphorylation and nuclear translocation of STAT1 and STAT2, and induction of expression of hundreds of interferon-stimulated genes (ISG) [2]. These ISG control viral infections through a diverse range of direct antiviral effector functions [3] and by modulating adaptive immune responses [4].
Type I IFN responses are essential for the controlling infection by West Nile virus (WNV) [5], [6], an encephalitic positive strand RNA virus of the Flaviviridae family that has emerged over the past decade as a significant cause of neuroinvasive disease [7]. IFNAR-/- mice are exquisitely vulnerable to WNV infection, with expanded tissue tropism, uncontrolled viral replication, and rapidly uniform death, with all animals succumbing within four days of infection after inoculation with a single plaque forming unit (PFU) of virus [8].
Apart from its function in controlling viral infection through cell-intrinsic antiviral gene induction, type I IFN has an established role in priming of B and T cell responses (reviewed in [9], [10]). Signaling through IFNAR regulates early innate and adaptive B cell activation in the lymph node and spleen [11]–[13] and induces dendritic cells to mature, express higher levels of co-stimulatory molecules, and present antigen more efficiently, which is required for optimal induction of a functional T cell response (reviewed in [14]). Diminished effector functions of memory CD8+ T cells in IFNAR-/- mice have been described after infection with influenza and vaccinia (VV) viruses [15], [16]. This could be due in part, to defects in cross-priming of CD8+ T cells, which is believed to require both virus-induced type I IFN [9], [13], [17] and CD8-α dendritic cells [18].
Although cell-type and tissue-specific conditional deletions of IFNAR have been described [19]–[22], the function of type I IFN at discrete stages of viral infection remains unknown. To define the temporal functions of type I IFN signaling in the context of infection by WNV, we utilized a previously reported blocking anti-IFNAR monoclonal antibody (MAb MAR1-5A3), which prevented type I IFN-induced intracellular signaling in vitro, was non-cell-depleting, and inhibited antiviral, antimicrobial, and antitumor responses in mice [23].
By administering MAR1-5A3 antibody at different times after viral inoculation, we separated the early innate from the later innate-adaptive functions of type I IFN. Treatment prior to WNV infection resulted in massive expansion of virus-specific CD8+ T cells by day 9. However, blockade of type I IFN signaling beginning at day 4 after WNV infection was associated with defects in virus-specific effector CD8+ T cells at day 9 including depressed IFN-γ and TNF-α responses and changes in phenotypic markers suggesting altered activation status and CD8+ T cell exhaustion that is usually seen during chronic viral infection [24]. This phenotype was not due to direct signaling effects through IFNAR on CD8+ T cells and was also observed after vaccinia virus (VV) infection under similar experimental conditions. Experiments in BATF3-/- mice, which lack CD8-α dendritic cells and have impaired antigen cross-presentation and CD8+ T cell priming capacity, showed a similar effect of temporal blockade of type I IFN signaling on CD8+ T cell maturation. Collectively, our results suggest that cell non-autonomous type I IFN signaling shapes maturation of antiviral CD8+ T cell response at a stage distinct from the initial priming event.
Previous studies established a critical requirement for type I IFN in controlling WNV-NY (strain New York, 2000) as infected IFNAR-/- mice showed expanded tissue tropism, uncontrolled viral replication, and rapidly uniform death within four days [8]. While these experiments suggested a dominant antiviral function of type I IFN in vivo, key roles in modulating adaptive B and T cell responses against viruses also have been described [13], [17]. One caveat of the antiviral and immunologic studies is that they have been performed primarily in complete or cell-type IFNAR-/- mice, which limits insight into the temporal function of IFN signaling in modulating immune responses. Also, because many viruses replicate to substantially higher levels in IFNAR-/- mice, it can be difficult to separate how enhanced antigen burden and lack of type I IFN signaling differentially impact adaptive immune responses in the context of live virus infection. To begin to define the temporal functions of type I IFN signaling, we utilized MAR1-5A3, a previously described MAb that potently blocks type I IFN receptor signaling and is non cell-depleting [23].
IFNAR-/- mice succumb to lethal WNV-NY infection within 4 days of infection after a dose of 102 PFU of virus [8]. We assessed whether treatment with the MAR1-5A3 MAb recapitulated this phenotype. We performed a dose titration of MAR1-5A3, in which mice were treated one day prior to infection with 102 PFU of WNV-NY and monitored for survival (Figure 1A). Similar to IFNAR-/- mice, all wild type mice treated with a single dose of MAR1-5A3 but not the isotype control GIR-208 MAb ranging from 0.3 to 2.5 mg succumbed to WNV-NY infection, although the mean time of death (MTD) was delayed (6.5 days versus 4 days, P<0.0001). Given this data, we chose a MAR1-5A3 dose of 1 mg per mouse for the remainder of the study. The difference in MTD was not unexpected as MAR1-5A3 is not expected to cross the blood-brain-barrier efficiently and type I IFN has direct antiviral effects on neurons in the central nervous system [8], [25], [26].
We hypothesized that type I IFN signaling may have distinct functions at different stages of viral infections. To test this, mice were treated with a single 1 mg dose of MAb at different days after infection and survival was monitored (Figure 1B). We observed a significant difference (P<0.0003) in survival of mice treated with MAR1-5A3 as late as four days after infection as compared to the isotype control MAb treated mice. The MTD after WNV-NY infection for mice receiving MAR1-5A3 between days 0 and 2 was ∼8 days whereas those receiving MAb on days 3 or 4 survived on average between 10 and 11 days.
To further characterize the impact of type I IFN signaling at different stages, we compared viral titers from organs of mice at day 6 after infection in mice treated with the MAR1-5A3 or the control GIR-208 MAb at days 2 or 4 post infection (Figure 1C to G). In mice treated with MAR1-5A3 two days after infection, we observed an increase in viremia (739-fold, P<0.02), and infection in the spleen (242-fold, P<0.02) and kidney (240-fold, P<0.002) compared to the isotype control MAb. This corresponded with markedly higher viral titers in the brain (325-fold, P<0.006) and spinal cord (2,650-fold, P<0.001) compared to the control group. In contrast, mice treated with a single dose of MAR1-5A3 at day 4 after infection showed substantially smaller increases in the spleen (4.4-fold, P<0.03) and brain (13-fold, P<0.006) with no detectable elevation in serum, kidney, or spinal cord (P>0.19) at day 6. Thus, although the relative timing (day 2 or 4) of MAR1-5A3 administration did not differentially affect clinical outcome, it impacted viral spread and tropism; earlier blockade of type I IFN signaling resulted in enhanced replication in all tissues examined, whereas later administration had a small effect in only a subset of organs.
Several groups have observed differences in antibody and CD8+ T cell responses in IFNAR-/- and STAT1-/- mice after infection or vaccination [13], [15]–[17], [19], [27], [28]. Because administration of MAR1-5A3 at day 4 had relatively minor effects on viral burden at day 6 (Figure 1) or day 8 (data not shown) yet still resulted in complete lethality, we hypothesized that blockade of type I IFN receptor signaling at later stages might impact early adaptive immune responses.
The development of an antibody response is critical for surviving WNV infection [29], [30]. To study the temporal effects of type I IFN signaling on the humoral response, wild type mice were infected with WNV-NY, treated with MAR1-5A3 or isotype control antibody two or four days later, and serum was harvested at day 6 or 9 after infection. We detected no statistically significant difference in WNV-specific IgM or IgG response between the MAR1-5A3 and control antibody-treated groups at any of the time points tested (Figure S1, P>0.2). Thus, blockade of type I IFN signaling at day 2 or 4 after infection had no major impact on induction of WNV-specific antibody responses during the acute phase of infection.
CD8+ T cells contribute to the rapid clearance of WNV infection from peripheral and central nervous system (CNS) tissues [31]–[34]. Analysis of CD8+ T cells at day 9 in the spleen of wild type mice treated with the MAR1-5A3 or control antibody at day 4 after infection showed a similar percentage and total number of WNV-specific CD8+ T cells when measured by intracellular IFN-γ and TNF-α staining after ex vivo incubation with an immunodominant Db-restricted NS4B peptide (Figure 2A) or direct tetramer staining (data not shown). Nonetheless, blockade of type I IFN signaling at day 4 resulted in a decrease in the amount of intracellular IFN-γ (P<0.0001) and TNF-α (P<0.006) produced by individual antigen-specific CD8+ T cells as judged by differences in the geometric mean fluorescence intensity of the positive cells. Correspondingly, the amount of granzyme B in NS4B tetramer positive cells was less (P<0.003) in MAR1-5A3 treated mice (Figure 2B). The differences in intracellular cytokines and granzyme B protease establish a late temporal role for type I IFN signaling in the maturation of the antigen-specific CD8+ T cells, even though initial priming, as reflected by the total percentage and number of WNV-specific IFN-γ+ CD8+ or TNF-α+ CD8+ T cells, remained intact.
To determine whether a similar effect on T cell maturation was observed if type I IFN was neutralized throughout infection, we pre-treated mice with MAR1-5A3 prior to infection with an attenuated lineage 2 WNV strain from Madagascar (WNV-MAD) [35], [36]. We used this less virulent WNV strain because mice treated with MAR1-5A3 and infected with WNV-NY did not survive past day 6 (see Figure 1). In comparison, mice treated with MAR1-5A3 before or after infection with attenuated WNV-MAD showed very limited mortality (data not shown). Accordingly, mice were treated with MAR1-5A3 or isotype control mAb one day prior to and four days after infection with WNV-MAD, and CD8+ T cells were analyzed at day 9. Notably, depletion of type I IFN signaling throughout the course of infection resulted in a substantial increase in the percentage (6 to 49%, P<0.008) and number (P<0.008) of WNV-specific IFN-γ+ CD8+ T cells (Figure 2C). Similar results were observed when intracellular TNF-α was measured (Figure 2D). The large increase in CD8+ T cell priming may be attributed to the greater WNV antigen burden in lymphoid tissues in mice lacking type I IFN signaling [8]. However, and consistent with that observed with MAR1-5A3 treatment at day 4 only with WNV-NY infection, the amounts of intracellular IFN-γ and TNF-α present in WNV-specific CD8+ T cells were lower (P<0.008) when type I IFN signaling was blocked throughout infection.
Blockade of type I IFN receptor at day 4 also modulated the CD4+ T response after WNV-NY infection. The percentage of IFN-γ+ or TNF-α+ CD4+ T cells, as measured after ex vivo stimulation with anti-CD3 antibodies, was decreased (P<0.007) in mice receiving MAR1-5A3 compared to the GIR-208 isotype control MAb (Figure S2). While we observed a significant decrease (P<0.004) in the total number of TNF-α+ producing CD4+ T cells in MAR1-5A3 treated mice, this was not observed in IFN-γ+ CD4+ T cells. Analogous to that seen with WNV-specific CD8+ T cells, decreased amounts (P<0.01) of IFN-γ and TNF-α were produced by activated CD4+ T cells in animals treated with MAR1-5A3.
Given that a blockade of type I IFN signaling resulted in WNV-specific CD8+ T cells that expressed lower levels of intracellular cytokines, we speculated that this could be due to an increase in CD4+CD25+FoxP3+ regulatory T cells. Type I IFN has been reported to alter regulatory T cell activity, which impacts CD8+ T cell function [37], [38], and decreased regulatory T cell levels augment WNV-specific CD8+ T cell responses [39]. However, we observed no difference in the percentage or numbers of CD4+CD25+FoxP3+ cells at day 9 in the blood (data not shown) or spleen (P>0.6), when MAR1-5A3 or control GIR-208 MAb was administered at day 4 after WNV-NY infection (Figure S3).
Our data suggested that type I IFN signaling at a later stage modulated WNV-specific T cell responses despite having limited effects on viral replication or initial priming. To determine whether this finding was typical of other viral infections, we repeated MAR1-5A3 treatments at day 4 after infection with an unrelated DNA (VV, Western reserve strain) virus. Mice were harvested eight days after infection (four days after treatment) and T cell populations were analyzed. Similar to that seen with WNV, the amount of intracellular IFN-γ produced by CD8+ T cells from the MAR1-5A3 treated mice was lower (P<0.02) after re-stimulation ex vivo with two different VV-peptides (A47L or B8R) compared to the isotype control GIR-208 treated mice (Figure 3A and B). Notably, and in contrast to WNV infection, we also detected a decrease in the percentage (P<0.02) and number (P<0.04) of IFN-γ producing VV-specific CD8+ T cells, suggesting that for VV infection, type I IFN signaling at day 4 or after also contributed to initial priming. Similar results were observed with TNF-α production with VV-specific CD8+ T cells after MAR1-5A3 treatment (Figure 3A and B). Thus, a temporal blockade of type I IFN signaling impairs antigen-specific CD8+ T cell maturation in the context of infection by WNV and VV, two unrelated RNA and DNA viruses.
Recent studies have suggested that type I IFN enhances the CD8+ T cell response during antigen cross-presentation [17], [22], [40], [41]. To evaluate whether the temporal effect of type I IFN signaling on CD8+ T cell responses occurred in mice with impaired cross-presentation capacity, we utilized BATF3-/- mice, which lack CD8-α and CD103+ dendritic cells [18], [42]. Consistent with earlier results from BATF3-/- 129SvEv mice [18], we observed a decrease in the percentage and number (P<0.008) WNV-specific CD8+ T cells in BATF3-/- mice on the C57BL/6 background although no substantive difference (P>0.06) in intracellular IFN-γ levels was detected (Figure 4A).
To determine whether mice with priming defects due to impaired cross-presentation still required late stage type I IFN for CD8+ T cell maturation, MAR1-5A3 or control GIR-208 MAb was administered to wild type or BATF3-/- mice at day 4 after WNV-NY infection. As expected, associated with the absence of CD8-α dendritic cells, the magnitude (percentage and number) of IFN-γ+ and TNF-α+ NS4B-specific CD8+ T cells at day 9 was markedly lower in MAR1-5A3 or GIR-208 MAb treated BATF3-/- mice compared to wild type animals (data not shown). Nonetheless, reduced intracellular levels of IFN-γ and TNF-α (P<0.009) in WNV-specific CD8+ T cells were still observed in BATF3-/- mice treated with MAR1-5A3 at day 4 compared to control MAb-treated animals (Figure 4B). Thus, the temporal effect of type I IFN blockade on CD8+ T cell maturation occurred both in the presence or absence of CD8-α dendritic cells and efficient antigen cross-presentation.
Studies with IFNAR-/- bone marrow chimera or conditionally deleted IFNAR on T cells showed reduced cross-presentation of ovalbumin peptides to CD8+ T cells, suggesting that direct stimulation of T cells by type I IFN enhances the antigen-specific CD8+ T cell response, at least for soluble antigens [22]. Blockade of type I IFN signaling four days after WNV infection results in a dysfunctional antigen-specific CD8 T cell population that nonetheless appeared to undergo a relatively normal priming phase. In comparison, MAR1-5A3 treatment at days -1 and 4 (essentially throughout infection) resulted in a dysfunctional antigen-specific CD8+ T cell population, but with a massive increase in the fraction and number of antigen-specific T cells. To establish whether the effect of type I IFN on CD8+ T cell functional development was cell-intrinsic in the context of viral infection, we adoptively transferred naïve purified IFNAR-/- (CD45.2) or B6.SJL (CD45.1) CD8+ T cells into RAG1-/- recipient mice. Immediately after WNV infection, blood was sampled to confirm transfer of T cell populations in the recipient mice (data not shown). At day nine after infection, spleens were harvested and the CD8+ T cell activation profiles analyzed. Notably, we did not detect a significant difference (P>0.06) in the intracellular levels of IFN-γ or TNF-α between the IFNAR-/- (CD45.2) and B6.SJL (CD45.1) CD8+ T cells donor cells in the IFNAR+/+ RAG1-/- mice (Figure S4). This result suggests that at least in the context of WNV infection, the effect of type I IFN on the development of a functional CD8+ T cell response is largely T cell non-autonomous in nature.
As our adoptive transfer experiments suggested that efficient WNV-specific CD8+ T cell activation did not require cell autonomous type I IFN signaling in CD8+ T cells, we assessed whether antigen-presenting cells in the spleen were differentially affected by MAR1-5A3 treatment. MAR1-5A3 was administered 2 or 4 days after WNV infection and APC were examined on days 6 and 9 after infection (Figure 5A, B, and C). When MAR1-5A3 was given on day 2 and splenocytes analyzed on day 6, no difference was observed in the percentage of CD11c+ cells or their relative expression of the co-stimulatory molecules CD80 and CD86 (Figure 5B). We did however, observe an increased percentage of CD11b+ splenocytes at this time point, and this was associated with reciprocal decreases and increases in expression of CD80 and CD86, respectively. In comparison, when MAR-5A3 was administered on day 4 after WNV-NY infection and splenocytes analyzed on day 6, we observed a reduced percentage of CD11b+ (P<0.01) and CD11c+ (P<0.02) cells, and this was associated with decreased expression of CD86 only on CD11c+ cells (Figure 5B, P<0.008). When MAR-5A3 was administered on day 4 after WNV-NY infection and splenocytes analyzed on day 9, we also observed a decrease in surface expression of CD86 on CD11b+ (P<0.05) and CD11c+ (P<0.007) cells relative to the control MAb treatment (Figure 5A and C). In comparison, MAR1-5A3 treatment had no effect on CD80 expression on CD11c+ cells although an increase (P<0.005) was noted in CD11b+ cells at this time. Thus, blockade of type I IFN signaling at day 4 after infection resulted in a distinct antigen-presenting cell activation phenotype compared to MAR1-5A3 treatment at day 2; this suggests that disruption of type I IFN signaling pathways at particular stages of infection might limit the ability of antigen-presenting cells to provide key temporal signals that allow optimal generation of antigen-specific effector CD8+ T cells.
We speculated that a specific absence of type I IFN signaling in amtigen-presenting cells impaired development of a WNV-specific CD8+ T cell response because of an altered production of cytokines required for maturation. To assess this, we measured the cytokine levels in mice that were treated with MAR1-5A3 at day 4 after WNV-NY infection. Two or five days after MAb treatment (day 6 or 9 after infection), serum was harvested and levels of relevant cytokines (IFN-γ, TNF-α, IL-10, IL-12 p40, IL-17, and IL-18) were measured by bioplex assay (Figure 6A–F). Two days after MAR1-5A3 treatment, significantly (P<0.04) reduced levels of IL-12 p40 were observed (Figure 6D). Within five days of MAR1-5A3 treatment, serum levels of IFN-γ, TNF-α, and IL-12 p40 were reduced significantly (P<0.01) and IL-10 was elevated (P<0.02). The increased level of IL-10 in mice treated with the blocking type I IFN MAb may be particularly relevant as IL-10 negatively impacts CD8+ T cell activation and function [37], [38].
Because blockade of IL-10 in chronic lymphocytic choriomengitis virus (LCMV) infection prevents functional exhaustion of CD8+ T cells and promotes viral clearance [43], [44], we hypothesized that the increased IL-10 levels in serum of MAR1-5A3 treated mice after WNV-NY infection might cause the CD8+ T cells to acquire an exhausted phenotype. To assess this, at day 5 after MAb treatment (day 9 after infection), we profiled Db-NS4B-tetramer+ CD8+ T cells for expression of PD-1, CTLA-4, CD43, CD44, CD127, and CD11a (Figure 7A). Notably, treatment with MAR1-5A3 compared to the control MAb resulted in reduced expression of CD11a (P<0.001) and increased expression of CD127, CD43, CD44, CTLA-4 and PD-1 (P<0.007). Thus, CD8+ T cells from mice treated at day 4 with MAR1-5A3 not only showed altered intracellular cytokine patterns (see Figure 2) but also displayed some of the phenotypic hallmarks of exhaustion. Similarly, BATF3-/- mice treated with MAR1-5A3 at day 4 after WNV infection also expressed elevated (P<0.02) levels of the exhaustion markers CTLA-4 and PD-1 on WNV-specific CD8+ T cells at day 9 compared to control MAb (Figure 7B). Whereas prior studies described CD8+ T cell exhaustion at later time points during chronic LCMV infection [24], [45], blockade of type I IFN signaling independent of the mode of priming appears to exhaust WNV-specific CD8+ T cells during the acute effector phase.
One of the earliest stages of CD8+ T cell exhaustion is characterized by a reduced capacity to lyse target cells [45], [46]. Although we observed reduced levels of granzyme B in Db-NS4B tetramer positive CD8+ T cells (Figure 2), we questioned whether WNV-specific effector cells during the acute immune response displayed a fully exhausted phenotype. We assessed how MAR1-5A3 treatment affected CD8+ T cells ability to lyse peptide pulsed targets in vivo (Figure S5). Splenocytes from naïve B6.SJL (CD45.1) mice were divided into two groups: one group was pulsed with NS4B immunodominant peptide and labeled with 500 nM carboxyfluorescein diacetate succinimidyl ester (CFDA), and the other was not pulsed with peptide and labeled with 5 nM CFDA. The two groups were mixed in equal numbers and injected into WNV-infected C57BL/6 (CD45.2) mice at day 9 that had undergone treatment with either MAR1-5A3 or control GIR-208 MAb at day 4. Six hours after labeled cells were transferred, splenocytes were harvested and in vivo killing was assessed. Notably, we observed no difference in killing between the MAR1-5A3 and the control MAb-treated mice (P>0.3). Thus, type I IFN blockade at a later stage of WNV infection produces an intermediate exhaustion phenotype with skewed cytokine production, surface expression of exhaustion markers, yet relatively intact cytolytic potential.
In this study, we evaluated the antiviral and immunomodulatory roles of type I IFN signaling after viral infection. While past studies in IFNAR-/- mice with virulent or attenuated WNV strains revealed enhanced susceptibility, dissemination, and lethality compared to congenic wild type mice [8], [36], [47], they did not address the temporal functions of type I IFN during infection. While administration of MAR1-5A3 at day 2 after infection resulted in markedly enhanced viral burden in multiple tissues as seen in IFNAR-/- mice [8], treatment at day 4 had more subtle effects on viral replication. Instead, detailed analysis established a key role for later type I IFN signaling in the maturation of effector CD8+ T cells. Blockade of type I IFN signaling at day 4 after infection with WNV resulted in depressed cytokine responses and changes in phenotypic markers suggesting altered activation and exhaustion.
Prior studies have reported that type I IFN signaling primes adaptive immune functions including cross-presentation of CD8+ T cells, enhancement of antibody responses, and maintenance of dendritic cells in a state competent for antigen-presentation [9], [13], [17], [48]. Depending on the experimental system, type I IFN can act directly on CD8+ T cells or indirectly on antigen-presenting cells to influence the fate of CD8+ T cells during the initial phases of antigen recognition (reviewed in [49]). Many of these studies used IFNAR-/- mice [50], adoptive transfer of wild type or IFNAR-/- immune cells into IFNAR-/- or wild type mice [27], or cell-type specific deletion of IFNAR [51]. While they have provided significant insight into the immunomodulatory effects of type I IFN and defined key cells involved in priming, they have not elucidated the stage-specific effects of type I IFN. In our experiments, when type I IFN signaling was blocked with MAR1-5A3 prior to infection with an attenuated WNV strain, we observed at day 9 paradoxically enhanced numbers of antigen-specific effector CD8+ T cell responses that had deficits in IFN-γ or TNF-α production, results that are consistent with prior infection experiments [52]. The increased numbers of WNV-specific CD8+ T cells in mice treated with MAR1-5A3 at day -1 could be due to increased antigen burden or a failure to produce IL-10 and negatively regulate T cell expansion [37].
Administration of a single dose of MAR1-5A3 at day 4 after infection with virulent or attenuated WNV strains revealed a distinct phenotype. Although the absolute percentage and number of NS4B-specific CD8+ T cells was similar compared to isotype MAb-treated or unmanipulated animals, the geometric mean fluorescence intensity of IFN-γ or TNF-α was consistently lower. Thus, in the context of WNV infection, the initial priming phase of virus-specific CD8+ T cells does not absolutely require type I IFN signaling whereas the later maturation phase does. In addition, MAR1-5A3 treatment on day 4 was associated with lower granzyme B expression, decreased surface levels of the adhesion molecule CD11a (LFA-1), and increased expression of CD44, CD127 (IL-7R α-chain), and CD43 on WNV-specific CD8+ T cells. These markers are significant because in mice activated, lytic CD8+ T cells display a CD44hi CD43hi CD127lo granzyme Bhi phenotype whereas memory CD8+ T cells express a CD44hi CD43lo/int CD127hi granzyme Blo phenotype [53]-[55]. Thus, stage-specific blockade of type I IFN signaling alters intracellular cytokine production of antigen-specific CD8+ T cells and promotes a transitional phenotype during the acute (day 9) phase that appears to fall somewhere between effector and memory populations.
Consistent with functionally dysregulated CD8+ T cells when type I IFN signaling was blocked at day 4, we observed increased expression of PD-1 and CTLA-4, two markers of T cell exhaustion [24], [56], which were originally described in the context of chronic, persistent infection of LCMV [46]. In chronic LCMV infection, there is a hierarchy to CD8+ T cell exhaustion with some functions exhausted early (IL-2 production, cytotoxicity, and proliferation) and others persisting longer (intracellular pro-inflammatory cytokines) [45]. In comparison, blockade of type I IFN signaling at day 4 resulted in WNV-specific CD8+ T cells at day 9 that retained the ability to kill targets in vivo although they expressed lower quantities of IFN-γ and TNF-α. Thus, stage-specific blockade of type I IFN results in dysfunctional CD8+ T cells with loss of some but not all effector functions during the acute phase. Although we cannot address what happens during later stages (evolution and maintenance of memory CD8+ T cells) in the context of type I IFN blockade and virulent WNV-NY infection because of complete lethality in the model, kinetic studies are planned with the attenuated WNV-MAD strain and MAR1-5A3 to determine how and when type I IFN signaling affects the transition to and establishment of memory phenotypes.
The dysfunctional CD8+ T cell phenotype observed after MAR1-5A3 treatment and WNV infection also was observed after VV infection. The change in CD8+ T cell profile with type I IFN blockade at day 4 was even more marked after VV infection, as the percentage, number, and mean fluorescence intensity of antigen-specific CD8+ T cells were all significantly reduced at day 9 for two independent immunodominant epitopes. Thus, for VV, late stage type I IFN blockade affected both priming and subsequent maturation.
Cross-priming of CD8+ T cells occurs after dendritic cells pick up soluble molecules or cellular debris [57] and are licensed by additional cellular or inflammatory signals [58]. Although type I IFN can license dendritic cells for cross-priming of CD8+ T cells with soluble ovalbumin [17], it remains unknown if it is essential in the context of the inflammatory milieu associated with viral infection. We speculated that stage-specific blockade of type I IFN signaling might have dominant effects on CD8+ T cells maturation if CD8-α dendritic cells and cross-presentation were required for priming and activation. To evaluate this, we infected BATF3-/- mice, which lack CD8-α dendritic cells, are defective in antigen cross-presentation, and fail to optimally prime CD8+ T cell responses [18]. While the percentage and number of WNV-specific IFN-γ+ CD8+ T cells was blunted in BATF3-/- mice, the remaining CD8+ T cells that were presumably primed by a distinct antigen presentation pathway showed reduced intracellular cytokine levels and enhanced expression of CTLA-4 and PD-1. Thus, at least during WNV infection, the temporal effects of type I IFN signaling on effector CD8+ T cell maturation occur regardless of the initial priming pathway.
Although prior studies have suggested that direct stimulation of T cells by type I IFN enhances ovalbumin-specific CD8+ T cell responses during cross-priming [22], we did not observe this in the context of WNV infection. CD45.2 CD8+ T cells lacking IFNAR showed roughly equivalent WNV-specific responses compared to congenic CD45.1 CD8+ T cells after transfer into and infection of RAG1-/- recipient mice. An analogous small impact of direct stimulation by type I IFN on CD8+ T cells was observed after infection with VV [59] but not with LCMV [27], [50]. The differential requirement for direct signaling on CD8+ T cells may be due to differences in local and systemic type I IFN production during infection with different pathogens [50].
Blockade of type I IFN at day 4 after WNV infection was associated with decreased expression of CD86 on antigen-presenting cells, which likely influences optimal antigen presentation to CD8+ T cells [14], [60]. Indeed, lower levels of pro-inflammatory dendritic cell-produced cytokines (IL-12) [61] that regulate CD8+ T cell expansion and activation state were observed in mice treated with MAR1-5A3 at day 4. Alternatively, blockade of type I IFN signaling at day 4 could affect CD8+ T cell activation because of the elevated levels of the inhibitory cytokine IL-10. Although our results point to a critical temporal role of type I IFN signaling in the functional activation of CD8+ T cells in the context of infection by WNV, future studies are required to define the precise spatial and cell-type specific cues that govern this process.
The administration of a neutralizing anti-IFNAR antibody at day 2 after infection limited the ability of the host to control WNV replication and spread to target tissues, thus confirming a dominant antiviral effect of type I IFN during the early stages of pathogenesis. In comparison, administration of the anti-IFNAR antibody at day 4 after WNV infection had marginal effects on viral replication, no effect on the magnitude of CD8+ T cell priming, yet profoundly impacted the functional CD8+ T cell responses during the acute effector phase, resulting in blunted cytokine production, and changes in phenotypic markers associated with altered activation status and CD8+ T cell exhaustion. Given that several studies have established a protective clearance role for CD8+ T cells in the brain after WNV infection with virulent North American strains [31], [33], [34], [62], it is not surprising that a temporally defective type I IFN response that affects optimal CD8+ T cell maturation resulted in enhanced lethality.
Future studies that administer neutralizing antibodies against IFNAR, other individual IFN subtypes, or other anti- or pro-inflammatory cytokines at different phases of acute virus infection may reveal stage-specific requirements for shaping effector CD8+ T cells, the contraction phase, and the transition to central and effector memory. Such studies, coupled with experiments in mice with cell-type specific deletions of IFNAR, will provide new insight into the spatial-temporal dynamics of CD8+ T cell expansion and development during infection by different viruses.
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 at the Washington University School of Medicine (Assurance Number: A3381-01). All inoculation and experimental manipulation was performed under anesthesia that was induced and maintained with ketamine hydrochloride and xylazine, and all efforts were made to minimize suffering.
The lineage 1 WNV strain 3000.0259 (WNV-NY) was isolated in New York in 2000 [63] and passaged twice in C6/36 Aedes albopictus cells. The lineage 2 WNV strain from Madagascar (DakAnMg798, WNV-MAD) was isolated in 1978 and passaged on C6/36 cells [35]. BHK21-15 cells were used for plaque assay experiments with WNV. VV (Western Reserve) was grown in Vero cells and purified by ultracentrifugation through a 36% sucrose cushion.
Wild type and RAG1-/- C57BL/6 mice were obtained commercially (Jackson Laboratories). C57BL/6.SJL-Ptprca/BoyAiTac (B6.SJL, CD45.1) mice were purchased (Taconic). IFNAR-/- mice were obtained from J. Sprent (Scripps Institute, San Diego CA) and were backcrossed ten times onto the C57BL/6 background. BATF3-/- mice [18] were backcrossed onto a C57BL/6 background for ten generations. All mice were housed in the pathogen-free mouse facility at the Washington University School of Medicine. Mice (8 to 12 week-old) were inoculated subcutaneously via footpad injection with 102 plaque-forming units (PFU) of WNV-NY or WNV-MAD. VV (104 PFU) was injected via an intraperitoneal route. MAR1-5A3 (mouse anti-mouse IFNAR, IgG1) or isotype control GIR-208 (mouse anti-human IFN-γ receptor 1, IgG1) MAbs [23] were administered as a single dose at 1 mg per mouse unless otherwise indicated by intraperitoneal (IP) injection at specific times with respect to viral infection. MAR-5A3 and GIR-208 MAbs were purchased (Leinco Technologies) and certified as free of endotoxin contamination and aggregates. The half-life of MAR1-5A3 is reported as 5.2 days when a sufficient amount is administered to saturate the receptor pool [23].
For analysis of viral burden MAR1-5A3 or GIR-208 was administered two or four days after infection, and organs were recovered on day 6 after cardiac perfusion with 10 ml of PBS. Tissues were weighed, homogenized using a bead-beater apparatus, and titrated for WNV by plaque assay on BHK21-15 cells as described previously [29]. Serum was obtained from whole blood after phlebotomy of the axillary vein immediately before sacrifice and viremia was measured by analyzing WNV RNA levels using fluorogenic quantitative RT-PCR (qRT-PCR) as described [25].
WNV-specific IgM and IgG levels were determined using an envelope (E) protein–specific ELISA as described [64].
Intracellular staining of TNF-α and IFN-γ from splenocytes was performed as described previously [33]. Briefly, spleens were harvested and homogenized to form a single cell suspension. Cells (2×106 cells) were added to a 96 well plate and incubated with 2 µg/ml brefeldin A (Sigma) for 6 h at 37°C with 10−6 M of immunodominant T cell peptides (WNV: Db-restricted NS4B 2488–2496 (SSVWNATTA) [33]; and VV: Kb-restricted A47L 138–146 (AAFEFINSL) and B8R 20–27 (TSYKFESV) [65]) or 2 µg/ml anti-CD3 (145-2C11) (BD Biosciences). After incubation, the cells were stained with directly labeled antibodies (all from BD Biosciences unless indicated) against CD4 (GK1.5), CD19 (6D5), CD43 (1B11), CD127 (SB/199), CD8β (YTS156.7.7), CD44 (MI7), PD-1 (RMP1-30), and CTLA-4 (UC10-4B9, Biolegend). Db-NS4B tetramer was obtained from the NIH tetramer core facility. Cells were washed, fixed, and permeabilized with FixPerm Buffer (eBioscience), and stained intracellularly for anti-IFN-γ (XMG1.2), anti-TNF-α (MP6-XT22, eBioscience), or anti-granzyme B (GB12, Invitrogen). Lymphocytes were processed on an LSRII (BD Bioscience) using FACSDiva 6.1.1 software (BD Bioscience) and analyzed with FlowJo (Treestar). The total numbers of IFN-γ or TNF-α expressing CD4+ or CD8+ T cells was determined by multiplying the percentage of IFN-γ+ or TNF-α+ CD4+ or CD8+ T cells by the total numbers of splenocytes. CD4+CD25+FoxP3+ regulatory T cells were measured using a specific staining kit (eBioscience) following manufacturer's protocol.
The cytokine bioplex assay was performed on serum samples from mice at day 6 and day 9 post-infection from WNV-infected mice that had received either MAR1-5A3 or GIR-208 (1 mg/mouse) at day 4 after infection. The BioPlex Pro Assay was performed according to the manufacturer's protocol (BioRad). The cytokine screen included IL-2, IL-4, IL-10, IL-12p40, IL-12p70, IL-15, IL-17, IL-18, IFN-γ, and TNF-α.
In vivo killing of target cells was performed as previously described [66]. Briefly, splenocytes from B6.SJL (CD45.1) mice were isolated. Half of the cells were labeled with carboxyfluorescein diacetate succinimidyl ester (CFDA) at 500 nM and the remainder was labeled with 5 nM CFDA. After labeling, cells labeled with 500 nM CFDA were pulsed for one hour at 37°C with 1 µM NS4B 2488–2496 peptide, whereas the 5 nM CFDA cells were not pulsed with peptide. Both sets of cells were counted and equal numbers were mixed and injected intravenously (107 cells total per mouse) into recipient WNV-infected (at day 9 after infection) or naïve mice that had received either MAR1-5A3 or GIR-208 (1 mg/mouse) at day 4 post-infection. After 8 hours, the mice were sacrificed and splenocytes were gated on CD45.1 cells (donor cells). The percent killing of target cells was calculated: (1 – (ratio immune/ratio naive)) x 100. Ratio equals the number of NS4B peptide-coated targets/number of reference targets [67].
For survival analysis, Kaplan-Meier curves were analyzed by the log rank test. Statistical significance of viral burden, antiviral antibody titers, and number of activated T cells were analyzed by the Mann-Whitney test. All statistical analysis was performed using Prism software (GraphPad Prism).
|
10.1371/journal.pntd.0005827 | Development of risk reduction behavioral counseling for Ebola virus disease survivors enrolled in the Sierra Leone Ebola Virus Persistence Study, 2015-2016 | During the 2014–2016 West Africa Ebola Virus Disease (EVD) epidemic, the public health community had concerns that sexual transmission of the Ebola virus (EBOV) from EVD survivors was a risk, due to EBOV persistence in body fluids of EVD survivors, particularly semen. The Sierra Leone Ebola Virus Persistence Study was initiated to investigate this risk by assessing EBOV persistence in numerous body fluids of EVD survivors and providing risk reduction counseling based on test results for semen, vaginal fluid, menstrual blood, urine, rectal fluid, sweat, tears, saliva, and breast milk. This publication describes implementation of the counseling protocol and the key lessons learned.
The Ebola Virus Persistence Risk Reduction Behavioral Counseling Protocol was developed from a framework used to prevent transmission of HIV and other sexually transmitted infections. The framework helped to identify barriers to risk reduction and facilitated the development of a personalized risk-reduction plan, particularly around condom use and abstinence. Pre-test and post-test counseling sessions included risk reduction guidance, and post-test counseling was based on the participants’ individual test results. The behavioral counseling protocol enabled study staff to translate the study’s body fluid test results into individualized information for study participants.
The Ebola Virus Persistence Risk Reduction Behavioral Counseling Protocol provided guidance to mitigate the risk of EBOV transmission from EVD survivors. It has since been shared with and adapted by other EVD survivor body fluid testing programs and studies in Ebola-affected countries.
| The 2014–2016 West Africa Ebola Virus Disease (EVD) epidemic was large and widespread, affecting thousands of people across Guinea, Liberia, and Sierra Leone. Prior to this epidemic, there were limited data on persistence of Ebola virus in body fluids of EVD survivors and the potential risk that viral persistence may pose for Ebola virus transmission, including possible sexual transmission. This paper documents the development and implementation of a behavioral counseling protocol to facilitate adoption of risk reduction behaviors among male and female EVD survivors enrolled in the Sierra Leone Ebola Virus Persistence Study. This behavioral counseling protocol, composed of pre-test, delivery of results, and post-test counseling, enabled study staff to translate the study’s body fluid test results into individualized information and preventive action for study participants. Risk reduction behavioral counseling became an important EVD epidemic control measure.
| The 2014–2016 West Africa Ebola Virus Disease (EVD) epidemic was large and widespread, with at least 28,616 probable and suspected cases of EVD and 11,310 deaths across Guinea, Liberia, and Sierra Leone [1]. Limited studies during outbreaks prior to 2014 showed evidence of persistent Ebola virus (EBOV) in semen [2]. In the 1995 Kikwit outbreak in the Democratic Republic of Congo, viable virus in semen was detected 82 days post symptom onset and EBOV ribonucleic acid (RNA) was detected 101 days post symptom onset [3]. Less evidence was found for short-term EBOV persistence in vaginal and rectal fluids as well as urine and sweat [2]. In West Africa, EVD survivors who recovered from disease were initially advised by the World Health Organization (WHO) to abstain from sex or to use condoms for at least three months after discharge from an Ebola Treatment Unit (ETU) [4]. As the number of individuals who survived EVD in West Africa grew into the thousands, questions were raised regarding EBOV persistence in semen and in other body fluids of EVD survivors, and potential implications of viral persistence on residual risk of EBOV transmission from EVD survivors beyond the original three month recommendations [2, 5, 6].
Given limited pre-existing data, the potential for EBOV to persist in survivors was identified as an important research topic early in the epidemic. There was also a programmatic need to develop methods to test body fluids of EVD survivors and counsel them on risk reduction practices. This need became even more clear when a woman from Liberia tested positive for EBOV infection, and epidemiologic investigation found that her only exposure was unprotected vaginal intercourse with a male EVD survivor whose semen tested positive for EBOV RNA by real time reverse transcriptase polymerase chain reaction (qRT-PCR) 199 days after he first became symptomatic with EVD [7, 8]. Sequencing of the RNA from the semen of the male EVD survivor closely matched the sequence recovered from the female patient’s blood, providing further evidence for male-to-female sexual transmission of EBOV long into convalescence [8].
Prior to the EVD epidemic, the Demographic Health Survey conducted in 2013 in Sierra Leone showed that condom use was low, with 5% of female and 13% of male respondents reporting having used a condom in the past 12 months. Twenty-five percent of men and 6% of women reported having two or more sex partners in the last 12 months, and 23% of men and 5% of women reported concurrent sexual partnerships. Multiple and concurrent sexual partnerships were highest among older married men with low education [9]. Human Immunodeficiency Virus (HIV) prevalence was 1.5% among adults aged 15–49 years [9]. There is limited information regarding the impact of the EVD epidemic on sexual risk behavior, but adolescent girls in Sierra Leone reported more unplanned pregnancies and engagement in transactional sex in the nine months during the epidemic when public schools were closed [10]. In addition, EVD survivors reported stigma and feelings of bereavement similar to those experienced by people living with HIV, possibly negatively affecting both their quality of life and intimate relationships, and their motivation to seek healthcare and other services [11–13]. The impact of the EVD epidemic on the sexual behavior of EVD survivors has not yet been fully explored.
In May 2015, the Sierra Leone Ebola Virus Persistence Study was launched to investigate EBOV persistence in the body fluids of EVD survivors in Sierra Leone [5]. The study consisted of two phases. The first phase assessed EBOV persistence in semen of 100 adult male EVD survivors, and the second phase assessed EBOV persistence in semen and additional body fluids (vaginal fluid, menstrual blood, urine, rectal fluid, sweat, saliva, tears, and breast milk as applicable by sex) in 120 male and 120 female EVD survivors. Male and female EVD survivors living with HIV were also invited to participate in the study to characterize EBOV persistence among EVD survivors living with HIV. The study took place in two sites: Military Hospital 34 (an urban facility in Freetown, Western District) and Lungi Government Hospital (a semi-rural facility in Lungi, Port Loko District). In this paper, we discuss the development and implementation of the Ebola Virus Persistence Risk Reduction Behavioral Counseling Protocol (henceforth referred to as the behavioral counseling protocol) used in the study. The test results from this study will be published separately and the overall study design has been described elsewhere [5].
The objectives of the behavioral counseling protocol within the study were to: (1) provide participants simple explanations of qRT-PCR and virus isolation testing and deliver individual test results; and (2) encourage participants to engage in risk reduction behavioral practices corresponding to their individual qRT-PCR test results until the participant received two consecutive negative qRT-PCR test results. Counselors also referred participants to available EVD survivor services in the community when necessary.
Although the authors note the importance of acquiring information on the persistence of EBOV in various body fluids within the pediatric population, for ethical reasons, this study was limited to adults aged 18 years or older. All participants provided written informed consent at the first study visit. The study was approved by the Sierra Leone Ethics and Scientific Review Committee and the WHO Ethical Review Committee (No. RPC736).
We sought to reduce sexual exposure to EBOV by behavior change strategies such as abstinence, increased condom use, and choice of partners. To do this, we developed a behavioral counseling protocol adapted from the Project RESPECT Brief Counseling intervention, an individual face-to-face counseling model that has been proven effective at reducing both new sexually transmitted infections (STIs) and risky behaviors in randomized controlled trials [14]. RESPECT’s Brief Counseling was implemented in the context of Human Immunodeficiency Virus (HIV) testing, and included two 20-minute counseling sessions (i.e. pretest/posttest) in which the counselor supported risk reduction behaviors by increasing the client’s perception of personal risks, emphasizing self-efficacy and personalized goal setting through identifying concrete, incremental and achievable risk-reduction steps that limited sexual HIV/STI exposure. STI clinic patients were asked to describe their own sexual risk behaviors, and misconceptions about risk were clarified. Counselors supported clients in identifying personal barriers to risk reduction and possible ways to overcome them, and helped clients to identify and negotiate behavioral risk reduction steps relevant to their personal risk behaviors. RESPECT’s Brief Intervention model has been previously adapted in behavior change interventions aimed at reducing STIs and pregnancy in vulnerable populations residing in high HIV prevalence settings [15, 16, 17, 18]. The intervention appears to be particularly effective among those with limited exposure to HIV/STI prevention guidance [15, 16, 17, 18]. In sum, the rationale for using a RESPECT model for reducing sexual EBOV exposure was that the approach (1) has demonstrated efficacy at reducing behaviors relevant to sexual EBOV transmission (e.g., increasing abstinence, increasing condom use, reducing risky sexual partnerships), (2) uses a flexible approach that meets the client at his or her own understanding, (3) has been shown to be effective using counselors that are trained in adherence to the model but do not have advanced degrees in counseling, and (4) has been successfully adapted in many international settings with varying cultural contexts.
The adaptation of Project RESPECT for male and female EVD survivors enrolled in the Sierra Leone Ebola Virus Persistence Study required several changes to HIV/STI prevention behavioral guidance. These changes were made in consultation with HIV service providers operating in Sierra Leone [e.g. Sierra Leone National AIDS Control Programme, National AIDS Secretariat, Dignity Association, Joint United Nations Programme on HIV/AIDS (UNAIDS)].
The standard operating procedure for the behavioral counseling protocol included EBOV persistence pre-and post-test counseling as well as HIV pre-and post-test counseling. EBOV persistence pre-test counseling is an introduction to EBOV testing and risk reduction advice prior to testing, while post-test counseling is a delivery of tailored guidance based on qRT-PCR test results. Fig 1 shows the typical participant visit flow.
Table 1 shows the behavioral risk reduction guidance corresponding to each body fluid tested. In the event of a positive qRT-PCR test result, relevant guidance was delivered to the participant in order to reduce transmission risk. All standard operating procedures associated with the delivery of guidance at pre-test and post-test counseling can be found in the full Ebola virus persistence study behavioral counseling protocol here: S1 Protocol.
As prior evidence showed that EBOV could persist in semen [2], male EVD survivors were encouraged at pre-test counseling to use condoms or engage in abstinence for the prevention of EBOV transmission as well as prevention of HIV/STI transmission and unwanted pregnancy. Given limited or no prior evidence for EBOV persistence in vaginal fluids, menstrual blood, urine, rectal fluids, sweat, saliva, tears, and breast milk, precautionary guidance was not provided at pre-test counseling for any of these body fluids. Women were advised at pre-test counseling to use condoms or abstain from sexual intercourse to prevent HIV/STI acquisition and unwanted pregnancy [2]. Female participants were given sexual risk reduction guidance to prevent EBOV transmission only in the event of a positive qRT-PCR test result.
Sexual risk reduction behavioral guidance was tailored for each type of body fluid. Participants were given different guidance for fluids that posed a greater risk for sexual transmission of EBOV due to contact during sexual activity (semen, vaginal fluids, menstrual blood, rectal fluid, urine, sweat, saliva) as compared to fluids for which there was likely to be less contact during sexual activity (breast milk, tears) [2]. Sexual risk reduction guidance was tailored to participants’ religious and cultural practices, particularly around genital washing. For example, counselors reported that Muslim participants preferred to wash the genital area or entire body after sexual intercourse, while Christian participants preferred to wipe genital areas with tissues after sexual intercourse.
For non-sexual risk reduction behavioral guidance, the study team reviewed general infection prevention and control (IPC) guidelines for other pathogens that are spread through contact with infectious body fluids such as viral meningitis and hepatitis B (http://www.cdc.gov/meningitis/viral.html). This guidance was modified for a low resource setting, taking into consideration limited access to clean water and sustained power sources, and lack of flushing toilets, bleach, or other materials and practices used in standard IPC protocols.
Special consideration was given to shaping guidance for female EVD survivors. A focus group with female EVD survivors was held prior to the enrollment of women in the second phase of the study. Participants discussed experience with stigma due to their status as an EVD survivor, fear of resuming sexual activity, and concerns that their partner would not accept condom use. Information from this focus group was shared with study counselors as preparation for discussions with participants regarding specimen collection, sexual activity, and condom use among female participants. Given high rates of sexual violence among women in Sierra Leone [9], a referral pathway for intimate partner violence (IPV) was established in accordance with IPV service provision guidelines from the government of Sierra Leone. Additionally, if a breast milk specimen of a breastfeeding female survivor were to test positive for EBOV, she would be immediately asked to switch to replacement feeding using ready-to-use infant formula provided free of charge with support from UNICEF, and receive ongoing counseling and support from Sierra Leone MoHS Food and Nutrition Directorate staff or a trained nurse.
Given the complex nature of the behavioral guidance for all body fluids, particularly for semen and other intimate fluids, counselors with a background in HIV/STI and/or mental health nursing with previous experience with the EVD epidemic were recruited for the study. At Military 34 Hospital, one female HIV/STI nurse and one male mental health nurse were chosen. At Lungi Government Hospital, one woman with a background in HIV/STI nursing and one woman with a background in midwifery were chosen.
For the first phase of the study, we developed a four-day training package that focused on the science of EBOV transmission and diagnostic testing methods, how to effectively discuss sexual behavior and risk reduction, and common mental health/psychosocial issues reported by EVD survivors. We held an additional four day-training focused on HIV testing and counseling. For the second phase of the study, we added a four-day training to address behavioral guidance for additional body fluids, and to assess for IPV among female study participants in accordance with national and international guidelines [20]. These trainings were primarily geared towards study counselors but all study staff were invited to attend trainings relevant to their roles.
The counselors used role-playing to practice using counseling scripts that had been translated into Krio language and then back-translated for the first phase of the study. The Sierra Leone National AIDS Control Programme, Dignity Association, and UNAIDS provided training on HIV testing and counseling, condom demonstrations, and stigma and discrimination. During and after the training sessions, suggestions from the counselors were incorporated in the counseling scripts and behavioral guidance.
The development and implementation of a novel Ebola Virus Persistence Risk Reduction Behavioral Counseling Protocol as part of the Sierra Leone Ebola Virus Persistence Study was an important part of translating semen testing science into individualized, preventive guidance for study participants. It presented a unique opportunity to provide health education and test-based risk reduction services to EVD survivors at a time when these services were not available in a national programmatic setting. Baseline data from the first phase of the study demonstrated that 49% of phase 1 participants had positive baseline qRT-PCR results, and that EBOV RNA could persist in the semen of male EVD survivor participants for at least 270 days (9 months) post-symptom onset [5]. Results from subsequent phases of the study will be published separately. Although we had limited pre-existing information regarding the infectiousness and transmission potential of qRT-PCR positive body fluids, the behavioral counseling protocol became a very important tool to translate laboratory test results into direct guidance for the study participants to potentially prevent EBOV transmission.
EVD survivors in West Africa have reported experiencing a myriad of difficulties including stigma from family, friends, and community members, loss of employment, housing, and social networks, and numerous mental or physical health sequelae [11, 13, 21, 22, 23, 24]. The behavioral counseling protocol was developed to provide EVD survivor participants a safe place to discuss often complex concerns regarding EBOV persistence, semen testing, and other issues related to EVD survivorship. It is possible that positive experiences with the behavioral counseling protocol contributed to participants’ willingness to maintain continued participation in semen testing activities, in addition to transportation compensation, access to referral medical care, and receipt of critical health information.
We observed great receptivity to condom demonstrations, with many participants informing research staff that they had not participated in a condom demonstration prior to joining the study. Given the low national rates of condom use among men and women in Sierra Leone [9], exposure to condoms via the behavioral counseling protocol may eventually be associated with gains in reproductive health and STI prevention in participants.
Communicating the risk of sexual transmission of EBOV to study participants was challenging as there was limited scientific evidence regarding the length of persistence and the infectiousness of positive qRT-PCR body fluids. In this study, we observed that communicating uncertainty in a transparent way appeared to facilitate trust between participants and the research study. For example, following the report of an EVD survivor in the United Kingdom who relapsed with EVD meningitis [22], the behavioral counseling protocol was adapted to ensure that participants were aware of possible severe relapses. Future development of similar counseling protocols may consider embracing scientific transparency as a trust-building communication tool.
In the course of the study, we identified a need to clearly and simply explain the detection method of the qRT-PCR assay to participants. We developed a “mango tree” analogy comparing detection of EBOV-specific RNA primers (short target sequences) in the body fluids of Ebola survivors via qRT-PCR testing to trying to detect a whole mango tree (the intact, viable Ebola virus) by being able to detect only the mango fruits or the leaves (the target RNA primers for EBOV). Detecting only fruits or leaves, one cannot determine whether the tree is indeed intact and alive just as qRT-PCR testing can only determine whether the target RNA primers are detected or undetected in a body fluid specimen, and not whether the virus is infectious. Detecting only one of the two target RNA sequences (an indeterminate test result), could also be explained using this analogy, and also to explain possible variation in test results using different qRT-PCR assays with differing target sequences. This addition to the counseling script was very helpful for study participants to conceptualize the test, the different results they encountered, and the associated counseling messages.
A flexible protocol was also instrumental in optimizing participant flow during the study. For example, some participants preferred to visit the counselor prior to specimen collection for more detailed instructions on specimen collection processes or for extra encouragement to continue with the testing process. Talking points were developed as reference materials for study counselors as the original scripts were lengthy. Because we believed that participants should receive their previous test results before deciding whether to donate specimens for testing at the current visit, there was a delay between delivery of test results and post-test counseling and explanation of guidance. Future iterations of similar counseling protocols may be able to concurrently deliver testing results and post-test counseling if a rapid diagnostic test for detection of Ebola RNA in the body fluids of Ebola survivors becomes available.
The EVD epidemic had a devastating impact on the economic and social ties of communities across Sierra Leone [12, 21, 24]. EVD survivors commonly reported medical and mental health issues post-convalescence for which there was limited assistance in Sierra Leone [25]. These needs were far greater than the services the study could offer. This was particularly true for semen testing, as there were far more EVD survivors in Sierra Leone than could be enrolled in the EBOV persistence study. To this end, the Government of Sierra Leone initiated a comprehensive program for EVD survivors (CPES) in October 2015 that aimed to provide services for EVD survivors in Sierra Leone including counseling, semen testing, eye care, myalgia, and treatment for other EVD sequelae [26].
The mental health infrastructure of Sierra Leone is underdeveloped. There is limited capacity to identify and treat mental health disorders in the general public or EBOV-affected populations [12]. When the study was initiated, counseling staff were less familiar with behavioral counseling methods that emphasized dialogue regarding participants’ experiences with EVD survivorship, stigma, and intimate relationships. Learning these skills helped study counselors better assist participants who received multiple positive qRT-PCR test results to maintain participation in semen testing and openly discuss feelings of frustration and anxiety.
After observing initial high rates of HIV testing uptake in the first phase of the study, acceptance of HIV testing declined among study participants during the first visit. We hypothesize that participants may have felt anxious about their EBOV qRT-PCR tests and did not want to learn their HIV test results at the same time. We re-trained study counselors on HIV testing and counseling procedures and encouraged them to offer HIV tests at follow-up study visits and observed a rise in uptake of HIV testing. More formative research is needed to determine how best to incorporate both HIV and EBOV testing and counseling in future EVD epidemics.
One limitation of our behavioral counseling protocol is that, due to time constraints during study visits, we did not employ couples’ counseling. Future development of similar EBOV persistence behavioral counseling protocols may consider the inclusion of couples’ counseling to facilitate correct and consistent condom use, particularly for those participants who repeatedly test positive [27]. Consideration could also be given on how best to recruit, enroll, and counsel EVD survivors who are men who have sex with men (MSM), and who may be reluctant to discuss their sexual behavior in an environment where same sex behavior is criminalized [9], as well as survivors who may object to body fluid testing due to religious objections to masturbation or other specimen donation procedure.
Another limitation of the protocol is that we developed and implemented risk reduction guidance when relatively little was known about the persistence of Ebola RNA in the body fluids of survivors, or how qRT-PCR results related to the risk of transmitting Ebola to others. Since the implementation of this behavioral counseling protocol, additional data on the persistence of Ebola RNA in semen has become available from multiple Ebola-affected countries [28–31]. These data show Ebola RNA persisting in semen for lengthy periods of time for some Ebola survivors, and further illustrate the critical role that semen testing and behavioral counseling can play in Ebola epidemic control [32]. Future development of similar behavioral counseling protocols should consider these additional data in adapting behavioral guidance for Ebola survivors, such as collecting body fluid specimens more frequently and increasing the number of consecutive negative tests needed before testing ceases following the receipt of positive test results.
A behavioral counseling protocol that pairs test results with risk reduction behavioral guidance might help mitigate transmission risks associated with body fluids of EVD survivors in which EBOV has been detected. Risk reduction behavioral counseling is rapidly becoming an integral part of addressing the sexual transmission risk in this EVD epidemic and in filovirus outbreaks moving forward; its utility will likely also be swiftly demonstrated for other pathogens where virus persistence in body fluids may pose a risk for continued transmission from survivors [32]. A qualitative evaluation assessing the impact of the behavioral counseling protocol on study participants as well as staff perceptions of the protocol has been performed to add to lessons learned presented in this manuscript.
As of July 2016, the Ebola Virus Persistence Risk Reduction Behavioral Counseling Protocol has been used for more than 220 male and 120 female participants. The protocol has been adapted for use by other body fluid testing programs for EVD survivors, including the Government of Sierra Leone’s CPES (Alpren, C., personal communication) and Liberia’s Men’s Health Screening Program [29]. Lessons learned from implementation of the behavioral counseling protocol in the Sierra Leone Ebola Virus Persistence Study have also been shared to advise the operations of these semen testing programs [29]. We hope that our experience implementing this behavioral counseling protocol in the midst of an EVD epidemic in Sierra Leone can inform similar future efforts so that robust and effective services can be provided to EVD survivors.
|
10.1371/journal.pntd.0000695 | A Mechanism for Chronic Filarial Hydrocele with Implications for Its Surgical Repair | Chronic hydrocele is the most common manifestation of bancroftian filariasis, an endemic disease in 80 countries. In a prospective study, we evaluated the occurrence of intrascrotal lymphangiectasia, gross appearance/consistency of the testis, and the efficacy of complete excision of hydrocele sac in patients living in a bancroftian filariasis endemic area who underwent hydrocelectomy at the Center for Teaching, Research and Tertiary Referral for Bancroftian Filariasis (NEPAF).
A total of 968 patients with uni- or bilateral filarial hydrocele (Group-1) and a Comparison Group (CG) of 218 patients from the same area who already had undergone hydrocele-sac-sparing hydrocelectomy elsewhere were enrolled at NEPAF. Twenty-eight patients from the Comparison Group with hydrocele recurrence were re-operated on at NEPAF and constitute Group-2. In Group-1 a total of 1,128 hydrocelectomies were performed (mean patient age of 30.3yr and mean follow-up of 8.6yr [range 5.3–12]). The hydrocele recurrence rates in Group-1 and in the Comparison Group (mean age of 31.5 yr) were 0.3%, and 19.3%, respectively (p<0,001). There was no hydrocele recurrence in Group-2 (mean patient age of 25.1yr and mean follow-up of 6yr [range 5–6.9]). Per surgically leaking or leak-prone dilated lymphatic vessels were seen in the inner or outer surface of the hydrocele sac wall or in surrounding tissue, particularly in the retrotesticular area, in 30.9% and in 46.3% of patients in Group-1 and Group-2, respectively (p = 0.081). The testicles were abnormal in shape, volume, and consistency in 203/1,128 (18%) and 10/28 (35.7%) of patients from Group-1 and Group-2, respectively (p = 0,025).
Lymph fluid from ruptured dilated lymphatic vessels is an important component of chronic filarial hydrocele fluid that threatens the integrity of the testis in an adult population living in bancroftian filariasis endemic areas. To avoid hydrocele recurrence the authors advise complete excision of hydrocele sac and when identified, leaking or leak-prone dilated lymphatic vessels should be sutured or excised.
| Chronic hydrocele is the accumulation of fluid around the testis leading to an increase in the volume of the scrotal contents. Depending on the volume of fluid, hydrocele can be disfiguring and even incapacitating. Chronic hydrocele has multiple etiologies, but irrespective of the cause, surgery is the standard form of treatment and this can be done using different surgical techniques. The prevalence of chronic hydrocele in bancroftian filariasis endemic areas—a parasitic disease transmitted by mosquito—is very high and represents the most common clinical manifestation of bancroftosis, following by swollen legs of lower limbs or lymphedema among women. In Greater Recife, northeastern, Brazil, a bancroftian filariasis endemic area, a pioneering, prospective surgical study proposes a new mechanism for filarial-induced hydrocele and presents evidence that the filarial hydrocele fluid may damage the testis. Thus, based on the findings presented, the authors propose that in bancroftian filariasis endemic areas hydrocele patients should be operated on using a specific surgical technique in order to avoid recurrence of the disease, and consequently, additional damage to the testicle.
| Bancroftian filariasis is a mosquito-borne parasitic disease that affects approximately 100 million persons worldwide. It is estimated that 40 million persons suffer from the chronic disfiguring manifestations of this disease, including 27 million men with testicular hydrocele, lymph scrotum, or elephantiasis of the scrotum. The genital pathology caused by bancroftosis is impressively debilitating and economically punishing for huge numbers of adult males throughout endemic countries [1].
There is consensus that hydrocele is the most frequent clinical manifestation of bancroftian filariasis [2]. On the other hand, few systematic studies of the prevalence or incidence of hydrocele in temperate climates have been published. In spite of this, there appears to be a difference between the hydrocele prevalence in temperate countries and tropical and subtropical ones. In the tropics, hydrocele is a very frequent condition [3]. In locations where bancroftian filariasis is highly endemic, up to 40% of adult males are reported to have testicular hydrocele [4], [5].
With exception of its posterior aspect, the testicle is covered by the tunica vaginalis formed by two layers. The visceral layer and parietal layer are in direct contact with the testis and the scrotum wall, respectively. Both layers have secretory ability and only the parietal layer has resorption property. Between the layers there is a space called vaginal cavity. The balance of secretory and absorptive functions of these layers results in a small volume (0.5 to 2.0 mL) of straw colored fluid in this cavity. Outside of congenital hydrocele, the general mechanism accepted for abnormal chronic accumulation of fluid in the vaginal cavity, known as acquired hydrocele, irrespectively of the etiology, is unbalanced process between fluid production by the mesothelial cells of the inner surface of the tunica vaginalis and fluid absorption by the draining lymphatic vessels of the parietal layer. Surgery using different techniques is the standard treatment form for chronic hydrocele, regardless of the etiology, and the rationale for its repair is to expose, permanently, the secretory surface of tunica vaginalis to the absorbing surface of scrotal wall.
In Greater Recife, northeastern, Brazil, a bancroftian filariasis endemic area [6] the prevalence of acquired hydrocele in the adult male population is high and hydrocelectomy represents a commonly performed procedure in public hospitals. At these hospitals, the surgeons' preference is for surgical techniques in which the hydrocele sac is opened, everted with or without partial resection of the sac, and the edges sutured behind the testis (J. Norões, personal communication).
The purpose of this study is to evaluate the occurrence of lymphangiectasia in scrotal contents, the morphology and consistency of the testicles and, hydrocele recurrence using the complete excision of tunica vaginalis in hydrocele patients living in a bancroftian filariasis endemic area.
This study was carried out between March 1994 and March 2006 as part of a larger comprehensive study on different aspects of urological manifestation of bancroftian filariasis based on clinical, parasitological, chemotherapeutical, ultrasonographic, biochemical, surgical and histopathological observations [7]–[16]. The study was approved by the Ethics Committee of Hospital das Clinicas at Federal University – Pernambuco, Brazil. It comprised three groups.
Patients who underwent their first operation at NEPAF. Patients were selected using the following criteria: (a) aged between 18 and 40 year-old at the time of surgery and had signed the inform consent forms, which included permission to use his medical information for scientific publication; (b) documented uni- or bilateral hydrocele by physical examination (inspection and bimanual palpation) and ultrasonography of the scrotal area; (c) currently live or have lived in a house with at least one person with documented W. bancrofti microfilaremia: (d) do not have ipsilateral inguinal hernia; (e) no history of past ipsilateral inguinal hernia surgical repair, urological surgery, or intra-scrotal inflammation from testicular trauma or bacterial infection; (g) no current evidence of intrascrotal neoplasia; (g) no current ultrasonographic evidence of severe testicular damage which could anticipate the need of orchiectomy during hydrocelectomy; (h) demonstrated straw-colored hydrocele fluid during hydrocelectomy; (i) have at least five years of postoperative follow up; (j) have lived in a filariasis-endemic area for their entire life; (k) no medical contraindication for hydrocelectomy. Eligible patients were operated on by the same surgeon.
Between January 1999 and March 2001 any patient with bancroftian filariasis active infection/disease referred to NEPAF underwent a family protocol investigation that included the collection of information about current or past filarial infection/disease in family members of the patient. One of the questions to address hydrocele occurrence was whether or not hydrocelectomy had been performed in any family member. The family member(s) identified for previous hydrocelectomy was/were invited to NEPAF and selected according the following inclusion criteria: (1) if they lived in a house where at least one person had been positive for Wuchereria bancrofti microfilaria in the blood, (2) if they had had hydrocelectomy performed between the ages of 18 and 40 at public hospitals in Greater Recife and, (3) if they have signed the inform consent form. For each individual, information regarding the hydrocelectomy technique used was retrospectively gathered from the patient's medical records. All included, consenting individuals underwent physical and ultrasound examinations of the scrotal area.
Patients from CG who underwent hydrocelectomy at NEPAF for hydrocele recurrence. Inclusion criteria were the same used for G1 except that the previous hydrocelectomy was performed elsewhere.
Hydrocelectomies were performed under NEPAF protocol, which included total excision of tunica vaginalis parietal, regional anesthesia and prophylactic antibiotic therapy. A povidone-iodine scrub was used for preoperative skin preparation. After the surgical area was prepared and draped an incision was made parallel to the median raphe and deepened in the layers of the scrotum wall. The dissection of the hydrocele sac was made until it was completely separated from its surrounding tissue. After meticulous hemostasis, the hydrocele fluid was drained completely from the sac using a 16–18 gauge intravenous catheter and a syringe. Only in patients presenting testes abnormalities (see below) was the volume considered in the analysis. In patients presenting with two hydrocele sacs, the hydrocele volume was considered the sum of the volume of both sacs. Between two clamps, a middle line incision was made in the anterior aspect of the sac wall in its midpoint along its vertical axis and extended, cephalically, until the limit of the cord close to the head of the epididymis and, caudally, until the proximity of its tail. After the hydrocele sac had been opened it was excised all around, close to its reflexion onto the visceral layer. Pre and post hydrocele sac excision a careful examination was done to look for dilated lymphatic vessels, whether leaking or not, especially in the vicinity of the excised sac wall. After identification, all visible dilated lymphatic vessels were sutured or ligated and excised. In thin-walled hydrocele sac any visible bleeding at the incised site was coagulated. In thick-walled cases the cutting margin was marsupialized by suturing it with a 4-0 chromic catgut continuous interlocking suture. Once hemostasis was achieved, the testis was returned to the scrotum and the wound was closed in two layers. The inner layer was closed with an interrupted 4-zero chromic catgut suture and the skin and dartos with a continuous 3-0 monofilament nylon suture. As a rule, no drains were used. A small dressing and a slight compressing scrotal bandaging were then applied. In cases with bilateral hydrocele, the same procedure was performed on the contra-lateral side using the same surgical procedure. The post-operative management included oral analgesic if necessary, scrotal support and, “ice bag” four times a day during the hospitalization period. At discharge the patients were instructed to perform personal hygiene twice a day with soap and clean water with changing of underwear. Underwear and soap were provided by NEPAF when necessary. After hospital discharge three days postoperatively, the patients were seen for follow up on the fifth day. If no complications were recorded, the patients were followed up on the ninth (when suture was removed), sixteenth and thirtieth post-operative day, and each two months for six months, and at least every twelve months thereafter. In cases of any complication the patients were seen as needed.
Testicles were considered to be abnormal if they were seen per-surgically to have both of the following characteristics: (1) loss of the typical ovoid shape and noticeable volume reduction by inspection and (2) thickness of vaginal visceral and albuginea layers, both by inspection and palpation (increased consistency). During follow-up, reappearance of intra-scrotal fluid of any volume suspected by physical examination and confirmed by ultrasound was defined as hydrocele recurrence. Recurrence was also considered to be present if intrascrotal fluid was detected only by ultrasound.
A two-tailed binomial test was used to compare proportions with a hypothesized value. Differences in proportions were tested using Fisher's exact test or Pearson's chi-square test. Differences in means were tested using a two-tailed Student's t test.
All patients from Group 1 and Group 2 were operated on between March 1994 and March 2001. No patients dropped out. It was not possible to define the duration of disease. Patients gave different answers when asked on several occasions during the study and their responses did not refer to the onset of the disease, but to the time when the volume of the hydrocele began to bother them. Thus, the duration of the disease could not be obtained in a reliable manner and is not presented. Only patients from Group 2 presented with more than one hydrocele sac. Orchiectomy was not performed on any of the patients included in the present study. Oral analgesic was needed only in 31 patients within 24 hours postoperatively (30 in Group 1 and one in Group 2). There was no postoperative haematoma or infection. Chronic edema, elephantiasis or lymphscrotum were not seen in the scrotal wall in any patients during the follow-up period (from five to 12 years).
General information about the patients, the characteristics of their hydroceles, hydrocelectomy and recurrence are found in Table 1. Two of the 968 patients experienced recurrence. One of them recurred twice (case 1). They were re-operated and the findings are described below.
Of the 436 patients presenting at NEPAF in whom previous hydrocelectomy was performed in another hospital, 93 presented hydrocele recurrence (21.3%). It was possible to obtain the information about the surgical technique used to repair the hydroceles in 218. Characteristics of the CG are found in Table 2. In patients for whom the medical records could not be retrieved, the hydrocele recurrent rate was 23.4%. No abnormalities in the scrotal wall such as chronic edema, elephantiasis or lymphscrotum, were seen in any of the 436 patients at physical examination.
Of the 42/218 patients with recurrence and medical records retrieved, 28 agreed to a second operation at NEPAF. The surgical technique used previously was eversion with and without partial excision of the sac in 12 (42.8%) and 16 (57.2%) patients, respectively. The surgical approach for recurrent cases was the same used in G1. The twenty eight unilateral cases in G2 could be classified in three categories according to the location of the recurrent sac, as seen at surgery: (1) anterior recurrence: In five patients (17,8%) the recurrent sac was covering the anterior surface of the testis (Figure 1); (2) posterior recurrence. In eleven patients (39,3%) the returning sac was located on the posterior aspect of the testis and the epididymis (Figure 2) and (3) mixed recurrence. Twelve patients (42,8%) presented at least two separate recurrent sacs. They were located anterior and posterior to the testicular surface (Figure 3).
During surgery in G1 and G2, in patients with slight or mild fibrosis in hydrocele sac and surrounding tissue, intact dilated lymphatic vessels and/or a small quantity of slowly dripping clear and slightly yellow fluid, could often be identified (Table 3). These lymphatic vessels and/or dripping process were seen in the inner or outer surface of the sac wall or in the retrotesticular area, near what became the resection-line when removing the sac (Figure 4).
Marked changes in shape, volume, and consistency of the testis due to inflammatory reactions associated with fluid collection around the testis in G1 and G2 were seen significantly more so in G2 than G1 (p = 0,025). Testes seen at surgery to be abnormal, were affected with a range of severity (Table 3, Figure 1D and Figure 3C). The hydrocele volume was significantly lower in G2 patients (p = <0,001).
Hydrocelectomy dates from remote antiquity. The procedures employed have been devised and modified over the years resulting in a multiplicity of techniques with many variations and modifications of the original methods. It is generally believed that the pathogenesis of acquired hydroceles, irrespective of the etiology, is an unbalanced process between transudate production and reabsorptive activity of the tunica vaginalis parietal lymphatics [17], [18]. Thus, the rationale for open hydrocelectomies is to expose, permanently, the hydrocele fluid to an absorbing surface [19]. The mechanism of generating hydrocele has been considered to be the same in all acquired hydroceles and the etiology, per se, has not been considered when choosing the surgical approach. Fundamental factors influencing the choice of surgical procedure are efficacy, simplicity, safety, and, cost-effectiveness of the treatment. In addition to the surgeon's preference, the size of the hydrocele and the thickness of its sac wall have been considered elements that could be taken into account when deciding which technique to choose [19]–[24].
The hydrocele recurrence rates vary among studies [25]–[26] using different sample sizes, different criteria for recurrence, the type of the study (prospective or retrospective), degree of thickness of the hydrocele sac, the chosen surgical technique, different inclusion and exclusion criteria, different backgrounds of the health personnel involved and follow up periods. As a consequence, it is not easy to make accurate comparisons across studies.
The very significant difference in hydrocele rate recurrence (p<0,001) observed between patients from G1 and CG after two different surgical approaches – (1) complete excision and (2) eversion with or without partial excision of sac – could reflect more than differences in surgical technique. It may also signify that the pathogenesis of chronic filarial hydrocele is rather complex. The intrascrotal lymphatic vessels appear to be the preferred site for the adult worms of W. bancrofti in infected men. Extensive clinical, surgical and histological observations indicate that in almost 90% of infected men, adult W. bancrofti can be detected in the lymphatic vessels of the scrotal area [9], [10]. The primary lesion of bancroftian filariasis, while adult worms are alive, is non obstructive lymphatic vessel dilation without inflammation [8], [11]–[13]. Norões et al. [7] demonstrated that, by contrast, acute filarial hydrocele is a consequence of acute interruption of lymph flow from the tunica vaginalis of the testis. This obstruction is caused by filarial granuloma (corresponding to formation of palpable nodules detected by physical examination of intrascrotal contents [15]) resulting from death of W. bancrofti adult worms in the lumen of intrascrotal lymphatic vessels. They observed that 22% of patients who experienced nodule formation also developed acute hydroceles, and reabsorption of the granuloma led to resolution of 76% of such acute hydroceles within seven months during an eighteen month follow up period. Risk of acute hydrocele development, following a single filarial granuloma “event”, was increased by (1) the presence of nodules located in the superior paratesticular area (or adjacent to the posterior part of the upper pole of the testicle), a critical site where the lymphatic drainage of the tunica vaginalis, epididymis and testis converges [27] and (2) the occurrence of multiple filarial nodules. However, the risk factors and mechanism that lead acute filarial hydrocele to persist and to progress toward a chronic condition are not completely understood. Norões et al. [7] speculate that factors such as adult worm burden, formation of additional nodules, the speed of the granulomatous recanalization process, and the degree of pre-existing sub-clinical lymphatic disfunction could contribute to the chronicity of the process. The low progression rate of this “obstructive” hydrocele stands in contrast to the high prevalence of chronic hydrocele in filariasis endemic areas.
Based on observations after operating on patients with chronic hydrocele from non-endemic and endemic areas, and comparing surgical findings, we believe the accumulation of fluid in the vaginal cavity of the testis, in a large proportion of chronic filarial hydrocele cases, may be due to a different pathogenetic mechanism.
The surgical findings in recurrent and non recurrent hydroceles in G2 and G1 respectively, support the conclusion that lymph fluid composes the hydrocele fluid, based on the following evidences: (1) per-operatively, intact dilated lymphatics and/or a small quantity of slowly dripping clear and slightly yellow fluid were continuously seen in approximately one-third of the patients in G1 and G2. When a dilated lymphatic vessel ruptures, it is very difficult to visualize the vessel itself, but with a careful examination it is possible to see the clear fluid flowing from the lymphatic fistula. On the other hand, the lack of leak-prone dilated lymphatic vessels and lymphatic fistula in patients with a thicker tunica vaginalis and albuginea can be explained, in principle, by the important inflammatory reaction triggered by lymph fluid leading to fibrosis; (2) in 82,1% (22/28 ) patients from posterior recurrence of the hydrocele recurrence cases (G2) where the fluid collection was in a cavity located behind the testis and epididymis, the cavities were formed by the everted sacs where the fluid collection was in direct contact with the outer surface of the everted hydrocele sac. As well known, this outer surface is not a serous lined layer which is not capable of producing fluid. In everted sacs, the lined serous layer was in direct contact with the scrotal wall where lymphatics are able to drain fluid production. This evidence argues against primacy for the concept that transudate production/reabsorption problems represent the essence of the majority of chronic hydrocele fluid accumulation in lymphatic filariasis.
Based on results of the current study, it seems reasonable to propose that, first, the straw colored “filarial hydrocele fluid” consists of a combination of clear lymph fluid and transudate, in various proportions. The major mechanism of chronic filarial hydrocele proposed in the present study, is the same for milky fluid accumulation in cases of chylocele. In chyloceles the difference is the presence of chylomicrons caused by diffuse and extensive high retroperitoneal lymphangiectasia promoting the retrograde flow of milky lymph in ruptured intrascrotal lymphatics (J. Norões, personal communication). Second, based on G2 findings the high hydrocele recurrence rate in CG was likely to be due to the presence of lymphatic fistula in everted hydrocele sac and/or in surrounding tissue.
Based on the current findings it is suggested that the term “filaricele” be introduced for chronic straw colored fluid accumulation in the vaginal cavity in endemic areas. This term would help to emphasize the differences in pathogenesis and in the recommended surgical techniques that are appropriate for acquired chronic filarial and non-filarial hydroceles. On the other hand, the term chylocele should be kept for the milky appearance of hydrocele fluid rich in chylomicrons.
Two other practical lessons were also learned from this study: (1) in the second recurrent case from G1, the sac extended up to the spermatic cord, and a small posterior segment of the sac wall, adherent to the cord, was inadvertently not excised at the first hydrocelectomy. During the re-operation a small portion of tunica vaginalis in the inferior part of the cord was found to contain a small cystic hydrocele. Thus, excision of the tunica vaginalis should be as complete as possible; (2) in hydrocele among patients from non-endemic areas, the patient's personal preference as to whether or not he desires treatment should be given strong consideration since the indication for treatment depends, particularly, upon how much the hydrocele bothers him. By contrast, chronic filarial hydrocele could threaten the integrity of the testis even in small volume cases as shown in this study. Thus the medical indication for surgical treatment is stronger and the patient should be advised accordingly.
It was beyond of the scope of this study to compare the volume of all hydrocele cases, echogenicity of the fluid, spermogram profiles and microfilaraemia status, which are planned to be published separately.
One potential limitation of the study was the small sample size of Group 2. In spite of that, this group provided unprecedented detailed findings leading to a pioneering classification of recurrent hydrocele in endemic area occurring after eversion technique with and without partial excision of the hydrocele sac.
In conclusion, in bancroftian filariasis endemic areas, lymphatic fistulae are likely to be an important mechanism responsible for chronic hydrocele, whether recurrent or not. Particularly with the intent to avoid hydrocele recurrence and testicular damage, complete excision of the hydrocele sac with its dilated lymphatic vessels and/or lymphatic fistula is advised as is identification and suturing or excision of any visible dilated lymphatic vessels in surrounding tissue, whether leaking or not.
|
10.1371/journal.ppat.1005059 | Analysis of the SUMO2 Proteome during HSV-1 Infection | Covalent linkage to members of the small ubiquitin-like (SUMO) family of proteins is an important mechanism by which the functions of many cellular proteins are regulated. Sumoylation has roles in the control of protein stability, activity and localization, and is involved in the regulation of transcription, gene expression, chromatin structure, nuclear transport and RNA metabolism. Sumoylation is also linked, both positively and negatively, with the replication of many different viruses both in terms of modification of viral proteins and modulation of sumoylated cellular proteins that influence the efficiency of infection. One prominent example of the latter is the widespread reduction in the levels of cellular sumoylated species induced by herpes simplex virus type 1 (HSV-1) ubiquitin ligase ICP0. This activity correlates with relief from intrinsic immunity antiviral defence mechanisms. Previous work has shown that ICP0 is selective in substrate choice, with some sumoylated proteins such the promyelocytic leukemia protein PML being extremely sensitive, while RanGAP is completely resistant. Here we present a comprehensive proteomic analysis of changes in the cellular SUMO2 proteome during HSV-1 infection. Amongst the 877 potentially sumoylated species detected, we identified 124 whose abundance was decreased by a factor of 3 or more by the virus, several of which were validated by western blot and expression analysis. We found many previously undescribed substrates of ICP0 whose degradation occurs by a range of mechanisms, influenced or not by sumoylation and/or the SUMO2 interaction motif within ICP0. Many of these proteins are known or are predicted to be involved in the regulation of transcription, chromatin assembly or modification. These results present novel insights into mechanisms and host cell proteins that might influence the efficiency of HSV-1 infection.
| Proteins are subject to many types of modification that regulate their functions and which are applied after their initial synthesis in the cell. One such modification is known as sumoylation, the covalent linkage of a small ubiquitin-like protein to a wide variety of substrate proteins. Sumoylation is involved in the regulation of many cellular pathways, including transcription, DNA repair, chromatin modification and defence to viral infections. Several viruses have connections with sumoylation, either through modification of their own proteins or in changing the sumoylation status of cellular proteins in ways that may be beneficial for infection. Herpes simplex virus type 1 (HSV-1) causes a widespread reduction in uncharacterized sumoylated cellular protein species, an effect that is caused by one of its key regulatory proteins (ICP0), which also induces the degradation of a number of repressive cellular proteins and thereby stimulates efficient infection. This study describes a comprehensive analysis of cellular proteins whose sumoylation status is altered by HSV-1 infection. Of 877 putative cellular sumoylation substrates, we found 124 whose sumoylation status reduces at least three-fold during infection. We validated the behavior of several such proteins and identified amongst them several novel targets of ICP0 activity with predicted repressive properties.
| Herpes simplex virus type-1 (HSV-1) is an alphaherpesvirus which causes vesicular oral and genital lesions, and has the capacity to cause more severe diseases such as meningitis and encephalitis, particularly in immunocompromised individuals and neonates (see [1,2] for general reviews). Characteristic of an alphaherpesvirus, HSV-1 establishes latency within sensory neurons, from which reactivation occurs periodically. The lytic, latent and reactivation states are governed by the innate, intrinsic and adaptive immune responses and the mechanisms by which HSV-1 has evolved to counteract these immune responses.
The attachment and entry of HSV-1 into a cell causes the activation of the innate and intrinsic immune responses. The former involves production of interferons (IFNs) which activate signal transduction pathways, resulting in the expression of IFN stimulated genes (ISGs) (reviewed in [3]). Intrinsic antiviral resistance, on the other hand, is mediated by constitutively expressed proteins. Amongst the various factors that have been identified as contributing to intrinsic resistance are certain components of promyelocytic leukaemia (PML) nuclear bodies (PML NBs, also known as ND10), including the PML protein itself and other major components such as Sp100, hDaxx and ATRX [4]. Both PML and Sp100 are heavily modified by the SUMO family of ubiquitin-like proteins [5], and both sumoylation and interaction with sumoylated proteins are key factors in the assembly of PML NBs [6,7]. These proteins are recruited to sites of incoming HSV-1 genomes very early in infection [8] and they have the potential to restrict HSV-1 replication as soon as the cell becomes infected [9–12]. The mechanism of this recruitment is incompletely understood, but it is clear that both sumoylation and SUMO-mediated interactions play important roles [13].
The HSV-1 regulatory protein ICP0 reduces the sensitivity of HSV-1 to IFN [14–16] and also counteracts the restrictive effects of PML NBs through its ubiquitin E3 ligase activity [17]. ICP0 induces the degradation of the sumoylated forms of PML through an activity that has similarities to those of SUMO-targeted ubiquitin ligases (STUbLs) [18,19], and it also degrades the unmodified forms of the most abundant isoform of PML in a SUMO-independent manner [20]. In addition, at later times of HSV-1 infection a widespread loss of high molecular weight cellular SUMO conjugates occurs in an ICP0-dependent manner [18,19]. These activities combine to cause complete disruption of PML NBs, dispersal of those components such as hDaxx and ATRX that are not degraded, and inhibition of the recruitment of PML NB components and other as yet uncharacterized sumoylated proteins to the sites of HSV-1 genomes. Given that depletion of the only known SUMO E2 conjugating enzyme Ubc9 also diminishes intrinsic resistance to HSV-1 infection (and hence augments the replication of ICP0-null mutant HSV-1) [18], there is accumulating evidence that mechanisms that are regulated by sumoylation play an important part in intrinsic resistance to HSV-1 infection.
The SUMO family of proteins includes three members (for general reviews of the SUMO pathway, see [21,22]). By sequence, SUMO1 is 18% related to ubiquitin and is conjugated to specific lysine residues in substrate proteins through an isopeptide bond between its C-terminal glycine carboxyl group and the lysine side chain in the substrate. SUMO2 and SUMO3 are closely related to each other (and share about 50% identity with SUMO1) and they include an internal lysine residue to which other SUMO moieties can be conjugated, and hence they can form poly-SUMO chains. The SUMO conjugation pathway involves a sequence of events that are analogous to those of ubiquitin conjugation, with SUMO first forming a thioester bond with the SAE1/2 SUMO E1 activation enzyme, followed by the activities of Ubc9 and in some cases SUMO E3 ligases that catalyze conjugation to specific substrate proteins. Several previous studies have analyzed the diversity of cellular proteins that can be sumoylated under a variety of conditions [23–28].
In this study we set out to characterize changes to the cellular SUMO2 proteome in response to infection with wild type (wt) HSV-1 using Mass Spectrometry (MS)-based quantitative proteomics. Our experimental design allowed the identification of sumoylated proteins that are preferentially degraded in the presence of ICP0, and it also revealed a number of cellular proteins that were not candidate sumoylated species that were also degraded. We investigated a number of the identified proteins, including known controls such as PML and Sp100, to determine their sumoylation status and whether or not their degradation was ICP0 dependent. The data revealed widespread changes in the SUMO2 proteome during HSV-1 infection, revealing many proteins that are potentially involved in the regulation of gene expression that have hitherto not been previously identified or examined in the context of a viral infection. Of particular note, we identified several members of the ZBTB family of proteins, and other proteins with related BTB domains, that are susceptible to degradation by ICP0. We present examples of these proteins that are degraded by mechanisms that are influenced by modification by SUMO2/3, or on the presence of a SUMO2/3 interaction motif within ICP0. Our results provide detailed insight into the numerous and complex changes in protein stability that occur during HSV-1 infection.
Cells expressing poly-histidine tagged SUMO1, -2 or -3 were isolated after transduction with lentivirus vectors. HepaRG hepatocytes were used because they are readily infected by HSV-1 and, unlike human diploid fibroblasts (HFs), their growth was not compromised in isotopically labeled SILAC medium. We concentrated on poly-histidine tagged SUMO2 (His-SUMO2)-expressing cells (herein named HA-HisSUMO2) because of the potential for poly-SUMO2 chains and because of the existing depth of knowledge of the SUMO2 proteome [23–28]. Infection of HA-HisSUMO2 cells with wt HSV-1 caused a reduction in the overall abundance of His-SUMO2 conjugated proteins (Fig 1A), which was more marked at later times of infection (S1A Fig). This was not as pronounced as in previous studies using HFs [18], perhaps because of over-expression of His-SUMO2 (S1B Fig), or because reductions in overall sumoylation levels are less pronounced in HepaRG cells compared to HFs [18]. Despite the increased expression of SUMO2, the sumoylated forms of PML were not substantially more abundant in HA-HisSUMO2 cells (S1C Fig) and they were readily degraded by HSV-1 (Fig 1A). Wt HSV-1 gene expression (Fig 1B) and plaque formation efficiency (Fig 1C) were as efficient in HA-HisSUMO2 cells as in parental HepaRG cells and control transduced cells expressing the His-tag only (HA-His only cells). Surprisingly, the plaque formation efficiency of ICP0-null mutant HSV-1 was increased in HA-HisSUMO2 cells (S1D Fig). Given previous results on the role of Ubc9 in restricting this virus [18], it might be expected that over-expression of SUMO2 could be inhibitory. On reflection however, this result does not necessarily challenge the hypothesis that sumoylation of repressive proteins, and/or their interactions with sumoylated proteins, contributes to the regulation of HSV-1 infection. Over-expression of a protein can affect the proper functioning of the pathway in which it is involved, and in this case over-expression of SUMO2 might affect the balance of interactions between sumoylated proteins and those containing SUMO interaction motifs (SIMs). Support for the role of sumoylation in intrinsic resistance to HSV-1 infection comes from a study of HepaRG cells highly depleted of SUMO2/3, in which PML NBs are disrupted and ICP0 null mutant HSV-1 replicates with increased efficiency (M. Glass, manuscript submitted for publication). These issues are clearly open for further experimentation, but they are beyond the scope of this particular paper.
A method involving nickel affinity purification of His-tagged sumoylated proteins under denaturing conditions [27] gave efficient recovery of both overall sumoylated proteins (Fig 1D) and the specific example of sumoylated PML (Fig 1E). HA-HisSUMO2 cells were grown in isotopically normal (light; L) or heavy (H) SILAC media (in which lysine and arginine have the heavy isotopes of nitrogen (15N) and carbon (13C)). Uninfected HA-His only control cells were grown in isotopically intermediate (M) medium, as defined in the methods section. The L cells were infected with wt HSV-1 at MOI 10, then all but one plate of cells from each condition were harvested directly into guanidinium denaturing buffer 12 h later. The H, L and M lysates were mixed in equal protein amounts then used for nickel affinity purification of SUMO2-modified proteins. To allow analysis of total protein abundance changes and confirm efficient virus infection and degradation of previously characterized cellular proteins, cells from one plate of each set were harvested directly into SDS-PAGE loading buffer. Samples of the crude and affinity purified mixtures were subjected to SDS-PAGE, then the gels were stained and cut into slices for in gel tryptic peptide production. The peptides were analyzed by LC-MS/MS and the data analyzed using MaxQuant software (See Materials and Methods for details). Fig 2A shows a flow diagram of the experimentation and images of the resulting SDS-PAGE stained gels. Data derived from the crude and purified samples gave information on the relative levels of total protein and putative sumoylated protein species respectively. Analysis of the H/M ratios in the purified sample allowed the separation of likely sumoylated species from non-specifically purified proteins, while the H/L ratios of these protein IDs in the same sample can be used to assess changes in relative abundance during HSV-1 infection.
A total of 6128 cellular proteins were identified, with 5508 and 2842 from the crude and purified fractions respectively, of which 2222 were in common (S1 Table, sheet 1 and Fig 2B). A number of proteins were detected only in the affinity purified sumoylated fraction, probably because their intrinsic abundance is too low for detection in the crude (Fig 2B). Because the majority of proteins would not be expected to change in abundance during infection, and also the presence of non-sumoylated proteins in the purified fraction, MaxQuant internally calculated normalised ratios could be used to correct for any errors in the mixing of the various samples prior to gel electrophoresis. Frequency plots for log2 HA-HisSUMO2 to HA-His only ratios (log2 H/M) showed little variation in abundance of proteins in crude extracts (Fig 2C, blue line), with the frequency plot forming a tight normal distribution around the 1:1 ratio region. Two distinct sub-populations can be seen by the same analysis of data derived from purified samples (Fig 2C, red line). While there is a large peak of proteins with ratio 1:1 (log2 = 0), consistent with these being non-specific purification contaminants, there is also a smaller, broader peak in the region of log2 ratio of 1 to 6. These proteins are much more abundant in nickel purifications from HA-HisSUMO2 expressing cells compared to those from HA-His only cells, and so are likely to be SUMO2 conjugates. Ratio cut-offs for H/M and M/L were defined such that an estimated false discovery rate of less than 1% for SUMO2 substrates was applied (see Methods for further details), giving 877 putative SUMO2 conjugates (including multiple isoforms of some proteins such as PML), as listed (S1 Table, sheet 2). This method of SUMO2 substrate identification was validated by assessing the difference between the apparent MW of proteins based upon gel retention, and their predicted MW by sequence alone (Fig 2D) (see [29] for details of the method). The substrates and non-substrates clearly form two independent distributions in frequency plots, with substrates running in gels on average 20 kDa heavier than expected. Furthermore, comparison with a previous SUMO2 proteome analysis at the level of identified modification site [26] indicated 324 proteins in common with this set, while comparison with several major SUMO2 proteome studies revealed 521 proteins in common [23–28] (S1 Table, sheet 2, column P). In summary we can be confident that this list of 877 proteins represents true cellular SUMO2 substrates under these experimental conditions. A complete listing of all the data on the cellular proteins identified is presented in S1 Table. In addition to these cellular proteins, 71 viral proteins (i.e. all but one of the major viral polypeptides, the exception being US5) were identified in the crude sample (S2 Table).
H/L ratios can be used to study changes to either the total proteome (via ‘crude’ data), or the SUMO proteome (via ‘pure’ data) upon HSV-1 infection. Frequency distribution charts comparing ‘crude’ ratios from infected cells and uninfected cells showed few changes in total protein levels (Fig 2E, blue line). However, while most proteins in purified preparations also showed no change in abundance, the distribution is skewed towards larger ratios (Fig 2E, red line), with putative SUMO2 substrates being mostly responsible for this high ratio tail (Fig 2E, insert). This shows that the non-substrates are largely unchanged during HSV-1 infection, while SUMO2 conjugates have a tendency toward high ratios, indicative of a widespread loss of SUMO2 conjugation during infection. This is consistent with western blot data (Fig 1A). To test the reproducibility of these data a similar quantitative proteomic experiment was undertaken, this time only including Light infected and Heavy uninfected samples (S2A Fig). Although the total number of proteins was lower in this compared to the triple labeled experiment (S2B Fig), there was considerable overlap between the proteins in the purified fractions (S2C Fig), and the details of their ratio changes correlated substantially (S2D Fig; S3 Table). Because the triple labeled experiment was both larger and included the His-only control, subsequent sections mostly refer in detail only to this dataset.
To determine which SUMO2 substrates changed significantly during HSV-1 infection, Significance B (SigB) values were calculated using Perseus from the MaxQuant suite of software [30]. SigB is calculated using both signal intensity and SILAC ratio, and is indicative that a protein abundance change significantly deviates from the bulk of the quantified proteins. Of the 877 putative sumoylated proteins, 260 changed in abundance (either up or down) in the purified fraction of the triple SILAC experiment with SigB values of less than 0.1 (S1 Table, sheet 3). This number includes duplicate entries for proteins, such as PML, for which more than one isoform was detected. Removal of these duplicates and restricting the list to entries with H/L ratio increases of 2-fold or more results in a list of 185 proteins, shaded according to degree of change, and listed in order of decreased abundance (Fig 3). An additional 18 proteins on the putative sumoylated substrates list were detected with H/L ratios of 2 or greater and SigB values of less than 0.1 in the purified fraction of the replicate experiment, but which were not recorded as significantly regulated substrates in the primary experiment (S4 Table). As shown below, at least one of these is an authentic sumoylated substrate whose abundance decreases during HSV-1 infection. Overall therefore, up to around 200 cellular proteins identified in the purified fractions of the experiments decreased significantly in abundance during infection.
The degree of change in abundance of the putative sumoylated forms of these proteins varies considerably, up to a maximum of greater than 20-fold. In broad view, these data are consistent with previous observations on the decreased stability of sumoylated cellular proteins during HSV-1 infection. They also support the idea that there is considerable specificity or selectivity to the extent of sensitivity to desumoylation, as two-thirds of sumoylated proteins remain largely unaltered during infection while only 14% and 1% decrease in abundance by over 3-fold and 7-fold respectively.
We also identified a further 72 proteins that were not defined as potentially sumoylated on the basis of H/M ratios, which nonetheless had H/L ratios of greater than 2 and SigB of less than 0.1 in the purified fraction (S5 Table), although only 32 of these also complied with these criteria in the double labeled experiment. These may represent proteins that are non sumoylated, but which have an affinity for the Ni-agarose beads and whose abundance decreases during HSV-1 infection.
A small number of putative sumoylated proteins listed in S1 Table sheet 3 also exhibited H/L ratios of greater than 2 in the crude fraction (S6 Table), indicating candidate SUMO2 substrates whose total protein levels also reduced in infected cells. In the cases of IFI16, CENPB and PML, this has been reported previously [19,31–33]. Another protein on this list (NACC1) will be considered below, and further data presented below suggests that this list does not include all proteins that behave in this manner.
The levels of a number of proteins were changed significantly in the crude samples (S1 Table, sheet 4). Of these, and excluding those already noted above as regulated SUMO2 substrates, 128 proteins gave H/L ratios of greater than 2 in the crude samples (S7 Table). Given that HSV-1 infection causes the shut-off of host protein synthesis through induction of mRNA instability [34], this list is perhaps shorter than might be expected. It is possible that low abundance proteins with short half-lives (and thus those most susceptible to decreased host transcription) have not been detected efficiently in the ‘crude’ preparations by this approach, which will naturally favor the most abundant cellular proteins.
An unexpected finding concerns a small group of cellular proteins whose degree of sumoylation appears to contradict the general trend of deconjugation, and actually increases during infection. S8 Table shows proteins with H/L ratios in the purified fraction of less than 0.5 and SigB values of less than 0.1. Several of these proteins are components of the basic transcriptional apparatus (MED9, TAFs 1, 9 and 12) or transcription factors (MAFA, MAFB). This may indicate overall changes in transcription complexes as infection progresses. Validation of an example of a protein in this category (ZBTB7A) will be presented below. Similarly, in the crude fraction a small number of cellular proteins increased in overall abundance by a factor of 2-fold or more and with SigB values of less than 0.1 (S9 Table). Perhaps surprisingly, no induction of interferon-stimulated genes was detected under these infection conditions.
By applying the same method for monitoring apparent in gel molecular weights of cellular proteins (see Fig 2D) we were able to investigate the difference between apparent and predicted molecular weights of viral proteins. Most viral proteins were found in gel slices that were consistent with their predicted molecular weights, although the UL26 capsid maturation protease exhibited higher gel mobility than expected by sequence alone, consistent with its known cleavage during capsid assembly and also the production of its C-terminal half (protein VP22a) as a separate protein from an independent transcription unit [35,36]. Several proteins, however, had lower gel mobilities than predicted. In the case of glycoproteins gC, gL, gK and gM this is likely due to glycosylation (S2 Table). A small group of viral proteins exhibited decreased gel mobility over that predicted in the purified but not the crude sample, and in most of these cases the size difference could be consistent with sumoylation (Fig 4A). Note that the predicted molecular weight of some of these proteins differs considerably from their established gel mobilities, and in the case of ICP0, for example, this has been attributed to the nature of the primary sequence rather than post-translational modification. Therefore we investigated whether putative sumoylated species could be detected by western blotting of purified extracts from infected HA-HisSUMO2 cells.
Analysis of UL42, the processivity factor for the viral DNA polymerase, in the crude and purified samples of infected HA-HisSUMO2 cells revealed clear evidence of slower migrating species whose mobility is consistent with sumoylation (Fig 4B). This is of interest because the analogous protein (UL44) of HCMV is sumoylated [37]. However, these putative sumoylated UL42 bands were not clearly detectable in the crude fraction of HA-HisSUMO2 cells, nor were they evident during a normal wt HSV-1 infection of control HepaRG cells (Fig 4C). It is possible that over-expression of SUMO2 in the HA-HisSUMO2 cells forces a sumoylation event, or shifts the sumoylation equilibrium so that such species become more detectable. Therefore, while the evidence indicates that sumoylation of UL42 can occur, the likely sumoylated species seem to be in very low abundance during the course of a normal infection. We also detected likely sumoylated forms of UL6 and ICP0 in the purified fraction (albeit for the latter only on very long exposures of the blot) (Fig 4D), and extended exposure of the UL12 samples also revealed a possible sumoylated form. For the other proteins on the list of Fig 4A, we were either unable to detect sumoylated species by western blot of purified fractions (Fig 4D), or we lacked the reagents required to perform the analysis. Considering the scale of the proteomic method, it is conceivable that the sensitivity of the mass spectrometric approach is higher than the western blots shown here, and the possibility that all these proteins have a sumoylated component cannot be excluded. Analysis of the sequences of these proteins for potential sumoylation sites revealed consensus modification sites in US3, UL12 and UL42, but not the others.
As an initial step towards functional analyses of the proteins showing the greatest degree of change in sumoylation during HSV-1 infection, we grouped the 124 proteins of Fig 3 and S3 and S4 Tables with greater than 3-fold increases in H/L ratios in the purified fractions (pooling the data from experiments 1 and 2), then grouped them into broad, sometimes overlapping, categories (Fig 5). The proteins in each category are listed in order of degree of H/L ratio change and shaded as in Fig 3. The largest group of proteins include those with zinc finger domains, followed by transcription factors and chromatin-related proteins. BTB proteins, many of which have an additional zinc finger domain (the ZBTB proteins), form another marked group. There is a group of nuclear structure components such as lamins, and the PML NB components PML, Sp100 and MORC3, and also three centromere proteins. Several centromere proteins are already known to be degraded during HSV-1 infection in an ICP0-dependent manner [33,38,39]. Other groups of proteins have functions in RNA metabolism, interferon related pathways, and general metabolism (such as kinases).
It is striking that so many ZNF and ZBTB proteins were identified as regulated substrates, opening the question whether the zinc finger or the BTB domain of itself is contributing to the sensitivity of the sumoylated forms of these proteins to HSV-1 mediated degradation. We identified approximately 200 proteins with ZNF in the gene name (this will not include all proteins that include a zinc finger) of which 59 were sumoylated candidates and 24 of these were reduced in abundance by 3-fold or more. For ZBTB proteins, 19 in total were detected, 18 of which were sumoylation candidates and 11 were reduced in abundance by 3-fold or more. Thus the presence of a zinc finger or ZBTB domain itself does not generally confer sensitivity to HSV-1, but rather it seems that the proportion of these classes of proteins that are subject to sumoylation is increased compared to the bulk of cellular proteins.
Analysis of the functional consequences of these changes in abundance of these proteins during HSV-1 infection is obviously beyond the scope of this study, but the results certainly provide many novel avenues to pursue. Especially for those proteins undergoing the most dramatic changes in sumoylation, and in some cases overall abundance, there will inevitably be substantial disruption of the pathways in which they are involved during HSV-1 infection. Known examples of this include disruption of PML NBs and centromeres. But it is also reasonable to expect that the effects on chromatin related proteins and transcription factors will have consequences to chromatin structure or modification and transcriptional activity. Given that previous SUMO proteomic studies have highlighted that sumoylation of the proteins that are involved in these pathways is common [23–28], it is not surprising that these pathways feature prominently amongst those that are potentially disrupted by HSV-1 infection. While many such proteins may be innocent victims that are affected simply because of their sumoylation status, it is likely that this analysis includes previously unrecognized proteins which impact on the efficiency of HSV-1 infection.
The proteomic data for several previously studied proteins were consistent with their established behaviour during HSV-1 infection. For example, the H/L ratios in both crude and purified fractions were increased for PML (for which both sumoylated and unmodified forms are known to be degraded [18,19]), while the Sp100 H/L ratio increased only in the purified fraction (consistent with the loss of only the sumoylated forms [40,41]), and there was no change in H/L ratio for RanGAP1 (which is neither degraded nor regulated at the level of sumoylation during HSV-1 infection [19]) (S1 Table, sheet 2; and S3 Fig). We analyzed a number of proteins listed in Fig 3 that had not been previously investigated, selected on the basis of being amongst those with the greatest changes in H/L ratios, or being representatives of groups of related proteins, and on antibody availability. ZBTB4, ZBTB10, ZBTB38, NACC1 and MORC3 all exhibited high H/L ratios in the purified sample, while NACC1 and to a lesser extent MORC3 and ZBTB4 also showed high H/L ratios in the crude samples, indicative of reduced total protein amounts (peptides for ZBTB10 and ZBTB38 were not detected in the crude sample). Total protein and His-purified extracts of uninfected and infected HA-HisSUMO2 cells and uninfected HA-His only cells were blotted for the above proteins, using PML and RanGAP1 as controls (Fig 6). The sumoylated forms of PML were readily identified in the purified sample of uninfected HA-HisSUMO2 cells, and both these and the major unmodified form were degraded during infection. In contrast, sumoylated RanGAP1 was stable (Fig 6, upper left panels). Similarly, a sumoylated form of NACC1 was detected, and both this and the non-sumoylated form were degraded. Analysis of the other proteins was complicated by likely non-specific bands, but in all cases bands consistent with sumoylated forms were detected in the purified fraction, and in all cases except ZBTB7A (see below) these were lost during infection. For ZBTB10 and MORC3 (and ZBTB4 to a lesser extent), likely unmodified forms (marked by asterisks) were also diminished during infection. The anti-ZBTB38 antibody was particularly prone to detection of potentially spurious bands, but likely sumoylated forms were clearly detected in the uninfected sample and lost during HSV-1 infection (Fig 6, upper right panel, see also below). Therefore, where reagents of sufficient quality are available, these results validate the SILAC data with a high degree of success. They also reveal a number of proteins whose apparently unsumoylated forms are also degraded during HSV-1 infection, and may therefore constitute previously unrecognised substrates of ICP0.
ZBTB7A was selected as a representative protein with a low H/L ratio in the purified fraction, potentially indicating an increase in abundance of sumoylated forms following infection. Potential sumoylated species of ZBTB7A were detected, albeit weakly, in the purified sample, and these were of increased abundance in the infected sample (Fig 6), again consistent with the proteomic data.
By analyzing total protein extracts over time following HSV-1 infection of normal human fibroblasts, we found that the major forms of NACC1, ZBTB10, ZBTB38 and MORC3 were all degraded within 3 or 6 h (Fig 7). In these cells, the slower migrating sumoylated forms were not generally detected in the total protein extracts, although in the case of ZBTB4 the major form of the protein appeared stable while a potential sumoylated species was rapidly lost. The difference between the fate of the major form of ZBTB4 in this Fig compared to that in Fig 6 may be due to cell type, as the latter was performed in HA-HisSUMO2 cells. CITED2 was included in this analysis as an example of a protein with a high H/L ratio in the purified fraction, yet for which there was no evidence of authentic sumoylation (S1 and S3 Tables). This protein was also rapidly degraded during HSV-1 infection (Fig 7).
While confidence in the reliability of the proteomic data is strengthened by the above results, the analysis is limited by the quality of available antibodies. Therefore we selected further candidate proteins for study using an inducible expression system [42]. This allows addition of an epitope tag and expression in a high proportion of transduced cells at levels that could be controlled by the length of time of induction. The proteins selected included some analyzed in Figs 5 and 6 (ZBTB4 and ZBTB10) and several more for which antibodies either gave ambiguous results or were unavailable (BEND3, ETV6, MBD1, ZBTB12 and ZBTB20, all of which are amongst those with extreme H/L ratios; Fig 3). ARID3A was included in this set because preliminary analysis of the data of the experiment of S2 Fig identified it as a protein in the purified fraction that was sensitive to HSV-1 infection, consistent with published studies [43]. Although the H/L ratio of ARID3A was not significantly reduced in the experiment of Fig 2, it was identified as a likely sumoylated substrate (S1 Table, sheet 2) and a related protein (ARID4A) was reduced during infection (Fig 3). NACC2 was included in the analysis because it was the highest scoring protein ID of the experiment of Fig 2 that did not achieve the cut-off values of H/L ratio change in the triple labeled experiment. All the proteins were expressed in the inducible system (most by 2 h after induction), and most gave a major band close to the predicted molecular weight plus minor slower migrating species that are likely sumoylated products (S4 Fig). ZBTB4 was the least efficiently expressed of these proteins, and any sumoylated forms were below the level of detection in the presented exposure (but see Fig 8, in which putative sumoylated forms are visible).
The various cell lines were then treated with doxycycline for the appropriate length of time, the doxycycline was then washed out and the cells infected at moi 10 for 8 h with either wt or ICP0 null mutant HSV-1 (lanes marked dl). The extracts were also analyzed for the efficiency of viral gene expression in each instance (Fig 8). Where visible on the blot exposures presented, the slower migrating probable sumoylated forms were invariably lost during wt but not ICP0 null mutant infection, indicating that their loss is, directly or indirectly, dependent on ICP0. The major likely unmodified forms of ZBTB4, ZBTB10, ZBTB12, MBD1, BEND3 and NACC2 were also reduced to a greater or lesser extent in the wt virus infected cells (Fig 8). In contrast, potential sumoylated forms of ZBTB4, ZBTB10 and ETV6 appeared to increase in abundance in the ICP0-null mutant infected samples, which may be related to the overall accumulation of sumoylated species that occurs during the mutant virus infection [18,19]. The reduction in the major form for ZBTB4 observed here is consistent with the data of Fig 6, with both experiments being conducted in HepaRG-based cells, and in contrast to the infection time course of Fig 7 (performed in HF cells), supporting the possibility of cell type differentials in the behaviour of certain proteins during HSV-1 infection.
Taking into account all the proteins analyzed in Figs 5, 6 and 8, of the 124 proteins listed in Fig 5, five had been defined previously as decreasing in abundance during HSV-1 infection (although not all have well characterized sumoylated forms). We have also analyzed 11 further proteins that had not been studied in HSV-1 infection, finding that all showed evidence of sumoylated forms which were sensitive to HSV-1 infection, and in some cases their likely unmodified forms also. Thus of the 13% of the proteins in Fig 5 investigated, 100% were confirmed as behaving as predicted from the proteomic analysis. This very high confirmation rate gives much confidence about the overall validity of the proteomic analysis.
There are several mechanisms that could reduce the amounts of the candidate proteins identified in this study. The prime aim of the project was to identify sumoylated forms of proteins that are degraded in an ICP0-dependent manner, and it is likely that these constitute a major grouping. However, proteins that alter in abundance during HSV-1 infection that are either unsumoylated or modified to only a very minor degree may also be detected by this methodology, allowing the possible identification of substrates of ICP0 that are degraded in a SUMO-independent manner. Host cell proteins may also become less abundant during HSV-1 infection in an ICP0-independent manner, either through induced degradation as a consequence of virus infection in general or more passively because of reduced rates of host transcription. Distinguishing between substrates that are degraded by ICP0-dependent and-independent mechanisms during HSV-1 infection is not always straightforward due to the low infectivity of ICP0-null mutant HSV-1. For example, initial studies indicated that IFI16 was degraded by ICP0 during HSV-1 infection [32], but it later emerged that the degradation could occur in the absence of ICP0 in conditions in which the mutant infection was progressing as rapidly as the wt [31]. Therefore we have investigated the degradation of selected proteins of interest in cells induced to express ICP0 in the absence of infection (as described in [42]).
Control HA-TetR and ICP0 inducible cells were treated or not with doxycycline then the whole cell extracts were analyzed by western blotting. As reported previously [42], PML is efficiently degraded in this system (Fig 9, top left). The stability of the endogenous forms of proteins MORC3, ZBTB10, ZBTB38, NACC1, ZBTB4 and CITED2 after induction of ICP0 expression was analyzed because of the availability of antibodies that detect endogenous levels of the proteins. It is clear that the first four of these proteins can be degraded by ICP0 in the absence of infection, thereby identifying a number of previously undescribed ICP0 substrates. CITED2, on the other hand, appears only be degraded in HSV-1 infected cells, while the major form of ZBTB4 (but not a more slowly migrating, potential sumoylated form of the protein; see also Fig 7) appears to be relatively stable in the presence of ICP0. The relative loss of the major form of ZBTB4 during infection seen in Figs 5 and 8 does not seem to occur in the presence of ICP0 but the absence of infection, even though the protein was stable during infection with an ICP0-null mutant virus.
To expand the repertoire of proteins for which we could test the effect of ICP0 specifically we constructed vectors that expressed blasticidin resistance and myc tagged versions of BEND3 and MBD1 constitutively. ICP0 inducible cells were transduced with these vectors, then induction of ICP0 expression revealed that both proteins could be degraded by ICP0 alone (Fig 9B).
We investigated whether the degradation of selected proteins was influenced by the presence of SUMO2/3 by analyzing cells transduced by a lentivirus expressing multiple shRNAs that target both SUMO2 and SUMO3. These cells were highly depleted of SUMO2/3 (Fig 10A) and they exhibited highly reduced levels of sumoylated PML, disrupted PML NBs and increased replication of ICP0-null mutant HSV-1 (M. Glass, submitted for publication). In parallel, cells depleted of SUMO1 were also examined (Fig 10B). Depletion of the SUMO isoforms did not affect the efficiency of wt HSV-1 gene expression during high multiplicity infections (Fig 10C) and the remaining SUMO1 and SUMO2/3 proteins in these cells were sensitive to HSV-1 mediated reductions (Fig 10A and 10B). We analyzed examples of proteins for which antibodies that detect the endogenous proteins were available, finding that degradation of NACC1 was unaffected by depletion of SUMO2/3 (Fig 10D), while ZBTB10 was more stable in the SUMO2/3 depleted cells (but not the SUMO1-depleted cells) (Fig 10E). These results indicate that degradation of ZBTB10 is ICP0-dependent and is influenced by the abundance of SUMO2/3. ZBTB38 gave a marked interesting result, in that the bands detected by the antibody were clearly shifted in mobility in the SUMO2/3 depleted cells, but not in the SUMO1 depleted cells, and the novel ZBTB38 bands were lost in the infected cells (Fig 10F). These data indicate that ZBTB38 may be preferentially modified by SUMO2/3 compared to SUMO1, and its HSV-1 induced degradation is dependent on ICP0 but not on modification by SUMO2/3.
ICP0 includes several candidate SUMO interactions motifs (SIMs), one of which was shown by yeast 2 hybrid assay to bind to SUMO2, and was hence termed SIM-like sequence (SLS) -4 [18]. A recombinant HSV-1 mutant (mSLS4) was constructed with mutations in SLS4, which, although causing only a slight defect in itself, resulted in a highly defective phenotype when present in conjunction with mutations in other SLS motifs [44]. Mutant SLS4 was also found to be defective in degrading sumoylated forms of PML isoforms other than PML.I (when these were expressed in isolation), although it retained the ability to degrade PML.I in a sumoylation-independent manner [44]. As PML.I is the most abundant PML isoform and it interacts with all the others, mSLS4 retains the ability to degrade endogenous PML isoforms. We therefore investigated whether ICP0 lacking functional SLS4 could degrade a selection of the proteins of interest during infection either of normal cells or of transduced cell lines expressing a myc-tagged version of selected proteins, depending on antibody availability. The results indicated that the degree of degradation of a given protein could be influenced by the SLS4 mutation in ICP0 when expressed in the context of infection. For example, endogenous NACC1 appeared more stable during mSLS4 than wt virus infection whereas MORC3 was equally sensitive to the two viruses (Fig 11). We present three examples from the myc tagged protein data, including ZBTB20 (for which the sumoylated form is more resistant to loss in mSLS4 than wt virus infection, while the major band is relatively unchanged), MBD1 (for which all forms are reduced to a lesser extent in the mSLS4 compared to wt infection), and ZBTB10 (for which the major band is equally sensitive in the two infections while the abundance of likely sumoylated species increases) (Fig 11). The reasons for this last observation will be considered in the Discussion, but note that the sumoylated forms of tagged ZBTB10 also increased in abundance in cells infected with ICP0-null mutant HSV-1 (Fig 9).
Taken together, the results of Figs 9 to 11 indicate a range of factors which influence the reduction in abundance of these proteins during HSV-1 infection. These include via mechanisms for which ICP0 is insufficient (such as CITED2), or ICP0-dependent mechanisms which can be influenced or not by the abundance of SUMO2/3 (ZBTB10 or NACC1 respectively) and/or by the presence of the SUMO2 interaction motif of ICP0. This analysis illustrates the complexities that regulate cellular protein stability during HSV-1 infection, and each example requires careful analysis to determine all the factors involved; there is no one single simple mechanism at play.
This study is the first to report a comprehensive analysis of the cellular proteins whose abundance is affected by HSV-1 infection, with specific focus on proteins modified by SUMO2. The loss of SUMO-modified PML and Sp100 in an ICP0-dependent manner during HSV-1 infection has been reported previously [18,19,40,41,45], and these proteins play important roles in an intrinsic immune response to HSV-1 infection [9,10]. The bulk of SUMO-conjugated proteins are also reduced following infection with wt HSV-1 but increased during ICP0 null mutant infection [18,19]. The biological relevance of these observations was supported by the finding that disruption of the SUMO pathway through knockdown of Ubc9 enhanced the replication of an ICP0-null mutant HSV-1 [18]. These results prompted the question as to the identity of the affected SUMO-modified proteins and their role in the context of HSV-1 infection. To address these questions we used Mass Spectrometry (MS)-based quantitative proteomics analysis of HepaRG cells expressing His-tagged SUMO2 (HA-HisSUMO2 cells) with and without HSV-1 infection. The use of SILAC Light, Heavy and Medium media for HSV-1 infected and uninfected HA-HisSUMO2 cells and uninfected control HA-His only cells, respectively, allowed for the relative fold changes to be calculated for His-SUMO2 purified proteins and unmodified proteins following infection. This analysis identified with high confidence 877 cellular sumoylated proteins under these experimental conditions, of which 521 (59%) were in common with a compilation of the largest previous SUMO substrate identification proteomics studies [23–28]. Following HSV-1 infection 260 of these proteins changed in abundance with SigB values of less than 0.1, indicating some specificity in the targeted group of proteins; relative loss of sumoylated species is not simply a consequence of sumoylation per se, but the identity of the sumoylated species is also important. This is further illustrated by the fact that only 14% of these proteins decreased in abundance by over 3-fold and only 1% by 7-fold or more. Our analysis of specific proteins in this group has therefore been restricted to a subset of those most highly affected by HSV-1 infection.
Identification in our MS dataset of the sumoylated forms of PML and Sp100 as proteins whose abundance is significantly reduced during infection supported the validity of the experiment, as did the lack of change in sumoylated RanGAP1, which is not affected by HSV-1 [19]. Of the other proteins whose sumoylated forms reduced by a factor of at least 3-fold, a significant proportion may be linked to transcriptional regulation or chromatin related pathways (Fig 5). This supports the view that one of the most recognized functions of sumoylation is to regulate transcription [46–49]. SUMO-modification of transcriptional regulators is often described as having an inhibitory effect on transcription [47,50–52], however, there have been instances where SUMO-modification enhances transcription factor activity [53]. SUMO-modification of proteins in the context of transcriptional regulation may include transcription factors themselves, transcriptional co-regulators, and chromatin-remodeling proteins [46]. Sumoylation of these proteins may regulate their DNA binding activity, subcellular localization, assembly of multi-component complexes, interaction between transcription factors and co-regulators, and also DNA repair pathways and chromatin structure [46,54,55]. The reduction in the levels of the sumoylated forms of the number of proteins linked to these functions, and in some cases also their total abundance, would likely in normal circumstances have profound effects on the cell. In the samples analyzed here, the cells are subject to an extremely active and soon to be fatal infection, so even drastic changes to cellular gene expression may be inconsequential. The more interesting questions concern the consequences these changes may have on viral gene expression, and the potential identification of previously unrecognized preferential substrates for ICP0-mediated degradation. Therefore our analysis was driven first by authentication of a number of potentially interesting proteins, then by investigation of the role of ICP0 in their desumoylation or degradation.
We verified several example proteins on the list of sumoylation substrates that change during HSV-1 infection through a combination of western blotting of total protein extracts and His-SUMO2 purified proteins from mock and infected HA-HisSUMO2 cells (Fig 6), time course of degradation of endogenous proteins in whole infected cell extracts (Fig 7), and analysis of epitope tagged protein expression in stable cell lines (Fig 8). Evidence in favor of the sumoylation of these proteins, and the loss of these sumoylated forms (and in some cases their unmodified species) was obtained in one or more of these approaches. Overall, 13% of the proteins listed in Fig 5 were subjected to verification, with entirely positive results.
The loss of bulk sumoylated species following HSV-1 infection is more pronounced in human fibroblasts than HepaRG cells [18], and therefore we also analyzed the fate of several designated sumoylated proteins (ZBTB4, ZBTB10, ZBTB38, NACC1, and MORC3) over a time course of HSV-1 infection of HFs (Fig 7). All of these proteins decreased in abundance following wt HSV-1 infection, some as early as 3 h p.i. This suggests many proteins may be targeted much earlier than the 12 h time point used for the MS experiments (chosen to ensure complete infection and maximal effects). The available antibodies to these endogenous proteins do not always recognize a clearly sumoylated form, but it is clear that in most cases even the major unmodified form is sensitive to HSV-1 infection. Similar results were obtained in HepaRG cells, and these proteins remained stable during infection with ICP0-null mutant HSV-1.
In some cases, problems with lack of availability, specificity, or affinity of available antibodies were overcome using an inducible lentiviral expression system, in which the level of protein expression can be regulated by time of induction. In all cases, likely sumoylated bands were detected (S4 Fig) and these, together in some cases with the unmodified forms, were sensitive to HSV-1 infection (Fig 8). These proteins were however stable during infection with an ICP0 null mutant virus used at a multiplicity allowing an equivalent level of infection between wt and ICP0 null virus infected cells (Fig 8). To investigate whether ICP0 alone in the absence of infection is sufficient to decrease putative SUMO-modified protein bands, we utilized cells that can be induced to express ICP0 at levels equivalent to those during wt infection [42]. These experiments illustrated that ZBTB4, ZBTB10, ZBTB38, NACC1, MORC3, BEND3 and MBD1 all suffered a loss of total protein or likely sumoylated bands in the presence of ICP0, illustrating that ICP0 can cause this phenotype in the absence of other viral proteins. It is likely that the reduced abundance of many other proteins identified in this study is also ICP0-dependent. Whether ICP0 induces desumoylation or degradation of the sumoylated forms of these proteins requires further investigation, but in those instances where there is a loss of total protein suggests that the sumoylated species are being degraded rather than desumoylated.
We investigated the role of the sumoylation itself in substrate targeting in a limited number of instances through two approaches, firstly using SUMO depleted cells, and secondly analysis of the effects of a mutant form of ICP0 with a SUMO2 interaction motif inactivated. A complete analysis of these issues is beyond the scope of this paper, but the results demonstrate that sumoylation can play a role in the response of a protein to HSV-1 infection, but that the details may differ between proteins. It is intriguing that infection with the mSLS4 virus causes an increase in overall SUMO conjugate levels [44], and this is reflected in either reduced loss of presumed sumoylated forms of certain proteins during mSLS4 infection, or indeed an increase in their abundance (as in the case of ZBTB10; Fig 11). Presumed sumoylated forms of ZBTB10 and ZBTB4 also increase during ICP0 null mutant HSV-1 infection (Fig 8). These results confirm the role of sumoylation and SUMO-SIM interactions in some of the effects we observe. It is more difficult to determine whether these functions also impact on the degradation of the unmodified forms of selected proteins, as ICP0 can also target proteins in a sumoylation-independent manner [44], and it also includes other potential SIMs. We did not analyse the activity of a form of ICP0 that lacks multiple potential SIMs in this study because this mutant is highly defective in degrading all substrates previously analyzed [44].
Proteins containing a zinc finger (ZF) and/or BTB (broad-complex, tramtrack and bric-à-brac) domain were prominent on the list of those showing a greater than 3-fold reduction in the purified fraction following wt HSV-1 infection (Fig 5). These ZF and BTB domains are likely to bind DNA and mediate protein:protein interactions, respectively. Apart from being predicted to play roles in transcriptional regulation, the precise functions for most ZBTB proteins, the proteins they interact with and the genes they regulate are yet to be discovered [56]. Thus loss of sumoylation of ZBTB proteins may control transcription of cellular genes required for a successful immune response to the infection, or prevent repression of transcription of viral genes. In addition to potential roles in chromatin and transcription-related pathways, many BTB proteins are now known to act as substrate-specific adaptors for Cullin3-based E3 ligases [57] and hence play a role in modulating protein stability. Some of the known functions of the BTB and ZBTB and other proteins validated in this study are summarized in Table 1.
In addition to the ZBTB and BTB proteins listed in Table 1, we also validated some other proteins implicated in chromatin-related pathways and transcriptional repression, such as BEND3, MORC3, MBD1 and ETV6. It is intriguing that BEND3 promotes heterochromatinization and that sumoylation is important for its repressive activities [58], while MORC3 is a PML NB component that interacts with PML in a SUMO-dependent manner [59,60]. MBD1 is the largest of the methyl-CpG binding domain (MBD) family of proteins [61] that can also bind to unmethlyated DNA to mediate transcriptional repression [62,63] via its interaction with HDACs [63,64]. ETV6 is also a transcriptional repressor that can be sumoylated [65,66] and which interacts with transcriptional co-repressors and histone deacetylases [67–71].
While the biological significance of the changes in abundance and/or sumoylation status of these and the other proteins listed in Figs 3 and 5 remains to be determined, we note that of the 124 proteins whose sumoylation status changes by 3-fold or more during HSV-1 infection (Fig 5), four (PML, Sp100, ATRX and IFI16) have already been reported to affect the efficiency of ICP0 null mutant HSV-1 infection [9,10,12,31,72]. Given the number of proteins listed in Fig 5 that have potential roles in gene expression and chromatin-related pathways, it seems likely that future studies will reveal more such examples.
In addition to cellular proteins, we also identified a number of viral proteins that had higher than predicted molecular weights in the His-SUMO2 purified fraction. Excluding glycoproteins, these included nuclear protein UL3, capsid portal protein UL6, protein kinase US3, deoxyribonuclease UL12, tegument protein US11, DNA polymerase processivity subunit UL42, transcriptional regulator ICP4, ICP0, and regulatory protein ICP22 (Fig 4). There is no previous report of HSV-1 proteins being modified by SUMO, although the UL42 homologues expressed by HCMV (UL44) [37], and possibly that of EBV (BMRF1) [73] can be sumoylated. Thus the potential to be sumoylated may be a feature of the DNA polymerase accessory subunits expressed by herpesviruses. Of the HSV-1 proteins listed above, US3, UL12 and UL42 contain potential SUMO conjugation sites and analysis of His-SUMO2 purified proteins with available antibodies revealed likely sumoylated species, albeit in low amounts, for UL6, UL12, ICP0 and UL42. Further analysis will be required to investigate the extent and consequences of this potential sumoylation. In general, there are several reports of viral proteins subject to sumoylation, and in some cases this is an important aspect of their activity (reviewed in [74,75]).
In summary, post-translational modification of proteins with SUMO has important implications for protein function, and has roles in many pathways of importance for viral infections, including transcriptional regulation, and innate and intrinsic immune responses (reviewed in [74,75]). This proteomics study has provided a large data set opening many avenues of research for not only herpes virology, but also other areas where the sumoylation of proteins is of heightened interest. The validation using independent methods of the changes in abundance of many proteins identified by the MS approach lends considerable confidence to the overall utility of this study, and we have documented several previously unrecognized examples of proteins that are subject to ICP0 mediated degradation. In that a number of known biologically relevant substrates of ICP0 were identified in our study, it is likely that future studies on the basis of this analysis will reveal important novel aspects of the regulation of herpesvirus infection.
Human diploid foreskin fibroblasts (HFs, obtained from Dr Thomas Stamminger, University of Erlangen), HEK-293T human embryo kidney cells (American Type Culture Collection CRL-11268) and human osteosarcoma cells (U2OS, American Type Culture Collection HTB96) cells were grown in Dulbecco’s Modified Eagles’ Medium (DMEM) supplemented with 10% fetal calf serum (FCS). Baby hamster kidney cells (BHK-21, obtained from original Glasgow MRC Virology Unit stocks) were grown in Glasgow Modified Eagles’ Medium (GMEM) supplemented with 10% new born calf serum and 10% tryptose phosphate broth. HepaRG cells [76] were grown in William’s Medium E supplemented with 10% fetal bovine serum Gold (PAA Laboratories Ltd), 2 mM glutamine, 5 μg/ml insulin and 0.5 μM hydrocortisone. Derivatives of HepaRG cells expressing the tetracycline repressor (HA-TetR cells) and wt ICP0 in a doxycycline inducible manner (HA-cICP0 cells) have been described previously [18,42]. HepaRG cells transduced with lentiviruses expressing multiple anti-SUMO1 or a combination of anti-SUMO2 and antiSUMO3 shRNAs have been described elsewhere (M. Glass, manuscript submitted), as has the control lentivirus expressing an shRNA that does not target any human gene (shNeg) [11]. All cell growth media were supplemented with 100 units/ml penicillin and 100 μg/ml streptomycin. Lentivirus transduced cells were maintained with continuous antibiotic selection, as appropriate.
HSV-1 wild type (wt) strain 17 and mutant dl1403 [77] were the wt and ICP0-null mutant strains used. Virus mSLS4, which expresses a form of ICP0 with an inactivated SIM-like sequence SLS4, has been described previously [44]. These viruses were grown in BHK cells and titrated in U2OS cells, using 1% human serum in the overlay. ICP0 is not required for HSV-1 plaque formation in U2OS cells [78], therefore allowing a true comparison of the titres of wt and mutant virus stocks. Estimation of plaque formation efficiencies in cell lines expressing SUMO family members was performed using wt and ICP0 null mutant HSV-1 isolates (viruses in1863 and dl1403CMVlacZ respectively) that express a β-galactosidase marker gene linked to the HCMV promoter, as described previously [42]. Briefly, cells were seeded into 24-well dishes then infected with 3-fold serial dilutions of the viruses the following day. After 24 h incubation in the presence of additional 1% human serum, the cells were stained for β-galactosidase activity as described [42]. Relative plaque forming efficiencies were calculated by determining the number of plaques in each cell line at a given dilution of virus, then calculating fold changes in plaque number compared to controls cells at the same dilution. Averages and standard deviations were calculated from at least three independent determinations.
Lentivirus vector plasmid pLVX-6His-SUMO2, in which the human SUMO2 cDNA with a polyhistidine tag was inserted between the XhoI and XbaI sites of pLVX-IRES-Puro (Clontech), and pLVX-6His (a control with only the tag sequence) were kindly provided by Ben Hale. Plasmids with cDNAs of selected cellular proteins were either purchased from Source Bioscience or were gifts from Pierre-Antoine Defossez (MBD1 transcript variant 3, ZBTB4). The cDNAs were amplified by PCR using primers containing suitable restriction sites and encoding an in frame N-terminal myc tag, then the products were inserted in place of the ICP0 cDNA in doxycycline inducible lentiviral vector pLKO.DCMV.TetO-cICP0 (pLDT-cICP0) [42]. Lentiviral vectors expressing myc tagged versions of BEND3 and MBD1 in a constitutive manner were constructed by inserting the relevant cDNAs in place of the EYFP-PML cDNA in plasmid pLKOneo.gD.EYFP-PML.I [79] which had been modified by inserting the blasticidin resistance coding region in place of that for neomycin. Lentivirus transductions of were performed as described [9], with stable cell lines selected using puromycin (1 μg/ml, reduced to 0.5μg/ml for subsequent passage), G418 (0.5 mg/ml) or blasticidin (1 μg/ml), or combinations thereof, as relevant.
HA-HisSUMO2 and HA-His Only cells were cultured in SILAC DMEM lacking L-lysine and L-arginine, which were replaced with normal (light; L), heavy (H) or medium (M) stable isotopically labeled forms of these amino acids (all SILAC medium reagents were sourced from Cambridge Isotope Laboratories). These media were supplemented with 10% dialyzed fetal bovine serum (FBS), 100 units/ml penicillin and 100 μg/ml streptomycin and 0.5 μg/ml puromycin. Mock infected HA-HisSUMO2 cells were cultured in H medium (13C6 15N2-lysine, Lys8, and 13C6 15N4-arginine, Arg10), wt HSV-1 infected HA-HisSUMO2 cells were cultured in L medium (isotopically normal; Lys0, Arg0), and mock infected HA-His Only cells were cultured in M medium (4,4,5,5-D4-lysine, Lys4, and 13C6-arginine, Arg6). Cells were cultured for six population doublings in their respective SILAC media before being expanded into 12 x 150 mm dishes for each condition. HA-HisSUMO2 cells cultured in SILAC Light medium were infected at an MOI of 10 plaque forming units per cells for 12 h.
The method used was essentially as described [27]. Cells were washed twice with phosphate buffered saline (PBS), then lyzed in denaturing nickel sample buffer [6 M guanidinium hydrochloride (Merck), 94.7 mM Na2HPO4 (VWR Prolabo), 5.3 mM NaH2PO4 (VWR Prolabo), 10 mM Tris/HCl (Roche) pH 8.0, 20 mM imidazole (Sigma), 5 mM β-mercaptoethanol (Sigma), complete EDTA free protease inhibitor cocktail (Roche)], and stored at -70°C. Once thawed, equal amounts of protein from each Light, Medium and Heavy SILAC treatment (determined by BCA assay (Pierce) and confirmed by Silver staining after SDS-PAGE) were mixed and sonicated. Sonicated lysates were then centrifuged at 1,000 x g at 4°C for 10 min followed by passage through a 0.45 μm filter. Lysates were then added to 50 μl Ni2+ NTA agarose beads (Qiagen), pre-equilibrated with denaturing nickel sample buffer, and incubated with rotation at 4°C for 24 h. Beads were centrifuged out of suspension 1,000 x g at 4°C for 10 min. Beads were washed by centrifugation at 720 x g for 2 min in 1 ml buffer in a Lo-Bind Eppendorf tube as follows: once in denaturing nickel sample buffer, twice with wash buffer pH 8.0 [8 M urea (Sigma), 94.7 mM Na2HPO4, 5.3 mM NaH2PO4, 10 mM Tris/HCl pH 8.0, 20 mM imidazole, 5 mM β-mercaptoethanol, complete EDTA free protease inhibitor cocktail], twice in wash buffer with the pH reduced to 6.3, and once with a final wash in a new Lo-Bind Eppendorf tube in wash buffer pH 8.0. Bound proteins were then eluted from beads in 40 μl nickel resin elution buffer [2x LDS (Invitrogen), 1x reducing reagent (Invitrogen), 200 mM imidazole] at room temperature with agitation for 10 min, followed by boiling for 2 min. Samples were stored at -20°C.
For both quantitative proteomic experiments a ‘crude’ sample was prepared by TCA precipitation of proteins from a sample of the mixed lysates prior to nickel affinity purification; 35 μl of the protein mixture in the 6 M guanidine hydrochloride buffer (see above), containing about 70 μg of protein, was mixed with 400 μl 10% trichloroacetic acid (TCA), incubated in ice for 20 min and centrifuged at 19000 x g for 15 min at 4°C. The pellet was washed with 1 ml 100% ethanol at 4°C and re-centrifuged at 19000 x g for 15 min at 4°C. Supernatants were aspirated and the pellets dried in a gyrovap before resuspension in 80 μl 1.5 x LDS sample buffer containing reducing agent (Invitrogen). Then 35 μl of this and the nickel affinity chromatography elutions (representing between 20 and 40 μg of total protein) were fractionated by polyacrylamide gel electrophoresis containing SDS (NuPage 10% polyacrylamide, Bis-Tris with MOPS buffer—Invitrogen) and stained with Coomassie blue. For each experiment both crude and pure lanes were excised into identical slices according to apparent MW of markers, as indicated in Figs 2 and 3. Peptides were extracted from each slice by tryptic digestion [80], including alkylation with chloroacetamide.
Peptide samples were analyzed by LC-MS/MS on a Q Exactive mass spectrometer (Thermo Scientific) coupled to an EASY-nLC 1000 liquid chromatography system via an EASY-Spray ion source (Thermo Scientific) running at 75 μm x 500 mm EASY-Spray column. Elution gradient durations of 150 min and 240 min were used. Data were acquired in the data-dependent mode. Full scan spectra (m/z 300–1800) were acquired with resolution R = 70,000 at m/z 400 (after accumulation to a target value of 1,000,000 with maximum injection time of 20 ms). The 10 most intense ions were fragmented by HCD and measured with a target value of 500,000, maximum injection time of 60 ms and intensity threshold of 1.7e3. A 40 second dynamic exclusion list was applied.
Raw MS data files were processed together with the quantitative MS processing software MaxQuant (version 1.3.0.5) [30,81] Enzyme specificity was set to trypsin-P as required. Cysteine carbamidomethylation was selected as a fixed modification and methionine oxidation, protein N-acetylation and gly-gly adducts to lysine were chosen as variable modifications. The data were searched against a target/decoy human database in addition to the HSV-1 database (GenBank accession number JN555585.1). Initial maximum allowed mass deviation was set to 20 parts per million (ppm) for peptide masses and 0.5 Da for MS/MS peaks. The minimum peptide length was set to 7 amino acids and a maximum of four missed cleavages. 1% false discovery rate (FDR) was required at both the protein and peptide level. The ‘match between runs’ option was selected with a time window of two minutes. Data were output twice; firstly separated by ‘crude’ and ‘pure’ conditions, and secondly such that each digestion of each gel slice was considered a single ‘experiment’. The former is used for overall protein ratio changes, and the latter for apparent MW analysis [29].
In the triple SILAC labeled experiment one condition represented non-infected cells expressing the 6His sequence only, (HA-His Only cells) (See Fig 2). This was in the isotopically ‘medium’ condition (M), and so any ratio comparing this condition with the ‘light’ (L) or ‘heavy’ (H) conditions where 6His-SUMO-2 was expressed, i.e. M/L and H/M can be used as comparison between HA-HisSUMO2 and HA-His Only purifications. As SUMO2 substrates by definition should be more abundant in H or L conditions than M, substrates will be characterized by large H/M and small M/L ratios. For log2 M/L and log2 H/M ratios, two cutoffs of <-1.790 and >0.822 respectively were used. If applied to pure ratios this defined 877 putative SUMO2 substrates, representing 30.8% of all identifications. The same criteria applied to the crude ratios shortlisted 36 proteins, representing 0.65% of all identifications. By this method the false discovery rate for SUMO2 substrate definition is estimated to be below 1%.
For analysis of whole cell extracts, cells were seeded into 24-well dishes at 1 x 105 cells per well, then infected or treated with doxycycline the following day, as described in the figure legends. Cell monolayers were washed twice with PBS before harvesting in SDS-PAGE loading buffer. Proteins were resolved on 7.5% SDS-polyacrylamide gels, then transferred to nitrocellulose membranes by western blotting. Antibodies directed against the following proteins were used: 6xHis monoclonal antibody (mAb) (ab18184, Abcam), SUMO2/3 rabbit polyclonal antibody (rAb ab3742, Abcam), RanGAP1 (mAb 33–0800, Invitrogen), PML 5E10 mAb [82], Sp100 rAb SpGH [5], actin (mAb AC-40, Sigma), myc tag (mAb 9E10, Santa Cruz), β-tubulin (mAb T4026, Sigma), NACC1 (rAb ab29047, Abcam), ZBTB10 (rAb A303-257A, Bethyl), ZBTB38 (affinity purified rAb prepared by PRIMM), ZBTB4 (rAb #120/4, a gift from Pierre-Antoine Defossez), MORC3 (rAb NBPI-83036, Novus Biologicals rAb), CITED2 (rAb EPR3416(2) Abcam), ZBTB7A (rAb ab123075, Abcam). The sources of antibodies to HSV-1 proteins ICP0 (mAb 11060), ICP4 (mAb 58S) and UL42 (mAb Z1F11) have been described previously [42]. Monoclonal antibody 175 to detect UL6 and rabbit antibody BWp12 for UL12 were kindly provided by Frazer Rixon and Nigel Stow, respectively. Secondary antibodies include horse radish peroxidase conjugated goat anti-rabbit IgG (whole molecule) (Sigma A0545) and goat anti-mouse IgG (whole molecule) (Sigma A4416).
|
10.1371/journal.ppat.1002084 | The Intrinsic Antiviral Defense to Incoming HSV-1 Genomes Includes Specific DNA Repair Proteins and Is Counteracted by the Viral Protein ICP0 | Cellular restriction factors responding to herpesvirus infection include the ND10 components PML, Sp100 and hDaxx. During the initial stages of HSV-1 infection, novel sub-nuclear structures containing these ND10 proteins form in association with incoming viral genomes. We report that several cellular DNA damage response proteins also relocate to sites associated with incoming viral genomes where they contribute to the cellular front line defense. We show that recruitment of DNA repair proteins to these sites is independent of ND10 components, and instead is coordinated by the cellular ubiquitin ligases RNF8 and RNF168. The viral protein ICP0 targets RNF8 and RNF168 for degradation, thereby preventing the deposition of repressive ubiquitin marks and counteracting this repair protein recruitment. This study highlights important parallels between recognition of cellular DNA damage and recognition of viral genomes, and adds RNF8 and RNF168 to the list of factors contributing to the intrinsic antiviral defense against herpesvirus infection.
| The cellular DNA damage response pathway monitors damage to genomic DNA. We investigated whether cellular DNA damage response proteins can also respond to incoming viral genetic material and how they impact virus growth. Using Herpes Simplex Virus type 1 (HSV-1), we present evidence that DNA repair proteins are activated at the earliest times post-infection, and that they physically accumulate at sites associated with incoming viral genomes. A subset of these DNA repair proteins deposit repressive ubiquitin marks, recruit other DNA repair proteins, and limit transcription from the viral genomes. We demonstrate that the virus overcomes this anti-viral defense by targeting key DNA repair proteins for degradation. Our study adds these DNA repair protein mediators to the list of intrinsic antiviral defense factors active against HSV-1, and demonstrates that many aspects of the cellular recognition of foreign DNA parallel the recognition and response to cellular damage.
| Mammalian cells have evolved complex defenses to protect themselves from viral infections. Innate and adaptive immune responses are well-characterized, but resistance mediated by pre-existing cellular factors has recently emerged as another important arm of antiviral defense. In contrast to the canonical immune responses, which are slower acting and initiated by virus-induced signaling cascades, the pre-existing cellular factors are poised to protect the cell before the virus has even entered [1]. This mechanism of resistance is called intrinsic antiviral defense, and is characterized by the fact that the antiviral proteins are intracellular and constitutively expressed, and that the restrictive factors can be overcome by viral countermeasures. These intrinsic defense pathways provide a primary protective mechanism in the first cell infected in an immunologically naive host, making them an important front line of defense against viruses.
Herpes simplex virus type 1 (HSV-1) is a common human pathogen that causes life-long recurrent disease. Lytic HSV-1 infection is characterized by transcription in a temporal cascade of immediate-early (IE), early (E), and late (L) gene products. The immediate early (IE) genes create a favorable intracellular environment for the virus, and regulate the expression of the E and L genes. The IE protein ICP0 is one of the first viral proteins expressed during HSV-1 infection (reviewed in [2]). Although ICP0 is a not an essential viral protein, its deletion significantly impairs productive replication, especially at low multiplicity of infection (MOI) [3]–[6]. ICP0 is a RING finger E3 ubiquitin ligase that induces degradation of several cellular proteins including the catalytic subunit of DNA-dependent protein kinase (DNA-PKcs) [7], the cellular DNA damage ubiquitin ligases RNF8 and RNF168 [8], components of the nuclear domain structures known as ND10 (or PML nuclear bodies) [9], [10], and centromeric proteins [11]–[13].
Prototypic intrinsic antiviral defense proteins, such as APOBEC3 proteins, are known to be active against a variety of viruses [1]. However, to date, the only proteins demonstrated to mediate intrinsic defense against herpesviruses are all components of ND10. The first evidence that these proteins may mediate intrinsic immune defense against herpesviruses came from the observation that depletion of PML increased the plaque-forming efficiency of both human cytomegalovirus (HCMV) [14] and ICP0-null HSV-1 [15]. Similarly, it was found that the ND10 proteins hDaxx and ATRX induce a repressive viral chromatin structure on incoming HCMV genomes that is prevented by the viral tegument protein pp71 targeting hDaxx for degradation [16]–[21]. Depletion of either hDaxx or ATRX also improves the plaque-forming efficiency of ICP0-null HSV-1, providing further evidence that ND10 proteins have a general role in mediating intrinsic antiviral defense against herpesviruses [22]. In the case of HSV-1, the repressive ND10 proteins have been detected at sites juxtaposed to incoming viral genomes [22]–[24]. During wild-type HSV-1 infection ICP0 rapidly disperses these inhibitory proteins, ensuring that replication can proceed. In the absence of ICP0, the recruitment of the ND10 proteins into novel structures associated with the viral genomes is readily observable as a very early cellular response, detectable within the first 30 minutes of infection [25]. ICP0 has therefore emerged as one of the key viral counterattacks to the cellular attempt to limit the early stages of infection.
Cells have elaborate machinery in place to monitor damage to genomic DNA and ensure the fidelity of replication [26]. Recent work has demonstrated that the cellular DNA repair machinery can also recognize viral genetic material [27]. HSV-1 has a complex relationship with the DNA damage response, in that it appears to activate many components of the ATM-dependent arm of the signaling pathway, while inhibiting the DNA-PKcs- and ATR-dependent arms [7], [28]–[30]. During lytic infection, HSV-1 recruits several cellular DNA repair proteins into viral replication compartments where they enhance viral replication [28]–[31]. Despite global activation of the ATM-dependent signaling pathway, we recently reported that RNF8 and RNF168, which are key mediators in this pathway, are targeted for proteasome-mediated degradation by ICP0 [8].
During HSV-1 infection, the viral capsid docks at the nuclear pore and the linear viral genome is released into the nucleus [32]. In this study, we asked whether the cellular DNA repair machinery recognizes this incoming viral DNA, and we explored the significance of ICP0-mediated degradation of RNF8 and RNF168 for the virus. We report that cellular DNA repair proteins respond to incoming HSV-1 genomes and we identify RNF8 and RNF168 as novel components of the intrinsic antiviral defense against HSV-1.
In order to investigate effects of incoming HSV-1 genomes on localization of DNA damage proteins, we utilized a previously described assay to visualize nuclei at the earliest stages of infection [23], [24]. In this assay, cells are infected at low multiplicity so that directional viral spread through developing plaques can be analyzed. This directionality, combined with the fact that incoming viruses often congregate near the microtubule organizing center, means that nuclei of cells at the edge of plaques frequently display an asymmetric arc of incoming viral genomes [23], [24]. Human foreskin fibroblast (HFF) cells were infected at low MOI with wild-type or ICP0-null HSV-1, fixed 24 hours post-infection (hpi) and processed for immunofluorescence. Sites of incoming viral genomes were detected by staining with antiserum to the viral DNA binding protein, ICP4, which has been previously shown to co-localize with viral genomes in this assay [23], and the localization of certain cellular DNA repair proteins (Figure S1A) was assessed. In mock infected cells there was minimal γH2AX staining, and the damage checkpoint mediators Mdc1, 53BP1, and BRCA1 were localized in a diffuse nuclear pattern. In cells infected with ICP0-null virus, we detected that γH2AX, Mdc1, 53BP1, and BRCA1 accumulated in distinct asymmetric arcs in close proximity to incoming viral genomes (Figure 1A). In cells infected with wild-type virus, γH2AX and Mdc1 still re-localized to sites associated with viral DNA, but 53BP1 and BRCA1 remained diffusely nuclear (Figure 1B, Figure S1C). 53BP1 accumulated at sites associated with incoming ICP0-null viral genomes when high MOI infection was performed in the presence of α-amanitin, suggesting that viral transcription may not be essential (Figure S1B). These data indicate that redistribution of 53BP1 and BRCA1 in response to incoming viral genomes is an early response to HSV-1 infection that is inhibited by ICP0. We quantified the effect using 53BP1 and γH2AX as examples of DNA repair proteins that accumulated near incoming HSV-1 genomes. We observed that γH2AX accumulated near incoming HSV-1 genomes in over 80% of cells in both the presence and absence of ICP0 (Figure S1C). In contrast, while 53BP1 accumulated near incoming HSV-1 genomes in approximately 90% of cells infected with ICP0-null virus, this was reduced to approximately 25% of cells in the presence of ICP0 (Figure S1C).
It has previously been reported that components of ND10, including hDaxx, PML, ATRX, and Sp100 accumulate at sites overlapping, but not precisely co-localizing with, incoming HSV-1 genomes [22]–[24]. We wished to determine if the virus-induced accumulation of DNA repair proteins we observed co-localized with either viral genomes or ND10 proteins. We found that while the γH2AX and 53BP1 staining co-localized, these DNA repair proteins did not co-localize with either ICP4 (representing viral genomes) or PML (representing ND10 proteins) (Figure 2A; see Figure S2A for the corresponding cytofluorograms). Despite this lack of co-localization, we observed a degree of overlap between the different structures. To analyze this, a Manders' overlap co-efficient [33] was determined for each image (Figure S2B). We observed that on average, approximately 50% of the PML signal overlapped with the ICP4 signal, whereas only 20% of the 53BP1 or γH2AX signal overlapped with the ICP4 signal. These data suggest that incoming viral genomes are more closely associated with ND10 proteins than DNA repair proteins, and that all three structures have subtly distinct sub-nuclear localizations.
Next, we investigated whether the accumulation of DNA repair proteins at sites of incoming viral genomes was dependent on major ND10 proteins. HepaRG cells depleted of PML or Sp100 [34] were infected with wild-type or ICP0-null HSV-1 and processed for immunofluoresence at 24 hpi. Infections in cells depleted for PML or Sp100 were indistinguishable from control cells with respect to γH2AX accumulation near incoming viral genomes in both the presence and absence of ICP0 (data not shown), while 53BP1 accumulated only in the absence of ICP0 (Figure 2B). Therefore, the recruitment of 53BP1 to incoming HSV-1 genomes and the ability of ICP0 to block this process are not dependent on either PML or Sp100. Taken together, these observations suggest that accumulation of ND10 proteins and DNA repair proteins are independent events occurring at distinct physical locations.
We recently reported that ICP0 expression leads to proteasome-mediated degradation of the cellular DNA repair proteins and histone ubiquitin ligases RNF8 and RNF168 [8]. We therefore investigated whether these proteins were responsible for coordinating the recruitment of 53BP1 to sites associated with ICP0-null viral genomes. We infected RNF8 depleted cells (Figure S3), or cells derived from a patient who has a biallelic mutation in RNF168 (RIDDLE cells, [35]) with wild-type or ICP0-null HSV-1 and assessed the recruitment of DNA repair proteins to incoming viral genomes (Figure 3A and B and Figure S4). During infection with wild-type virus, ICP0 expression prevented 53BP1 recruitment in the presence or absence of RNF8 and RNF168. However, in cells infected with ICP0-null virus, 53BP1 was not recruited to sites associated with incoming viral genomes in the absence of RNF8 or RNF168 (Figure 3A and B), despite the fact that γH2AX still accumulated (Figure S4). To determine if RNF8 and RNF168 themselves were recruited to sites associated with incoming viral genomes, we generated a cell line that could be induced to express GFP-tagged RNF8, or utilized RIDDLE cells complemented with a cDNA expressing HA-tagged RNF168. We observed that RNF168 clearly accumulated near incoming ICP0-null viral genomes (Figure 3C). Redistribution of RNF8 to the vicinity of HSV-1 genomes was also detectable, although this was weaker and more variable than recruitment of RNF168 (Figure 3C). Together, these data suggest that accumulation of RNF8 and RNF168 at sites associated with incoming viral genomes coordinates 53BP1 recruitment. This implies that the reason ICP0 targets RNF8 and RNF168 for degradation is to prevent recruitment of specific DNA repair factors to viral genomes, suggesting that this recruitment is detrimental to incoming virus during early stages of lytic infection.
In uninfected mammalian cells, a tightly controlled hierarchy of events occurs following the induction of DNA double strand breaks [36], [37]. RNF8 and RNF168 coordinate the recruitment of 53BP1 to sites of cellular damage [38] and also to sites associated with incoming viral genomes. We therefore predicted that the latter process would be disrupted by depletion of factors upstream of RNF8 and RNF168 in the DNA damage response pathway. Phosphorylation of the histone variant H2AX is one of the first events to occur after induction of a double stranded DNA break [39], [40] and it is required for sustained accumulation of factors such as 53BP1 at damage sites [41], [42]. Phosphorylated H2AX binds MDC1, which in turn recruits RNF8 in a phosphorylation-dependent manner, and this interaction tethers 53BP1 and other downstream mediators at damage sites [38]. H2AX is therefore upstream of RNF8 and RNF168, and stable foci of 53BP1 do not form in H2AX-null cells. We infected cells from mice deleted for H2AX or matched control cells [43] with wild-type and ICP0-null HSV-1, and examined cells at the edge of developing plaques. As predicted, 53BP1 was recruited to ICP0-null viral genomes in wild-type mouse embryonic fibroblasts (MEFs), but did not accumulate during infection of cells lacking H2AX (Figure 4A).
ATM and the Mre11 complex are also upstream regulators of the cellular response to DNA damage. The Mre11 complex senses DNA double strand breaks and facilitates activation of ATM by recruiting it to the break sites [44]–[46]. However, despite this upstream role, 53BP1 still accumulates at sites of cellular DNA damage in cells deficient in Mre11 complex members or ATM [47]. We therefore assessed the requirement for ATM and Mre11 in coordinating the recruitment of 53BP1 to sites associated with ICP0-null viral genomes. We infected cells from patients with ataxia telangiectasia (A–T) and ataxia telangiectasia-like disorder (A-TLD) that lack functional ATM and Mre11 respectively, and compared them to matched controls in which ATM or Mre11 had been reconstituted. We observed that neither ATM (Figure 4B) or Mre11 (Figure 4C) were required for the accumulation of 53BP1 at sites associated with incoming ICP0-null viral genomes. These data demonstrate that H2AX, RNF8 and RNF168 are required for accumulation of 53BP1 at sites associated with incoming viral genomes, but ATM and Mre11 are not required. This hierarchy of signaling and recruitment events in response to viral genomes parallels the response to cellular DNA damage. Our data therefore suggest that the host cell recognizes either the incoming viral genomes themselves, or the resultant changes in local chromatin structure induced by incoming viral genomes, as DNA damage.
RNF8 and RNF168 are ubiquitin ligases for the histone H2A [35], [48]–[51], and we have previously reported that ICP0 expression leads to loss of uH2A, concomitant with the degradation of these two ligases [8]. We therefore examined ubiquitin conjugation at the sites associated with incoming ICP0-null viral genomes. We infected RNF8-null MEFs, RIDDLE cells, and matched controls, with ICP0-null virus and examined conjugated ubiquitin staining (FK2) at sites associated with incoming viral genomes at 24 hpi (Figure 5A). Asymmetric FK2 staining was detectable only in cells expressing RNF8 and RNF168, suggesting that this represents uH2A, which we also detected associated with incoming ICP0-null viral genomes (Figure S5A). The FK2 signal co-localized with 53BP1, but not PML, at sites associated with incoming viral genomes, suggesting that conjugated ubiquitin was a marker for sites of DNA repair protein accumulation rather than sites of ND10 protein accumulation (Figure S5B).
SUMO modification has also recently emerged as an important regulator of cellular DNA damage signaling [52], [53] and SUMO conjugates have been detected at sites associated with incoming ICP0-null genomes (Cuchet-Lourenco, Boutell and Everett, unpublished observations). In the case of cellular DNA double strand breaks, SUMO1 and SUMO2/3 recruitment is dependent on RNF8 and RNF168 [52]. We therefore determined whether SUMO recruitment to sites associated with incoming ICP0-null genomes was also dependent on RNF8 and RNF168. We infected cells depleted for RNF8 or lacking functional RNF168, and their matched controls, with ICP0-null virus and analyzed cells at the edges of developing plaques for asymmetric accumulations of SUMO. Both SUMO1 and SUMO2/3 were recruited to sites associated with incoming ICP0-null genomes even in the absence of RNF8 or RNF168 (Figure 6A and B). ND10 proteins are heavily SUMOylated, and SUMO modified forms of PML and Sp100 are known to be targets of ICP0 [15]. We therefore speculate that at least some of the SUMO conjugates we detected in the absence of RNF8 and RNF168 may represent sites of ND10 protein accumulation rather than DNA repair proteins, an idea supported by the observation that PML is still recruited to these sites in cells depleted for RNF8 and RNF168 (Figure 6C and D). Together, these data show that recruitment of ND10 components and DNA repair proteins are independent events, sharing the common themes of being disrupted by ICP0 and likely being coordinated by SUMO modification events.
Accumulation of cellular factors at sites associated with incoming HSV-1 genomes has been strongly linked to restricting the invading virus [22], [34]. We therefore wished to determine the biological significance of the accumulation of specific DNA repair proteins at sites associated with incoming viral genomes.
First, we assessed the ability of wild-type or ICP0-null virus to form plaques on cells deficient for H2AX or matched control cells expressing wild-type H2AX. We observed that both wild-type and ICP0-null HSV-1 were approximately 10-fold more likely to form plaques in the presence of H2AX (Figure 7A). This is similar to our previous data demonstrating that certain DNA repair proteins, such as ATM and Mre11, are beneficial for HSV-1 replication [28], possibly via processing of intermediates generated during viral replication/recombination [54]. Even though γH2AX is excluded from viral replication compartments [55], this histone variant is one of the master regulators of DNA damage signaling, and it is likely that H2AX phosphorylation is required to activate or recruit specific downstream proteins required during viral replication.
Our FK2 data (Figure 5) suggested that ubiquitination events at sites associated with incoming viral genomes are regulated by RNF8 and RNF168. Since uH2A has well-characterized roles in silencing [56]–[58] and ICP0 is a known transcriptional activator, we hypothesized that one reason for ICP0 to target RNF8 and RNF168 is to limit transcriptional repression of incoming viral genomes. To test this hypothesis, we compared the transcriptional competence of viral genomes in the presence and absence of RNF8 (Figure 7B). Cells from RNF8 null mice transduced with empty retrovirus or retrovirus expressing human WT RNF8 [8] were infected with wild-type or ICP0-null HSV-1 and harvested at 2 and 5 hpi. RNA was isolated and reverse transcribed, and qPCR was performed to detect ICP27 transcripts as a marker of viral transcription. We confirmed that input DNA was similar in all infections (data not shown) and analyzed the data by comparing transcription in the presence of RNF8 to transcription in the absence of RNF8 (Figure 7B; see Figure S6A for transcript levels across all samples). We observed that a) both viruses were transcriptionally repressed by RNF8, b) this repression was more significant in the absence of ICP0 and c) RNF8-mediated repression decreased over time during wild-type but not ICP0-null virus infection, presumably as a consequence of RNF8 degradation (Figure 7B). These data indicate that RNF8 is transcriptionally repressive to HSV-1 genomes and explains why HSV-1 forms plaques less efficiently in the presence of RNF8 [8] and/or RNF168 (Figure S6B and C).
In this study we discovered that RNF8 and RNF168 coordinate a repressive barrier to incoming HSV-1 genomes, and that ICP0 targets these cellular ubiquitin ligases to overcome this host antiviral effect. We describe structures marked by DNA repair proteins and conjugated ubiquitin that form de novo in response to incoming viral genomes. These DNA repair structures are associated with, but are independent of, similar ND10-like structures that also form near incoming viral genomes.
Our studies highlight the complexity of the interface between HSV-1 and the cellular DNA damage response. Previous work demonstrated that certain recombination and repair proteins, such as Mre11, ATM, ATR/ATRIP, and WRN are beneficial for HSV-1 replication [28], [29], [31]. Here we show that H2AX is also required for optimal replication of HSV-1, as previously suggested [59]. In contrast, the NHEJ proteins, DNA-PKcs and Ku70, have been reported to be detrimental to HSV-1 replication [7], [31]. We found that RNF8 and RNF168 also inhibit replication, likely by creating a repressive environment at the nuclear sites of incoming viral genomes. Together, these observations suggest that HSV-1 temporally dissects the DNA repair pathway; this ensures that repressive proteins are degraded, while repair proteins required to coordinate signaling and facilitate replication or processing of viral genomes are retained.
Although accumulation of many cellular factors has been strongly linked to restricting the incoming viral genomes [15], [22], [34], recruitment does not necessarily always correlate with repression. For example, some PML isoforms accumulate at sites associated with incoming HSV-1 genomes but do not inhibit the plaque-forming ability of ICP0-null virus (Cuchet-Lourenco, Boutell and Everett, unpublished observations). Similarly, we observe γH2AX accumulation at sites associated with incoming viral genomes, but find that H2AX is required for optimal HSV-1 replication. In contrast, proteins involved in intrinsic antiviral defense are not only recruited to incoming genomes, but limit viral progression, and are therefore inactivated by the virus during the earliest stages of infection. Our data identify RNF8 and RNF168 as new members of the host cell antiviral arsenal against incoming HSV-1.
When HSV-1 genomes enter the nucleus, they do so as naked DNA. However, the cell responds by depositing repressive chromatin marks on the incoming nucleic acid [60]–[64]. In turn, the virus recruits modification complexes containing histone demethylases and methyltransferases, and installs positive marks to facilitate IE transcription [65]. These demethylases may act in concert with histone deacetylases, such as HDAC1, which bind the transcriptionally repressive coREST/REST complex in the absence of ICP0 [66], [67]. We observed γH2AX and uH2A in association with sites of incoming viral genomes at the earliest detectable stages of infection, suggesting that these post-translational histone modifications are a very early response to incoming viral DNA. Our co-localization studies raise the possibility that these modified histones may be deposited on the displaced host chromatin around the incoming viral genomes.
Our data highlight the emerging parallels between cellular recognition of viral DNA and the cellular response to DNA damage. In both cases, γH2AX is activated, Mdc1 accumulates, and downstream repair factors such as 53BP1 are recruited. Furthermore, both processes are coordinated by the ubiquitin ligases RNF8 and RNF168, and ICP0 is thus able to disrupt both by inducing the degradation of these cellular proteins (Figure 7C). However, in contrast to the situation at sites of cellular DNA damage, we observed that SUMO conjugates still accumulate near incoming viral genomes even in the absence of RNF8 or RNF168. This accumulation likely reflects SUMO modification of ND10 components, which we show are still recruited in the absence of RNF8 or RNF168. Recent work has demonstrated that the SIMs of PML, hDaxx and Sp100 are essential for their recruitment to virus-induced foci (Cuchet-Lourenco, Boutell and Everett, unpublished observations) raising the possibility that these ND10 components are recruited in response to upstream SUMO-dependent events at these sites. RNF168 is known to contain SIMs and its recruitment to sites of cellular damage depends on the SUMO ligase PIAS4 [52]. It will therefore be interesting to determine whether the SIMs in RNF168 are required for its accumulation near incoming viral genomes, and whether disrupting the SUMO pathway can abrogate accumulation of both ND10 proteins and DNA repair proteins at these sites. Conversely, it will be interesting to see if the accumulation of ND10 components at sites of cellular DNA damage [68], [69] is SUMO-dependent and whether this still occurs in the absence of RNF8 and RNF168.
It has recently been shown that sites of cellular DNA damage are characterized by transcriptional repression [70]. The parallels we have uncovered between recruitment of DNA repair proteins to sites of cellular DNA damage and to incoming viral genomes raise the possibility that silencing is a defining characteristic of both sites. We propose that the accumulation of SUMO conjugates, ND10 components and DNA repair proteins are hallmarks of a repressive cellular response to both damaged and foreign DNA.
Vero and U20S cells were purchased from the American Tissue Culture Collection. MEFs from RNF8-/- knockout mice and matched wild-type controls were obtained from Razq Hakem [71] or Junjie Chen [72] and for some experiments RNF8-/- MEFs were complemented with human RNF8 [8] were used. Human foreskin fibroblasts (HFFs), obtained from the University of California San Diego Medical Center, were kindly provided by Debbie Spector. Cells were maintained in Dulbecco modified Eagle's medium (DMEM) containing 100 U/ml of penicillin and 100 µg/ml of streptomycin, supplemented with 10% fetal bovine serum (FBS) and selection antibiotics as appropriate. Cells were grown at 37°C in a humidified atmosphere containing 5% CO2. HepaRG hepatocyte cells [73] were grown in William's medium E supplemented with 2 mM glutamine, 5 µg/ml insulin, and 0.5 µM hydrocortisone. H2AX-/- MEFs were obtained from Andre Nussenzweig [43]. A-T cells (AT22IJE-T) and matched ATM put-back cells were obtained from Yosef Shiloh [74]. A-TLD-1 cells and matched cells with Mre11 reconstituted were described previously [75]. The inducible RNF8-GFP cell line was constructed by cloning RNF8 into the previously described tet-inducible pLKO based expression system [76]. Cells were induced with 0.1 µg/ml tetracycline for 8 hrs.
shRNA targeting RNF8 was 5′ ACATGAAGCCGTTATGAAT 3′ as previously described [49]. This sequence was incorporated into a pLKO based (for HepaRG cells) or GFP-tagged HIV vector plasmids [77].
Parental virus HSV-1 strain was 17 syn+ and the matched ICP0 deletion mutant was dl1403 [5]. Viruses were grown in Vero cells and titered in U2OS cells, in which ICP0 is not required for efficient plaque formation. Infections were performed on monolayers of cells in DMEM with 0% FBS. After 1 hr at 37°C, virus was removed and media containing 10% FBS was added. For plaque edge experiments, this media was supplemented with 1% human serum to limit spread of the virus. For plaque assays, 24 well dishes were infected with three-fold dilutions of wild-type or ICP0-null HSV-1. After adsorption, the cells were overlaid with medium containing 10% FBS and 1% human serum. Plaques were stained with crystal violet 24–36 h post-infection. Pseudotyped lentiviral stocks were generated by transfecting 293T cells with the appropriate vector plasmid and pVSV-G, pRev and pMDL plasmids as previously described [77].
Primary antibodies were purchased from Bethyl (PML), Abcam (SUMO1 and SUMO2/3), Rockland (ATM S1981-P), Santa Cruz (BRCA1, 53BP1), Millipore (H2AX S139, FK2, H2A, uH2A), Calbiochem (BRCA1), Research Diagnostics Inc. (GAPDH), Covance (HA), Transduction Laboratories (DNA-PKcs), and Sigma (FLAG). Rabbit antisera to Mdc1 was from J. Chen. The 58S monoclonal antibody to ICP4 was generated from an ATCC hybridoma cell line [78]. All secondary antibodies were from Jackson Laboratories or Invitrogen.
For immunoblotting, lysates prepared by standard methods. For immunofluorescence, cells were fixed with 4% paraformaldehyde for 15 min and extracted with 0.5% Triton X-100 in PBS for 10 min. For certain antibodies, cells were pre-treated with 0.5% Triton X-100 in PBS for 10 min prior to fixation. Nuclei were visualized by staining with DAPI. Images were acquired using a Leica TCS SP2 confocal microscope.
2X106 cells were infected with WT or ICP0-null virus at an MOI of 0.01 and harvested at 2 and 5 hpi. 75% of the cell pellet was used for RNA extraction and 25% for DNA purification. 1 µg RNA was reverse transcribed using SuperScriptIII RT (Invitrogen) and oligo dT in a 20 µl reaction. qPCR was run in triplicate with 3 µl cDNA or 100 ng genomic DNA using SYBR Green PCR master mix (ABI) on an ABI 7900HT system. ICP27 transcript was detected using primers GCATCCTTCGTGTTTGTCATT (F) and GCATCTTCTCTCCGACCCCG (R) [65] and normalized to endogenous RPLPO transcript detected using primers CTGGAAGTCCAACTACTTCC (F) and TGCTGCATCTGCTTGGAGCC (R).
|
10.1371/journal.pgen.1003669 | Genome Wide Association Identifies Novel Loci Involved in Fungal Communication | Understanding how genomes encode complex cellular and organismal behaviors has become the outstanding challenge of modern genetics. Unlike classical screening methods, analysis of genetic variation that occurs naturally in wild populations can enable rapid, genome-scale mapping of genotype to phenotype with a medium-throughput experimental design. Here we describe the results of the first genome-wide association study (GWAS) used to identify novel loci underlying trait variation in a microbial eukaryote, harnessing wild isolates of the filamentous fungus Neurospora crassa. We genotyped each of a population of wild Louisiana strains at 1 million genetic loci genome-wide, and we used these genotypes to map genetic determinants of microbial communication. In N. crassa, germinated asexual spores (germlings) sense the presence of other germlings, grow toward them in a coordinated fashion, and fuse. We evaluated germlings of each strain for their ability to chemically sense, chemotropically seek, and undergo cell fusion, and we subjected these trait measurements to GWAS. This analysis identified one gene, NCU04379 (cse-1, encoding a homolog of a neuronal calcium sensor), at which inheritance was strongly associated with the efficiency of germling communication. Deletion of cse-1 significantly impaired germling communication and fusion, and two genes encoding predicted interaction partners of CSE1 were also required for the communication trait. Additionally, mining our association results for signaling and secretion genes with a potential role in germling communication, we validated six more previously unknown molecular players, including a secreted protease and two other genes whose deletion conferred a novel phenotype of increased communication and multi-germling fusion. Our results establish protein secretion as a linchpin of germling communication in N. crassa and shed light on the regulation of communication molecules in this fungus. Our study demonstrates the power of population-genetic analyses for the rapid identification of genes contributing to complex traits in microbial species.
| Many phenotypes of interest are controlled by multiple loci, and in biological systems identifying determinants of such complex traits is challenging. Here, we genotyped 112 wild isolates of Neurospora crassa and used this resource to identify genes that mediate a fundamental but poorly-understood attribute of this filamentous fungus: the ability of germinating spores to sense each other at a distance, extend projections toward one another, and fuse. Inheritance at a secretion gene, cse-1, was associated strongly with germling communication across wild strains; this association was validated in experiments showing reduced communication in a cse-1 deletion strain. By testing interacting partners of CSE1, and by assessing additional secretion and signaling factors whose inheritance associated more modestly with germling communication in wild strains, we identified eight other novel determinants of this phenotype. Our population of genotyped wild isolates provides a flexible and powerful community resource for the rapid identification of any varying, complex phenotype in N. crassa. The success of our approach, which used a phenotyping scheme far more tractable than would be required in a screen of the entire N. crassa gene deletion collection, serves as a proof of concept for association studies of wild populations for any organism.
| In most filamentous ascomycete species, hyphae form an interconnected network or syncytium of multi-nucleate cells known as a mycelium [1]. In nature, the formation of a mycelium often occurs via the germination of wind-dispersed asexual spores (conidia) [2]. Upon landing on a suitable substrate, conidia germinate to form germlings that are capable of fusion via specialized structures called conidial anastomosis tubes (CATs) to form the interconnected mycelial network common in this group of organisms [3], [4]. The formation of mycelial networks by germling fusion increases cytoplasmic flow and is important for the distribution of nutrients, signals and organelles within the colony [5], [6].
Similar to cell fusion in other organisms, the process of germling fusion in the filamentous ascomycete fungus Neurospora crassa requires cell recognition and attraction, adhesion, cell wall remodeling and membrane merger [7]. Genetically identical germlings of N. crassa exhibit remarkable chemotropism to each other, which enhances the formation of the inter-connected hyphal network [8], [9]. A number of mutants have been identified in N. crassa that fail to undergo germling and hyphal fusion, including nrc-1, mek-2 and mak-2, which are components of a conserved MAP kinase pathway [3], [10], [11], [12]. Other mutants of unknown biochemical function, such as soft (so), also show defects in chemosensing and cell fusion [13], [14]. The components of the MAP kinase pathway (NRC1, MEK2 and MAK2) and SO are recruited in a rapid and oscillatory manner to the plasma membranes of germling pairs undergoing chemotropic interactions [12], [14]. The oscillation of MAK2 and SO to CAT tips has been proposed to allow genetically identical cells to alternate between two different physiological states associated with signal delivery or response [14], [15], [16]. Given the complex physiology of cell communication and fusion, many other genes and proteins likely play a role in this process.
N. crassa is a heterothallic, obligate outbreeding species that has been a model for the study of population structure and genetic variability of fungi in the wild [17], [18], [19], [20], [21]. Recent advances in nucleic acid sequencing technologies have allowed for large-scale sampling of wild populations in this model microbe, and we recently harnessed this strategy in a population structure analysis of N. crassa by RNA-seq [21]. Data from such a sequencing survey provides a dense map of genetic variants across the genome and raises the possibility of genome-wide association studies in N. crassa. Association mapping is a powerful tool to identify candidate cases in which genetic variation at the DNA level underlies differences between wild individuals in a trait of interest. This approach is in common use in human genetics but has had little application to date in model organism systems, although recent work has established the power of association studies in mapping the genetic basis of trait variation across wild individuals in Drosophila [22], [23], [24], Arabidopsis [25], [26], [27], [28] and sunflower [29]. In fungi [30], [31] and in most other organisms beside humans, studies seeking to use natural variation as a screening tool to map genotype to phenotype have been largely limited to experimental cross designs, which survey polymorphisms in only a few wild individuals.
Here we describe the results of the first genome-wide association analysis used to identify novel loci underlying trait variation in a microbial eukaryote. We applied an association strategy using wild isolates of N. crassa to identify the genetic basis of the complex trait of germling communication. Developing a detailed, quantitative assay well-suited to the medium-throughput association-mapping paradigm, we surveyed germling communication across wild N. crassa strains and mapped differences in this trait to DNA sequence variants. We subsequently tested the function of genes mapped in our association study by assessing the germling communication phenotype of deletion strains, revealing mutants that showed both decreased and increased germling fusion frequency. We also tested the effect of some gene deletions on MAK2 and SO oscillation during chemotropic interactions. And we localized within hyphae the protein product of the gene that showed the most significant association with germling communication phenotype, a homolog of mammalian neuronal calcium sensor-1 (NCS-1).
Our previous study of the relatedness of wild N. crassa isolates from the Western hemisphere by RNA-seq revealed a well-defined population of 20 individuals from Louisiana [21]. To establish a larger set of genotyped Louisiana strains suitable for use in association mapping, we transcriptionally profiled an additional 92 Louisiana strains (Table S1). Analysis of the regulatory variation across the Louisiana population detected in these data will be reported elsewhere; here we used the RNA-seq reads to identify 1.09 million single-nucleotide polymorphisms (SNPs) in coding regions of the seven N. crassa chromosomes (Dataset S1). Phylogenetic analysis of these SNPs (Figure S1) indicated a set of 100 strains with little population substructure, including the smaller sample of Louisiana isolates that we had previously characterized [21]. We identified 81,614 SNPs at which the minor allele was present in >25% of strains, and which were equally distributed throughout the euchromatic regions of all seven chromosomes of N. crassa (Figure S2 and Dataset S2). Across the 9,730 protein-coding genes of the N. crassa genome (http://www.broadinstitute.org/annotation/genome/neurospora/MultiHome.html), the average gene harbored ∼10 high-frequency SNPs.
To use our genotyped Louisiana strains to dissect the genetics of germling communication, we first developed a communication assay as follows. When genetically identical macroconidia of the N. crassa laboratory strain FGSC 2489 germinate near each other, ∼89% of the germlings within 15 µm of other germlings sense their neighbors, reorient their growth, and engage in cell fusion via CATs [3] (Figure 1A). The remaining germlings ignore each other, do not show chemotropism, do not form CATs and do not fuse (Figure 1B). We thus quantified communication by isolating macroconidia from each given wild strain, plating them on agarose minimal medium, and tabulating the percent of germling pairs exhibiting redirected CAT growth (communication) or fusion after 3–4 hours of incubation. Applying this procedure to 24 Louisiana strains showed that the germling communication trait varied among the wild isolates, from a high of 90% communication/cell fusion efficiency to a low of less than 40% communication (Table S2 and Figure 2).
To map loci underlying the variation in communication efficiency/cell fusion across our wild population, we first scored patterns of germling interactions as a qualitative, binary trait, such that the phenotype of a given individual was classified as either avidly or poorly communicating. We then used our set of genotypes at high-frequency SNPs to test each locus in turn for co-inheritance with the communication trait across the strains of the population, using a permutation strategy, described in Methods, to correct for multiple testing. This mapping calculation yielded 3 SNPs showing significant association with germling communication at a threshold at which we expected ∼0.01 SNP by chance (Figure 3 and Dataset S3). All three SNPs lay in the 3′ UTR of the gene NCU04379 with linkage disequilibrium decaying sharply around this peak (Figure 4); we detected no differential expression of NCU04379 between strains with avid germling communication and those whose germlings communicated poorly (data not shown).
NCU04379 encodes CSE1, a homolog of the vertebrate neuronal calcium sensor-1 (NCS-1) and of Frq1p in Saccharomyces cerevisiae [32]. Deletion of cse-1 in N. crassa results in a mutant that is viable, but sensitive to calcium stress and ultraviolet light, and which shows slightly impaired growth [33]. Similar to NCS-1 and Frq1p, CSE1 harbors a consensus signal for N-terminal myristoylation and four EF-hand domains (PF00036) predicted to be involved in calcium binding [32], [34], [35]. We hypothesized that CSE1 played a role in germling communication and that mutations in this gene would impact cell fusion behavior. Germling CAT fusion experiments validated this prediction, revealing a striking 3.6-fold reduction in the frequency of communication and cell fusion between Δcse-1 germlings relative to communication between germlings of the wild-type, isogenic strain from which the Δcse-1 strain was derived (Figure 5A). The defect was rescued by integration of a wild-type copy of cse-1 at the his-3 locus in the Δcse-1 strain, confirming the specificity of the phenotype to the cse-1 lesion (Figure S3). To evaluate the ability of Δcse-1 germlings to respond to communication with wild-type isolates, we assayed Δcse-1 germlings positioned alongside those of the isogenic fusion-competent strain, and observed a defect similar to that of Δcse-1 germlings interacting with one another (Figure 5A). Thus, CSE1 is essential for chemotropic interactions, including the sensing of and response to the presence of a fusion-competent partner.
We next sought to learn if CSE1 acts before or after a required, chemotropic interaction event in germling fusion, the observable oscillations of MAK2 and SO to the tips of communicating CATs [14]. To address this question, we obtained a wild-type strain expressing either MAK2-GFP or SO-GFP, and we visualized the subcellular localization of the latter proteins during interactions between wild-type germlings and those of the Δcse-1 mutant background. In the few cases in which a Δcse-1 germling showed chemotropic interactions toward a wild-type germling, we observed normal recruitment and oscillation of both MAK2 and SO to wild-type germling tips (every ∼4 minutes) (Figure 6). In the ∼75% of cases in which a Δcse-1 germling and a wild-type germling showed no evidence of chemotropic interactions, MAK2 and SO did not localize or oscillate to CAT tips, but remained in the cytoplasm. We conclude that CSE1 acts upstream of the signaling that underlies chemotropic interactions, because in the rare instances where Δcse-1 germlings commit to chemotropic interactions and cell fusion, they successfully drove MAK2 and SO oscillation.
The mammalian homolog of CSE1, NCS-1, functions during regulated exocytosis in response to calcium signaling [36], [37], and the yeast homolog Frq1p localizes to the Golgi membrane [38]. We reasoned that these attributes would likely be conserved in N. crassa. We first focused on the role of calcium; the Δcse-1 mutant shows growth sensitivity to excess calcium, as well as to calcium depletion [33]. We therefore hypothesized that calcium could be required for chemotropic interactions between N. crassa germlings, and to test this, we assayed fusion of wild-type germlings on growth medium depleted of Ca2+. The results (Figure 7B) bore out our prediction, with no detectable chemotropic interactions or CAT fusion in the absence of Ca2+. We next investigated the localization of CSE1 in N. crassa. For this purpose, we used a Δcse-1 strain in which the cse-1 allele with a C-terminal GFP tag had been integrated at the his-3 locus. The introduction of the GFP-tagged cse-1 allele restored wild-type growth and germling communication phenotype to the Δcse-1 strain (Figure S3). We compared the localization of CSE1-GFP to that of the late Golgi marker VSP52 tagged with RFP [39], [40]. The results, shown in Figure 8, revealed colocalization of the CSE1 and VPS52, with CSE1-GFP also present in the cytoplasm.
Mammalian NCS-1 and S. cerevisiae Frq1p interact with phosphatidylinositol 4-kinase (Pik1p) [32], [37], a protein involved in secretion from the Golgi to the plasma membrane. As Frq1p is required for regulated exocytosis through Pik1p [38], we hypothesized that N. crassa homologs of components of this secretion pathway would play a role in germling communication. To test this hypothesis, we first assayed germlings carrying a deletion of the Pik1p homolog in N. crassa, NCU10397 (pik1), and observed a 1.5-fold reduction of germling communication (Figure 5A). A communication defect of similar magnitude was apparent when Δpik1 mutant germlings were assayed for interactions with wild-type fusion partners (Figure 5A). We next investigated 14-3-3 proteins, regulatory molecules that bind diverse signaling proteins [41] and in S. cerevisiae transport Pik1p from the nucleus to the cytoplasm [42]. Two members of this family have been identified in N. crassa, NCU03300 (nfh-1, encoding the DNA damage checkpoint component RAD24) and NCU02806 (nfh-2, encoding a 14-3-3 protein); we assayed germling communication in strains harboring deletions in each of these genes in turn. The results revealed no effect of the Δnfh-1 mutation (data not shown), but Δnfh-2 germlings communicated with one another at a frequency 1.5-fold less than that of isogenic wild-type germlings (Figure 5A), and Δnfh-2 conidia mixed with those of a wild-type strain exhibited a similar defect (Figure 5A). Echoing our findings from the Δcse-1 mutant, we observed normal oscillation of MAK2-GFP and SO-GFP to the CATs of wild-type germlings when they participated in chemotropic interactions with Δnfh-2 germlings, while wild-type germlings that did not communicate with those of the Δnfh-2 background showed uniquely cytoplasmic localization of MAK2-GFP and SO-GFP (Figure 6). Taken together, these data indicate that CSE1, PIK1, and NFH2 are each required for the calcium-dependent initiation of germling communication and chemotropic interactions, strongly suggesting their joint function in a Golgi secretion pathway involved in signaling to initiate germling fusion.
Given the robust genetic association between cse-1 genotype and germling communcation in wild strains (Figure 3), we reasoned that additional determinants of germling communication could be revealed by mining our genome-wide association data at lower significance levels. For this purpose, we re-examined our association results using a permissive threshold of p<0.015. Permutation testing estimated that 22% of loci reaching this level would be true positives (see methods); as such, independent gene-by-gene validation could uncover bona fide communication genes among this set, potentially both activators and repressors of the communication trait. We focused on genes annotated in secretion, kinase signalling pathways, or peptide hydrolysis in which SNPs showed association reaching our permissive significance cutoff. Of the 18 genes that fit this description and for which deletion strains were available and viable (Table 1), deletion of six genes had significant impact on communication frequencies as compared to a wild-type strain (Figure 5B). The most extreme phenotype, a complete failure of chemotropic interactions and CAT fusion, was observed in the deletion strain for the exocyst complex component sec15 (NCU00117) (Table 1; Figure 5B). The latter mutant also exhibited slower growth, reduced conidiation, and slower conidial germination. Deletion of two additional genes, the protein transporter sec22 (NCU06708) and the acetylornithine-glutamate transacetylase arg-15 (NCU05622) [43], also compromised fusion frequency (68%±2 and 53%±4, respectively) (Figure 5B). Remarkably, deletion of each of three genes heightened germling communication and fusion frequencies (Figure 5B): a GTPase activating protein (NCU06362; 96%±2), the nonidentical kinase-2 nik-2 (NCU01833; 97%±0.7), and the secreted subtilisin-like serine protease spr-7 (NCU07159; 97±1.3). The elevated fusion frequency in each of these strains contrasts with any known germling fusion mutant, all of which reduce or eliminate chemotropic interactions or cell fusion, and highlights the ability of association mapping to pinpoint negative regulators as well as genes with a positive role in cell communication. In each mutant with heightened fusion frequency, germlings were also often involved in fusion events with more than one germling (multiple fusion events) (26.33%±5.24 in ΔNCU06362, 21.33%±1.8 in Δnik-2, and 20.66%±4.07 in Δspr-7; Figure 7C). By contrast, multiple germling fusion events was a phenotype only observed at a low level in a wild-type strain (2%±2).
To investigate further the novel gain-of-fusion phenotype, we focused on the putative secreted serine protease spr-7. We first confirmed that the introduction of an ectopic copy of spr-7 at the his-3 locus restored hyphal communication of the spr-7 deletion strain to wild-type levels, establishing the deletion as the sole cause of the increased communication phenotype (Figure S3). We next asked whether the presence of wild-type germlings would be sufficient to complement the Δspr-7 phenotype during communication. Assays of Δspr-7 germlings mixed with those of a wild-type strain confirmed this hypothesis, revealing a fully wild-type communication phenotype (fusion frequency 82%±3), a striking contrast to the failure to communicate with wild-type germlings we had noted in Δcse-1, Δpik1 and Δnfh-2 mutants (Figure 5 and see above). These results support a model in which secreted SPR-7 from wild-type germlings acts in a cell-non-autonomous fashion to restrict communication and CAT fusion between wild type germlings.
In N. crassa, genetically identical germlings chemotropically sense partner cells and undergo mutual recognition-directed growth and cell fusion [14], [15], [16]. The molecular basis of this phenotype is only partly understood, and tools to identify candidate genes involved in fusion are at a premium in the field. In this work, we genotyped more than 100 wild N. crassa isolates, advanced our understanding of germling communication and fusion, and established this population as a powerful resource for high-resolution association mapping that can be used with any variable phenotype. Our study is the first to illustrate the utility of genome-wide association mapping to identify novel loci underlying trait variation in a microbe. We anticipate that this methodology will be a powerful and generally applicable tool in future genetic study of many eukaryotic microbes, owing to the small genome sizes and deeply-sampled populations of a number of species, particularly filamentous fungi.
The top hit from our association analysis was cse-1, which is homologous to a neuronal calcium sensor gene in animals that shows nervous-system-specific expression and neuron-specific phenotypes; neurons, like hyphae in filamentous fungi, are a highly polarized tissue. Neuronal calcium sensor-1 (Frequenin) is a myristolylated protein with four EF hands that functions as a calcium ion sensor for modulation of syntaptic activity and secretion [34], [44], [45], [46]. Our analysis revealed a near-complete loss of cellular communication during germling fusion in a N. crassa Δcse-1 mutant. In animals and in S. cerevisiae, NCS-1/Frq1p and Bmh1p-Bmh2p regulate phosphatidylinositol 4-kinase/Pik1p, with Bmh1p-Bmh2p mediating the nucleocytoplasmic shuttling of Pik1p [42]. NCS-1/Frq1p promotes association of Pik1p with the Golgi membrane, which is required for its role in regulated exocytosis [37], [38]. Our results established that in N. crassa, CSE1 localized to the Golgi and that deletion of pik1 or nfh-2 phenocopied a cse-1 deletion strain. These observations together support a model in which, in N. crassa, CSE1, PIK1 and NFH2 regulate exocytosis of an unidentified ligand and/or receptor, perhaps initiated via calcium signaling, which is important for establishing communication between cells and subsequent chemotropic interactions (Figure 9). Recently, an essential kinase (MSS-4) involved in the generation of phosphatidylinositol 4,5-bisphosphate (PtdIns(4,5)P(2)) was found to localize to contact points between germlings during cell fusion [47], indicating that the generation of different phosphatidylinositol phosphate moieties may regulate different aspects of germling fusion.
A role for phosphorylation is suggested by our finding that the defect in germling communication observed in the Δcse-1, Δpik1 and Δnfh-2 mutants correlates with an absence of oscillation of MAK2 and SO to CAT tips, because MAK2 kinase activity has been shown to be required for chemotropic interactions and MAK2 and SO oscillation [14]. In S. cerevisiae, Pik1p is required for full activation of the MAP kinases Fus3p and Hog1p and repression of Kss1p [48], and the Fus3p ortholog in N. crassa is MAK2 [10]. It is therefore tempting to speculate that the activation of PIK1 by CSE1 may play an important role in germling communication by affecting activation of MAK2, thus modulating MAK2 phosphorylation targets as well as downstream transcriptional targets required for germling fusion (Figure 9).
In addition to our mapping of cse-1 as a determinant of variation in germling communication across wild N. crassa, further mining of our association results led to the identification and validation of six other genes associated with CAT fusion. Of these, one gene, sec15, is a homolog of a component of the exocyst complex in S. cerevisiae, a multiprotein complex that localizes at the bud tip and is associated with exocytosis [49]. Our results indicated that sec15 is essential for CAT fusion in N. crassa. Likewise, our results revealed a defect in germling communication and fusion frequency in a strain bearing a deletion in a homolog of SEC22 in N. crassa, NCU06708; in S. cerevisiae, Sec22p assembles into a SNARE complex and plays a role in ER-Golgi protein trafficking [50]. Our demonstration that cse-1, pik1, nfh-2, sec15, and sec22 are all required for germling communication establishes the importance of protein secretion and trafficking for chemotropic interactions and cell fusion in N. crassa.
Our results also established that mutation of the acetylornithine-glutamate transacetylase arg-15 [43] confers a defect in germling communication. The homolog of arg-15 in S. cerevisiae, Dug2p, is involved in degradation of the antioxidant glutathione and other peptides containing a gamma-glu-X. dug2 mutants show deficient utilization of glutathione [51], which reacts non-enzymatically with reactive oxygen species and detoxifies oxidatively stressed cells [52]. A role for redox reactions in germling communication through arg-15 would dovetail with reports that mutants in components of the NADPH oxidase complex, which is involved in redox signaling, are defective in CAT fusion [9].
Our work has uncovered a new category of fusion mutants that exhibited germling fusion frequencies higher than those of wild-type, and which displayed multiple fusion events. Of the genes whose deletions gave rise to this striking phenotype, one encoded an uncharacterized predicted GTPase activating protein (GAP) (NCU06362). NCU06362 contains a TBC domain (PF00566) and is a paralog of GYP5 in S. cerevisiae; Gyp5p is involved in the recruitment to sites of polarized growth of the BAR domain protein Rvs167p, which has been implicated in exocytosis at the bud tip [53]. Rvs167p interacts with a second BAR domain protein, Rvs161p, and together this complex plays a role in receptor-mediated endocytosis [54]. Gyp5p also has in vitro GAP activity towards Ypt1p, which is involved in ER-to-Golgi trafficking, and towards Sec4p, which regulates exocytosis [55]. Thus, the increase in germling fusion frequencies observed in the ΔNCU06362 mutant could be due to alterations in secretion or in the reduction of endocytosis of a receptor involved in germling communication.
A second gene whose deletion enhanced hyphal communication, spr-7, encodes a secreted subtilisin-related serine protease, part of a family whose members carry out a wide range of peptidase activities [56]. The increase in fusion frequency and germlings involved in mutiple fusion events in the Δspr-7 mutant suggests that SPR-7 may be responsible for the degradation of a peptide required for extracellular communication (Figure 9). The nature of the extracellular ligand and receptor(s) that guide chemotropic interactions during cell fusion in N. crassa is currently unknown. In fungi, secreted peptides involved in extracellular communication have not been reported, apart from peptide pheromones involved in mating [57], [58] or small secreted proteins with antifungal properties [59], [60]. The genes we have uncovered here will serve as targets for future genetic and biochemical efforts to identify extracellular ligands and receptors involved in germling communication and cell fusion in N. crassa.
Our results also revealed an increase in germling communication in a nik-2 deletion strain. This gene encodes a histidine kinase, a member of a canonical two-component signal transduction pathway and part of an 11-member family in N. crassa. No phenotype for the Δnik-2 mutant has been previously reported [61]. However, other histidine kinases affect MAPK signal transduction pathways in fungi, including nik-1, a member of the osmoregulatory OS-2 pathway in N. crassa [62], and the histidine kinase Sln1p, which regulates the Hog1p MAPK pathway in S. cerevisiae [63]. We hypothesize that the increase in fusion frequencies in the absence of nik-2 may stem from a defect in the regulation of the MAK2 phosphorylation pathway, leading to a hyper-activated state during chemotropic interaction (Figure 9). Further research will be necessary to elucidate the specific role of nik-2 in this process.
By identifying multiple novel determinants of germling communication, our results underscore the power of association studies for the mapping of genes to phenotypes in wild populations. Importantly, our N. crassa population is particularly amenable to GWAS, with little discernable population structure and low linkage disequilibrium, allowing the detection of strong association to finely resolved loci. These attributes of N. crassa stand in contrast to S. cerevisiae, where GWA studies are hampered by a mosaic and heterogenous population structure [64]. Our relatively modest, medium-throughput phenotyping of a quantitative phenotype in wild individuals compares favorably with the high-throughput approach that would be required to survey the >9000 strains of the N. crassa deletion collection [65], not only by saving 98% of the labor, but in enabling analysis of all genes, including those that are essential. However, our molecular follow-up of GWAS hits was aided by the availability of a near-full genome deletion strain collection for N. crassa. When the central question, as in our work, is to infer novel function for poorly annotated genes, comparing a given gene's deletion strain and the isogenic wild-type strain is a straightforward and precise approach that obviates potential complications from epistasis in allele-swapping experiments. Our GWAS method also compares favorably to two-parent crossing schemes for the dissection of natural variation [66]: first, because linkage blocks in our outbreeding population often contain a single gene, whereas more than 50 can be contained in those resulting from just one cross [67], and second, because we sample phenotypes that vary among multiple individuals and not just those that differ between two parents. With the availability of our collection of 112 genotyped individuals to the fungal genetic community, future studies will require only phenotyping to map the molecular basis of trait variation using the strategy we have pioneered here. And as population-genomic resources are developed in many taxa, we anticipate that association mapping will be successfully applied in other species, within and outside the fungal kingdom.
All 112 strains used in this study were isolated from Louisiana, USA (Table S1) and are available from the Fungal Genetics Stock Center (FGSC) [68].
The deletion mutants used in these study were generated by the Neurospora Genome Project [65], [69] and are administered by the FGSC [70]. The rfp-vps-52 transformant was generously provided by Barry Bowman [40]. All strains were grown on Vogel's medium [71] and all crosses were performed on Westergaard's synthetic cross medium [72]. The his-3 A mutant (FGSC# 6103) and a his-3 a strain (FGSC #9716) were used as females in crosses with deletion mutants. Progeny bearing the deletion mutations and the his-3 mutation were isolated and used in complementation experiments.
Total RNA was isolated for each of the 112 strains listed in Table S1. Strains were grown for 16 hrs on cellophane on Bird medium [73]. Mycelia were harvested and immediately added to 1 mL of TRIzol reagent (Invitrogen Life Technologies) [74] and zirconia/silica beads (0.2 g, 0.5-mm diameter; Biospec Products). Cells were disrupted using a MiniBeadBeater instrument (Biospec Products) at maximum speed for 30 seconds twice in succession. Total RNA was extracted according to the manufacturer's protocol for TRIzol (Invitrogen) and quantified on a Bioanalyzer (Agilent).
For polyA RNA purification, 10 µg of total RNA was bound to dynal oligo(dT) magnetic beads (Invitrogen 610.02) two times, using the manufacturer's instructions. Purified polyA RNA was fragmented by metal-ion catalysis [75] using fragmentation reagents from Ambion (AM12450). For first strand cDNA synthesis 1 µg fragmented polyA RNA was incubated with 3 µg random hexamers (Invitrogen 48190-011), and incubated at 65°C for 5 minutes and then transferred to ice. 1st strand buffer (Invitrogen 18064-014) was added to 1× final concentration (4 µL). Dithiothreitol (DTT), dNTPs and RNAseOUT (Invitrogen 10777-019) were added to 100 mM, 10 mM, and 20 U/20 µL respectively, and the sample was incubated at 25°C for 2 minutes. 200 U of Superscript II (Invitrogen 18064-014) were added and the sample was incubated at 25°C for 10 minutes, 42°C for 50 minutes and 70°C for 15 minutes.
For second strand synthesis, 51 µL of H2O, 20 µL of 5× second strand buffer (Invitrogen 10812-014), and dNTPs (10 mM) were added to the first strand cDNA synthesis mix and incubated on ice for 5 minutes. RNaseH (2 U) (Invitrogen 18021-014), DNA pol I (50 U) (Invitrogen 18010-017) were then added and the mixture was incubated at 16°C for 2.5 hours.
End-repair was performed by adding 45 µL of H2O, T4 DNA ligase buffer with 10 mM ATP (NEB B0202S) (10 µL), dNTP mix (10 mM), T4 DNA polymerase (15 U) (NEB M0203L), Klenow DNA polymerase (5 U) (NEB M0210S), and T4 PNK (50 U) (NEB M0201L) to the sample and incubating at 20°C for 30 minutes. A single base was added each to cDNA fragment by adding Klenow buffer (NEB M0212L), dATP (1 mM), and Klenow 3′ to 5′ exo- (15 U) (NEB M0212L). The mixture was then incubated at 37°C in for 30 minutes.
Standard Illumina adapters (FC-102-1003) were ligated to the cDNA fragments using 2× DNA ligase buffer (Enzymatics L603-HC-L), 1 µL of adapters, and DNA ligase (5 U) (Enzymatics L603-HC-L). The sample was incubated at 25°C for 15 minutes. The sample was purified in a 2% low-melting point agarose gel, and a slice of gel containing 200-bp fragments was removed and the DNA purified. The polymerase chain reaction (PCR) was used to enrich the sequencing library. A 10-µL aliquot of purified cDNA library was amplified by PCR. PCR cycling conditions were: a denaturing step at 98°C for 30 seconds, 14 cycles of 98°C for 10 seconds, 65°C for 30 seconds, 68°C for 30 seconds, and a final extension at 68°C for 5 minutes. All libraries were sequenced using an Illumina Genome Analyzer-II using standard Illumina operating procedures. RNAseq data for all strains used in these analyses has been deposited in Gene Expression Omnibus (http://www.ncbi.nlm.nih.gov/geo/; accession no. GSE45406; GSM1103708-GSM1103819).
Mapping of RNA-seq reads to the genome sequence of N. crassa strain FGSC 2489 [76] and calling of single nucleotide polymorphisms (SNPs) was carried out with Maq [77]. All RNA-seq reads that mapped to multiple locations were eliminated from analysis, as were SNPs located in regions of low consensus read quality. These variants were further filtered to retain only those that were bi-allelic, yielding a complete data set of 1.09×106 SNPs (Dataset S1) which were used as input into phylogenetic inference with FastTree; because patterns of inheritance in one strain, JW168, were suggestive of misclassification (data not shown) we did not include this strain in the tree shown in Figure S1. For markers used as input into calculations of genetic association with the germling communication phenotype (see below), we filtered the complete SNP set to retain only sites at which the minor allele was present at >25% frequency (Dataset S2).
For germling communication assays, each strain was grown on Vogel's minimal media [71] in slant tubes for 4–6 days or until significant conidiation occurred. Conidial suspensions were prepared by collecting conidia with wood sticks and suspending in 600 µl of sterile distilled water. The conidial suspension was filtered by pouring over cheesecloth to remove hyphal fragments. Conidia were diluted to a concentration of 3×107 conidia/ml and 300 µl of this final mixture were spread either on an agar or agarose minimal-medium plates. The plates were incubated for 3–4 hours at 30°. At each of 2–3 timepoints for each strain, agar squares of 1 cm were excised and observed with a Zeiss Axioskop 2 using a 403 Plan-Neofluor oil immersion objective. For image acquisition DIC images were taken with a Hamamatsu Orca 03 camera (Hamamatsu, Japan) using the iVision Mac4.5 software and a Zeiss Axioimager microscope. Fusion events were counted for 50 germling pairs in each of 2–3 biological replicates.
Complementation experiments were done using the pMF272 plasmid system [78] to insert a wild type copy of the deleted gene into the intergenic region 3′ of the his-3 locus; transformants were subsequently analyzed for germling fusion frequencies. Wild type copies of genes were amplified using Taq polymerase from New England Biolabs (Ipswich, CA, USA). Primers were designed to amplify the coding regions and also contained an added restriction enzyme site. The amplified DNA fragments were TOPO (Invitrogen) cloned, cut with restriction enzymes and ligated into restriction enzyme-digested pMF272 plasmid. The ligated DNA was used to transform Escherichia coli (DH5a), and the plasmid isolated from individual transformants. The DNA sequence of each plasmid was determined; plasmids containing wild type copies of the genes were used for complementation experiments.
Some mutants showing reduced fusion frequencies were further characterized by studying the ability of the mutant germlings to induce recruitment of MAK2-GFP or SO-GFP to the plasma membrane of opposing germlings as described by Fleißner et al [13]. Conidia from MAK2-GFP and SO-GFP strains were mixed with equal amounts of conidia from the respective deletion mutants and samples were prepared for microscopy as described above. Images were taken at two-minute intervals using a Leica SD6000 microscope with a 100×1.4 NA oil-immersion objective equipped with a Yokogawa CSU-X1 spinning disk head and a 488-nm laser controlled by Metamorph software (Molecular Devices, Sunnyvale, CA).
To visualize CSE1-GFP and RFP-VPS-52 localization, the strains were grown on Vogel's MM plates overnight and squares of 1 cm were excised and examined in the same confocal microscope explained above using the 488-nm laser for GFP and 563 nm laser for RFP. To study co-localization of both proteins, heterokaryons were made by mixing conidia from both strains in the center of a plate and incubating them overnight to allow cell fusion and cytoplasmic mixing from both strains. The samples were prepared and imaged as explained above.
We used germling communication phenotype measurements in biological triplicate from 24 Louisiana strains in a genome-wide association analysis as follows. For each strain, we first calculated the average communication frequency across all replicates and timepoints to yield a final quantitative communication measurement. We then converted the latter value to a qualitative score: we calculated the grand mean and standard deviation of communication frequency across all strains, and we classified a given strain as low-communicating if its communication measurement was more than one standard deviation below the grand mean, and high-communicating otherwise. We then tested each marker in turn, from our set of SNPs with >25% minor allele frequency (see above), for co-inheritance with this qualitative communication score using Fisher's exact test [79]. To evaluate the experiment-wise false discovery rate at a given Fisher's p-value threshold pthresh, we shuffled the vector of phenotype category values among strains, repeated the association test, and tabulated the number of SNPs with Fisher's p-value<pthresh, in this null data set. Averaging over 1000 such permutations yielded an expectation of 0.011 SNPs called at pthresh = 5.6×10−6 and 652 SNPs at pthresh = 0.015, under a null model of no true association. Given the 3 and 837 SNPs, respectively, reaching these levels in the real data (Dataset S3), false discovery rates at these thresholds were 0.4% and 78%, respectively. Linkage disequilibrium in Figure 4 was calculated between all high-frequency SNPs in the region of cse-1 using the LDcorSV package in R.
|
10.1371/journal.ppat.1003847 | KSHV 2.0: A Comprehensive Annotation of the Kaposi's Sarcoma-Associated Herpesvirus Genome Using Next-Generation Sequencing Reveals Novel Genomic and Functional Features | Productive herpesvirus infection requires a profound, time-controlled remodeling of the viral transcriptome and proteome. To gain insights into the genomic architecture and gene expression control in Kaposi's sarcoma-associated herpesvirus (KSHV), we performed a systematic genome-wide survey of viral transcriptional and translational activity throughout the lytic cycle. Using mRNA-sequencing and ribosome profiling, we found that transcripts encoding lytic genes are promptly bound by ribosomes upon lytic reactivation, suggesting their regulation is mainly transcriptional. Our approach also uncovered new genomic features such as ribosome occupancy of viral non-coding RNAs, numerous upstream and small open reading frames (ORFs), and unusual strategies to expand the virus coding repertoire that include alternative splicing, dynamic viral mRNA editing, and the use of alternative translation initiation codons. Furthermore, we provide a refined and expanded annotation of transcription start sites, polyadenylation sites, splice junctions, and initiation/termination codons of known and new viral features in the KSHV genomic space which we have termed KSHV 2.0. Our results represent a comprehensive genome-scale image of gene regulation during lytic KSHV infection that substantially expands our understanding of the genomic architecture and coding capacity of the virus.
| Kaposi's sarcoma-associated herpesvirus (KSHV) is a cancer-causing agent in immunocompromised patients that establishes long-lasting infections in its hosts. Initially described in 1994 and extensively studied ever since, KSHV molecular biology is understood in broad outline, but many detailed questions are still to be resolved. After almost two decades, specific aspects pertaining to the organization of the KSHV genome as well as the fate of the viral transcripts during the productive stages of infection remain unexplored. Here we use a systematic genome-wide approach to investigate changes in gene and protein expression during the productive stage of infection known as the lytic cycle. We found that the viral genome has a large coding capacity, capable of generating at least 45% more products than initially anticipated by bioinformatic analyses alone, and that it uses multiple strategies to expand its coding capacity well beyond what is determined solely by the DNA sequence of its genome. We also provide an expanded and highly detailed annotation of known and new genomic features in KSHV. We have termed this new architectural and functional annotation KSHV 2.0. Our results indicate that viral genomes are more complex than anticipated, and that they are subject to tight mechanisms of regulation to ensure correct gene expression.
| Kaposi's sarcoma-associated herpesvirus (KSHV) is a member of the gamma-herpesvirus family and the etiologic agent of Kaposi's sarcoma, primary effusion lymphoma (PEL), and multicentric Castleman's disease [1], [2]. This human pathogen, initially identified in Kaposi's sarcoma lesions from AIDS patients, has been extensively studied since its discovery and isolation in 1994 [3]. Shortly thereafter, the KSHV genome, a dsDNA molecule of ∼165 kb, was sequenced from the lymphoid cell line BC-1, allowing the in silico annotation of open reading frames (ORFs) that fit the following criteria: (1) they start with a canonical initiator AUG codon and (2) they encode polypeptides larger than 100 amino acids (aa). Many of these ORFs had functional homologues in herpesvirus saimiri (HVS), a gamma-herpesvirus related to KSHV [4]. This study identified a total of 81 such viral ORFs, and except for the more recent addition of microRNAs, non-coding RNAs, and a few small ORFs [5]–[8], the genome map of KSHV has changed little ever since. Gene expression profiling of KSHV transcripts using northern blots, custom oligonucleotide microarrays and real time PCR arrays have demonstrated extensive transcription of the viral genome, hinting at a complex transcriptional profile [6], [9], [10] (unpublished data). More recently, proteomic studies of KSHV-infected cells have assessed the expression of many of the predicted ORFs [11]. However, and in spite of all the aforementioned efforts, a detailed understanding of the genomic architecture, translational state, and biological functions of KSHV gene products remains incomplete.
In an attempt to extend our current knowledge of the coding capacity of the KSHV genome during the productive stage of infection, we employed an unbiased functional genomics approach to study the transcription and translation profiles of lytic KSHV using mRNA-sequencing (mRNA-Seq), ribosome footprinting (Ribo-Seq), and genomic DNA sequencing (DNA-Seq). When combined, these methods provide a comprehensive, high-resolution view of gene regulation and expression dynamics [12]–[14].
By employing these techniques in parallel, we have generated a state-of-the-art annotation of the KSHV genome. Our approach confirms the presence and timing of expression of the majority of previously annotated ORFs, while revealing several novel and, in some cases, unexpected genomic features including ribosome protection of non-coding RNAs, new splice variants, and a plethora of upstream and small ORFs. In addition, we have confirmed and expanded the annotation of transcription start sites, polyadenylation sites, and initiation/termination codons of multiple known ORFs. Our analyses have also uncovered new instances of viral mRNA editing, strongly hinting at a new layer of viral gene regulation during reactivation. The wealth of information generated by integrating the data obtained from our combined methods has expanded our understanding of the viral genome architecture and dynamics, revealing a surprising coding capacity of KSHV that goes well beyond what was initially described based on its genome sequence alone.
The life cycle of KSHV can be separated in two very distinct stages: the dormant state known as latency and the productive state referred to as the lytic cycle [2]. While viral gene expression in latency is limited and most of the genome is silent, the lytic cycle is a transcriptionally dynamic state where the timing of gene expression is tightly regulated to ensure the ordered synthesis of viral products [6], [15], [16]. We sought to study the kinetics of latent and lytic viral transcription in detail, as well as the translational fate of newly synthesized mRNAs. To this end, we employed a system developed by our group that allows the study of the KSHV lytic cycle in a tightly-controlled manner [17]. This system comprises the epithelial iSLK-219 cell line, which is latently infected with a heterologous KSHV strain (see below) and harbors a doxycycline (Dox) inducible transgene encoding the viral transcription factor RTA (replication and transcriptional activator). The exogenous expression of RTA by Dox treatment in iSLK-219 cells is sufficient to induce the lytic reactivation of latent KSHV. Notably, latency in iSLK-219 cells is very strict with less than 0.1% of the cells showing lytic markers in the absence of induction [17]. This is the principal experimental advantage of SLK cells, allowing the study of KSHV latency in the near total absence of contaminating lytically infected cells. The viral strain in iSLK-219 is the recombinant KSHV.219, which encodes a constitutive GFP reporter as well as an RTA-inducible RFP reporter in the viral genome, thereby facilitating the monitoring of viral reactivation [18].
To finely resolve the transcriptional profile and the ribosome occupancy of viral mRNAs, we induced iSLK-219 cells with Dox for 0, 8, 24, 48 and 72 hr (Figure 1A). We evaluated KSHV lytic reactivation by epifluorescence microscopy analysis of GFP and RFP expression, as well as by immunodetection of viral products and quantification of viral DNA replication (Figure 1B, Figure S1). In iSLK-219, lytic DNA replication, the traditional border between early and late times, commences at ∼48 hr post induction (hpi) (Figure S1B). The selected time points represent the different stages of the lytic cycle, known as latent (0 hr), immediate early- (8 hr), delayed early- (24 hr), and late-lytic (48 and 72 hr) (Figure 1A). At each time point, we recovered polyadenylated RNA (mRNA) and 3 sets of ribosome footprints (described in Materials and Methods). To map actively elongating ribosomes on viral transcripts, we isolated ribosome footprints from cells treated with cycloheximide (CHX), a translation inhibitor that binds the ribosomal-E-site and arrests elongating ribosomes [19]. In the same manner, we mapped initiating ribosomes by treating cells with harringtonine (Harr), a translational inhibitor that binds the 60S subunit and hinders the progression of the initiating ribosome, causing ribosomes to stall at translation start sites [20]. Finally we mapped releasing ribosomes accumulating at the stop codon in samples that were not treated with any drug as previously described [14]. We then constructed Illumina-compatible libraries from fragmented and size-selected mRNA segments (40–100 nt), or ribosome protected RNA (ribosome footprints ∼30 nt in length, Figure S2A) following the standard ribosome profiling protocol previously described [14], [21]. The libraries were deep-sequenced and the resulting reads aligned to the KSHV genome (GQ994935). As expected, the number of reads aligning to the KSHV genome increased as the lytic cycle progressed (Figure S2B). To annotate viral splice junctions, we used two splice junction mapping tools; TopHat and HMMSplicer [22], [23]. With these tools, we detected the majority of the known splice junctions and discovered 7 new events including one at the 3′ end of ORF57. Lastly, we annotated putative ORFs by training a support vector machine (SVM) to identify translation initiation sites throughout the KSHV genome based on characteristic peaks within the harringtonine Ribo-Seq data. The list of ORFs produced by extending each of the putative initiation sites to the next in-frame stop codon (taking into account any intervening splice junctions) was then finalized through manual curation.
Figure 1C shows an example of the data obtained with our combined approach. In this case the read coverage from mRNA-Seq and Ribo-Seq (CHX and Harr) libraries for the late-lytic viral gene K8.1, one of the best-mapped genes in KSHV. The data clearly illustrates the single-nucleotide resolution and high-coverage of our methods which here allow the delineation of transcription start and end sites, splice junctions, and coding region boundaries. Notably, the coordinates derived from our combined approach correspond precisely to those previously reported for this gene [24], [25], providing strong validation of our methodology. Together, the data obtained using our multipronged approach generated a high-resolution map of the viral genome architecture.
mRNA-Seq and Ribo-Seq allowed us to perform an unabridged temporal analysis of viral gene expression coupled to a blueprint of the viral episome architecture, granting the opportunity to develop a revised version of the KSHV genome annotation, which we have designated KSHV 2.0 (Figure 2 and Tables 1, 2 and 3). In KSHV 2.0 we annotate the coordinates for 49 viral transcripts and 70 ORFs, as well as those for non-coding RNAs, polyadenylation signals, and splice junctions. In addition, KSHV 2.0 incorporates information pertaining to the timing of expression of the aforementioned elements. Remarkable novel features of KSHV 2.0 include a set of 50 ribosome-loaded segments not previously annotated as bona fide ORFs because of their small size (3–100 aa) and the use of non-canonical start sites (Figure 2B). Together with novel peptide isoforms, and splice variants, these short and upstream ORFs (sORFs and uORFs, respectively) increase the coding repertoire of KSHV by more than 45% and add a new level of potential gene regulation to an already complex landscape. The novel features annotated in this study are summarized in file S1 and can be visualized using the mochiview database in file S2 [26]. In spite of the comprehensive annotation generated for KSHV 2.0, some known features of the viral genome were not detected or could not be rigorously assigned (Figure 2B), due to ambiguities generated by regions of low sequencing coverage, overlapping transcription and translation, and cell line specific patterns of gene expression.
Inspection of KSHV 2.0 reveals three prominent features: (1) coding and non-coding elements are densely packed in the episome, (2) multiple strategies are used to increase its polypeptide repertoire, including splicing, mRNA editing, and alternative start codon use, and (3) sORFs and uORFs populate many regions of the viral genome. The specific transcriptional and translational features of KSHV 2.0 are discussed in detail below.
The transcriptional capacity of the KSHV genome has been traditionally studied using northern blotting and gene expression profiling with oligonucleotide microarrays [6], [9], [27], [28]. While these studies have exposed many features of the viral transcriptome, the limitations of these methods prevent the fine mapping of transcripts, which can require single-nucleotide resolution. For this reason, we performed mRNA-Seq in cells lytically infected with KSHV, to explore the transcriptional landscape of the KSHV genome, resolve the boundaries of viral messages, and uncover novel cis-regulatory elements including transcription start sites (TSS), polyadenylation signals (PAS), and splice junctions.
Taking advantage of the peaks visible at the 5′ ends of transcripts in mRNA profiles (Figure 1C), which are a natural consequence of visualizing the 5′ ends of fragments produced via random fragmentation of multiple mRNA copies of any given transcript, we mapped 49 TSS upstream of 54 out of the 85 officially annotated genes (See Materials and Methods, Tables 1 and 2). The annotation of the TSS coordinates for the remaining 31 viral genes was impeded by low coverage or the presence of overlapping transcripts. Among the mapped viral genes, the discrepancy between the number of TSS and genes, stems from the existence of bi- and poly-cistronic mRNAs (Figure 3A). Of the 49 TSS mapped, 28 are novel while 21 correspond to annotated transcripts whose TSS were previously characterized. Of the 21 previously documented TSS described in our study, 13 are mapped exactly as in the literature and 8 are located within 50 nucleotides of their reported coordinates as previously resolved by 5′ rapid amplification of cDNA ends (5′RACE) (Tables 1 and 2). Interestingly, sequence alignment of the promoter regions corresponding to the TSS unveiled in our analyses, shows the presence of a TATA-box 30 nucleotides upstream of 41 TSS, which remarkably corresponds to the same location of this cis-regulatory element in human promoters. The remaining eight TSS are TATA-less (Table S1). Our observations clearly reflect the strict evolutionary dependence of the pathogen on the host's transcriptional machinery (Figure 3B) [29]–[31].
Like cellular messages, KSHV transcripts are protected by a 5′ 7-methylguanosine cap and a 3′ poly-adenylate (polyA) tail [32], [33]. To map functional polyadenylation signals (PAS) in the viral genome, we selected the RNA-Seq reads that contained a stretch of 5 or more adenosines at their 3′ end and, after trimming this poly-A sequence, aligned the reads to the KSHV genome. The 3′ positions of aligned reads were then marked as polyadenylation cleavage sites, except when the genome contained a complementary poly-T stretch at the same location as the poly-A stretch. Using this approach, we mapped 94 putative cleavage sites, corresponding to 42 transcripts and 74 genes (Tables 1, 2 and 3, Table S2). Our data, and recent studies mapping the 3′UTRs of KSHV genes [34], [35], support the existence of bi- and poly-cistronic messages, as well as transcript clusters with distinct transcription start sites (TSS) that end in a common PAS, suggesting the existence of alternative nested promoters upstream of such PAS's (Figure 3A, Table S2).
Interestingly, sequence analysis of a 60-nucleotide window centered on the predicted cleavage site for the polyadenylation machinery confirmed the presence of the canonical AAUAAA motif in 83% of mRNAs (35 out of 42), the alternative AUUAAA motif in 17% (7 out of 42), and the accompanying downstream GU rich element in all of these transcripts (Figure 3C and Table S2). Similar observations were recently reported in genome-wide analyses of polyadenylation sites in PEL cells infected with KSHV [36]. The presence of these conserved elements highlights once more the strict dependence of the virus on host factors that control the RNA processing of Pol II transcripts [37].
Besides the densely packed coding regions and regulatory features we annotated, our mRNA-Seq data also show the massive accumulation of sequence reads that map outside of previously annotated coding regions, thus indicating highly permissive transcription of most of the viral genome late in the lytic cycle (Figure S3). Two of these regions correspond to two long transcripts recently discovered by our group, the 10 kb antisense RNA to the latent transcripts (ALT) and the 17 kb K1-ORF11 antisense (K1/11-AS) [6]. Intriguingly, these long RNAs show short regions modestly populated by ribosomes, suggesting they may have a coding potential (Figure S4A and S4B).
It is noteworthy that this observation was not restricted to ALT and K1/11-AS. Surprisingly, our Ribo-Seq data also revealed the presence of ribosomes on the “non-coding” RNA PAN (polyadenylated nuclear RNA). PAN is the most abundant viral transcript during the lytic cycle and is required for viral gene expression and virion production [33], [38], [39]. Interestingly, and in spite of PAN's reported nuclear localization, we observed initiating ribosomes accumulating at the start codon in the harringtonine treated samples, elongating ribosomes throughout the body of the transcript in the CHX treated samples, and an accumulation of releasing ribosomes at the stop codon in the samples not treated with any translation inhibitor, starting at 8 hr following reactivation and throughout the lytic cycle (Figure 4A, S5A and S5B). Taking in consideration this pattern of ribosome protection, classical of coding regions [14], we identified three predominant sORFs hosted within the PAN transcript: PAN1.1 (37 aa, 28655), PAN1.2 (44 aa, 28831) and PAN1.3 (25 aa, 28888) (Figure 4B). Besides these, we also identified 3 minor sORFs at the 3′ end of PAN, with very low ribosome occupancy (data not shown). To evaluate the coding capacity of PAN, we calculated the ribosome release score (RSS) for the three main putative ORFs, PAN1.1, 1.2 and 1.3. The RRS is a metric that takes in consideration that ribosome protection within a coding region ends after the stop codon and that no ribosomes should be present at the 3′UTR of the transcript following an ORF. A recent report by Guttman et al. indicates that the RSS provides an indirect measure of translation that allows the reliable differentiation between coding and non-coding transcripts [40]. The RSS calculated for PAN1.1, 1.2 and 1.3 are comparable to those of known coding RNAs and are similar to scores previously determined for small ORFs within mammalian transcripts [40], further supporting the translation potential of PAN (Figure S5D, E). Notwithstanding the low translation efficiency of the three major PAN sORFs (0.05 at 8 hpi to 0.2 at 72 hpi), our mRNA-Seq and Ribo-Seq data suggest that, owing to the significant accumulation of the PAN transcript during the lytic cycle, the putative peptides encoded in these sORFs could be quite abundant. In fact, PAN RNA represented up to 92% of the total viral mRNA-Seq reads and the ribosome-protected RNA corresponding to the small PAN peptides represented up to 1.7% of the total cycloheximide Ribo-Seq reads (Table S3). Thus, our data strongly indicate that this transcript is available for ribosome binding, and that in addition to its documented functions as a “non-coding” RNA, PAN may also be a presumptive coding RNA. It is important to note that the putative coding regions for PAN are overlapping with the ORFK7 transcript. However, close inspection of the ribosome accumulation at the start codon of ORFK7 indicates that in order for the PAN peptides to be encoded by the K7 transcript, the translation efficiency of the internal peptides would need to be 1000 to 10000 times more efficient than that of the main ORF K7 (Figure S5C). Based on this observation and the vast number of reads seen for the mRNA and ribosome protected fragments in PAN1.1, 1.2 and 1.3, we conclude that these putative coding regions are within the PAN transcript.
We hypothesize that despite their minuscule size, the putative peptides encoded by PAN may be functional. Indeed, the number and characterized functions of such small peptides are continuously increasing, and there is overwhelming evidence in other viral systems, as well as in eukaryotic cells, for the abundance and relevance of small peptides [14], [21], [41]. Encouraged by these findings, we started to look for a possible function of the predicted peptides encoded by PAN1.1, 1.2 and 1.3. To such end, we used bioinformatics tools that included diverse motif finding and peptide-function prediction engines [42], [43]. Surprisingly, multiple independent analyses predicted a putative signal peptide in PAN1.1 (Figure 4B), thus suggesting that this peptide may traverse the secretory apparatus. Importantly, the sequence of PAN1.1 and the other small peptides predicted within PAN show 100% conservation at the nucleotide level between different isolates of KSHV (data not shown). Further studies of these putative gene products are underway.
Three additional regions show extensive mRNA-Seq coverage, particularly at 72 hr post reactivation. These are the antisense transcripts corresponding to ORFK5/K6, as well as the antisense transcripts for ORFK9-ORF62 and ORFK2 (vIL6)-ORFK4.2 (Figure 4C, data not shown). We confirmed the existence of a 6 kb RNA antisense to ORFK5/K6 by northern blotting (Figure 4D). This antisense RNA, which we have denominated K5/K6-AS, corresponds to the T6.1 RNA previously described by Taylor et al. [28]. Strikingly, K5/K6-AS is devoid of any initiating or elongating ribosomes, and therefore may represent a bona-fide long non-coding RNA in KSHV that is inherently distinct from PAN, ALT and K1-11AS (Figure 4).
Like other herpesviruses, KSHV makes widespread use of the cellular mRNA splicing machinery [14], [44]. To confirm known splice junctions and discover new ones, we annotated the possible viral splice junctions in a genome-wide fashion by employing HMMSplicer and TopHat on the mRNA sequences that did not align to either the viral or the human genome [22], [23]. Our results confirmed the presence of 27 splice junctions, corresponding to one or more introns in 17 viral genes (20% of genes). These included the well-characterized splice variants observed in ORF50 and ORF57, as well as the multiple splice variants of K8 and K8.1 (Figure 1 and Table S4). The coordinates of the splice junctions annotated using our experimental data confirm those from previous reports (Table S4), affirming the reliability of our combined methods.
Notably, our data not only correctly annotated known splice junctions but revealed 7 novel ones, thus increasing the number of splicing events from 20 to 27 (Table S4). One such splice junction is located at the 3′ end of the ORF57 transcript. ORF57 is a well-characterized KSHV protein thought to be an activator of mRNA maturation and transport, enhancing viral gene expression [45], [46]. Our data support the existence of canonical splice donor and acceptor sites in the new predicted junction (Figure S6A), which give rise to a novel splice variant of ORF57 in which the truncation of a second exon results in the accumulation of ribosomes on a previously uncharacterized third exon (Figure 5A,B). We confirmed the second splicing event of the ORF57 transcript by end-point PCR in iSLK (iSLK-219) and lymphatic endothelial cells (LEC-219) infected with recombinant KSHV.219, but not in infected B cells (BCBL-1) (Figure 5C and S6B).
The second splicing event in the ORF57 transcript results in the removal of a 571 nucleotide fragment encoding amino acids 266–455 within the second exon and leads to the loss of the canonical UAA stop codon of ORF57, resulting in the generation of a novel isoform, here named ORF57A, with a different C-terminus that contains 33 amino acids (Figure 5B). Interestingly, the stop codon of ORF57A (position 83464) is located downstream of the canonical polyA cleavage site in ORF57 (position 83453), suggesting the presence of a transcript with an extended 3′ end. We confirmed the existence of such longer mRNAs in iSLK-219 and BCBL-1 cells by PCR using a primer set annealing within the ORF57 and ORF57A coding regions, and downstream of the annotated polyA cleavage site (Figure S6C). ORF57A, the new alternative splice variant of ORF57, is 299 amino acids in length and lacks the C-terminal leucine zipper (aa 343–364), the second arginine-glycine-glycine rich domain (RGG) (aa 372–374), the zinc finger domain (aa 423–432) and the glycine-leucine-phenylalanine-phenylalanine (GLFF) domain (aa 447–450) (Figure 5A). While the expression of a truncated form of ORF57 could have functional implications, it is important to note that the ORF57A splice variant was detected only in cells infected with the recombinant KSHV.219 virus (iSLK-219 and LEC-219), and could reflect a secondary effect of the insertion of the GFP/RFP reporter cassette downstream of ORF57 [18], leading to the activation of this cryptic splice site within the ORF57 transcript. This observation demonstrates that elements inserted within the viral genome, even in regions that are seemingly devoid of regulatory/functional elements, may not be inert and could have repercussions on viral gene expression and/or function.
A second posttranscriptional mechanism employed by KSHV to expand its coding capacity is mRNA editing. The post-transcriptional recoding of RNA results in single nucleotide discrepancies between the genomic and transcript sequences [47], [48]. By comparing our mRNA-Seq and DNA-Seq data sets, we found 6 instances of mRNA editing in KSHV in two or more time points (Table S5). Three of such editing events include the previously reported A-to-G transition in genomic position 117,809 within the transcript encoding Kaposin [49] and two novel G-to-A transitions at genomic positions 72,795 and 74,281 of the mRNA encoding RTA, leading to amino acid substitutions in the corresponding encoded polypeptides (Table S5). We confirmed these editing events by end-point PCR amplification of cDNA followed by Sanger sequencing (Figure 5D, E, data not shown). Interestingly, we note that the Kaposin message is edited starting at 24 hr following reactivation, and that the relative amount of edited transcript increases dramatically as the lytic cycle progresses, leading to a highly penetrant A-to-G transition at 72 hr (Figure 5E, Figure S7A and B). The surge in the levels of Kaposin mRNA editing is concomitant with the up-regulation of all isoforms of the adenosine deaminase acting on RNA (ADAR), the enzyme implicated in A-to-I editing (Figure S7C). Our observations indicate that ADAR is at least partially insensitive to the generalized host shutoff mediated by the viral endonuclease SOX (Figure S7D) [50]. The consequence(s) of higher levels of ADAR on host mRNA and other viral transcripts, if any, remain to be determined. In addition to investigating the mechanisms of regulation and the activity of ADAR during lytic infection, it would be of great interest to ascertain the biological impact of the A638T substitution in RTA. The post-translational modification prediction tool NetPhosK [51] suggests that such a mutation improves the sequence context for S634 and S636 phosphorylation in the C-terminus of the protein [52]. The biochemical and functional consequences of this mRNA editing in RTA are yet to be determined and will be the focus of future studies. Two of the predicted events, a U-to-G transversion in position 6144 within ORF6 and a G-to-U transversion in position 96434 within ORF59, were not affirmed by Sanger sequencing, highlighting the importance of validation of putative mRNA editing sites identified through next-generation sequencing (data not shown).
A third mechanism used by KSHV to increase the coding capacity of its viral genome is independent of transcriptional control and involves the manipulation of translation. Ribo-Seq allowed us to accurately map most of the annotated viral ORFs while affording the opportunity to discover several dozen undocumented peptides and putative protein isoforms (Figure 2B). Our data show five coding regions that are of particular interest in that regard. ORF70, ORF K6, ORF54, ORF62 and Kaposin exhibit a remarkable accumulation of initiating ribosomes on multiple in-frame translation start sites, strongly arguing in favor of the presence of at least two protein variants for each one of these ORFs (Figure 5F, Tables 1 and 2, Table S6, Figure S8, File S1 and data not shown). We examined the expression of one of these proteins, ORF54 by immunoblotting in lysates from latent and lytic iSLK cells infected with Wt KSHV or an ORF54 deleted virus [53]. In perfect agreement with our Ribo-Seq data, we detected two isoforms of ORF54 using an antibody directed against the C-terminus of the protein. These migrate at ∼ kDa (318 aa -ORF54) and 32 kDa (291 aa-ORF54A) in denaturing SDS-PAGE gels, consistent with our finding of 2 polypeptides that share a common C-terminal domain but possess distinct amino-termini owing to the usage of alternative translation initiation sites. Furthermore, our sequencing data indicate that the previously uncharacterized short form of ORF54 is the most abundant one (Figure 5F), which is also in exact agreement with our immunoblot analysis (Figure 5G). The ORF54 and ORF54A products detected during lytic infection of iSLK cells are also clearly seen in HEK293 cells transfected with C-terminally tagged versions of the gene (data not shown). Taken together, our data affirm that KSHV can selectively use alternative start codons to amplify the peptide repertoire synthesized during the lytic cycle.
The peptide coding capacity of KSHV has been defined globally employing in silico approaches and proteomics studies [4], [11] and at a single-gene level by mutagenesis, epitope-tagging and immunodetection. We sought to obtain a unifying and comprehensive understanding of the viral peptide coding capacity using ribosome profiling. Using Ribo-Seq, we mapped most of the previously annotated viral ORFs with precision and, remarkably, discovered 63 new ORFs, representing a higher than 45% increase of the annotated coding capacity of KSHV to date. The vast majority of these new ORFs encode peptides smaller than 100 amino acids and, in 44% of the cases, peptides that are translated from initiation codons with consensus or near-consensus Kozak sequences [54]. Thus, we have reclassified the coding regions of KSHV into primary ORFs, alternative splice variants, internal ORFs, ORFs with alternative start codons, small (sORFs) and upstream ORFS (uORFs) (Figure 2, Tables 1 and 2, Figure S9B–C, File S1).
We defined sORFs as all of those regions encoding peptides of ∼100 aa or less that are not found at the 5′ of an annotated viral gene [55]. In total we found 14 sORFs within 6 transcripts (Table S7). Among these, we clearly detected ribosomes populating the 5′ end of the ORF50-antisense (50-AS) transcript at 24–48 hr post reactivation (Figure S9B), confirming recent reports that indicate that this mRNA is indeed present in polysomal fractions [7]. While previous transfection-based studies from our lab and others have characterized peptides ranging from 17 to 48 amino acids starting from multiple AUG initiation codons, our Ribo-Seq data indicate that the accumulation of initiating ribosomes in an authentic viral infection involves at least three non-canonical start codons giving rise to small peptides from 8 to 76 aa [7], [8].
The second class of small coding regions revealed by our Ribo-Seq data consists of a group of 36 upstream ORFs (uORFs). These uORFs are present in the leader sequence of annotated ORFs and encode peptides of ∼100 aa or less [56]. We noted that uORFs are very numerous and widely distributed across the whole genome (Table S7). In total, 24 genes have between 1 and 6 uORFs that are either in-frame or out-of-frame with the main ORF (Figure S9C). Interestingly, and as has been previously reported for HCMV and mammalian cells [14], [21], 44% of uORFs are translated from a non-canonical start codon and are highly detected at late times during reactivation (Figure S9A, Table S7).
An example of the regulatory capacity of uORFs in KSHV was recently documented by Kronstad and colleagues, who described the functions of two uORFs identified in our Ribo-Seq data as uORF35.1 and uORF35.2 (Figure 6A) [57]. uORF35.1 and uORF35.2 have opposing regulatory functions on the translation of the downstream ORFs ORF35 and ORF36, in a mechanism akin to that described for eukaryotic uORFs regulating cell-stress response genes [58]. These uORFs are located in the 5′ leader sequence of the ORF35–36 bicistronic transcript. The uORF35.1 small peptide (8 aa) is in-frame with respect to ORF35, while the uORF35.2 small peptide (10 aa) is an out-of-frame overlapping ORF with respect to ORF35 (Figure 6B). Both of these uORFs inhibit the expression of ORF35, as their deletion promotes accumulation of this protein. However, uORF35.2 has stimulatory effects on the translation of the most 3′ gene, ORF36, via a continuous scanning mechanism [57]. These data affirm the existence and functional significance of two of the uORFs identified by Ribo-Seq, and support the reliability of this method for identifying such elements.
The phase switch from latency to the lytic cycle is a highly regulated process that requires the temporally controlled expression of genes. Our strategy of mapping transcripts and coding sequences across different stages of the lytic cycle in an RTA-regulatable expression system revealed a built-in timer for viral reactivation that relies on the use of specific TSS during the distinct stages of the viral life cycle. Fine temporal mapping of viral transcription include 4 latent messages, 13 early genes that are expressed starting at 8 hr, 19 genes expressed between 24 hr and 48 hr after reactivation and 38 genes at 48–72 hr following DNA replication (Tables 1, 2 and 3, Figure 7A, Figure 8A and 8B). We took advantage of the restricted latency and protracted lytic cycle observed in iSLK-219 cells when compared to cells of lymphoid origin (unpublished observations) to study the kinetics of viral transcription in much finer detail. As expected, our results show that only a handful of transcripts are expressed during latency, namely the K1-ORF4 bicistronic message, vIL6, Kaposin and the LANA-vCyclin-vFLIP tricistronic transcript. Interestingly, the ribosome profiling of latent cells shows that only vIL6 and LANA are protected by ribosomes (Figure 7A and 7B). Furthermore, we confirmed by immunoblot the presence of LANA and the absence of Kaposin and vCyclin proteins in latently infected cells (Figure 7D and 7E). These observations raised questions about whether the K1, vCyclin and vFLIP proteins might be importantly regulated at the level of translation. Consistent with this, the Kaposin, K1, vCyclin and vFLIP transcripts are abundantly protected by ribosomes upon induction of the lytic cycle, and their cognate proteins can be detected by immunoblotting after such induction (Figure 7C, 7D and 7E). However, subsequent Northern blot analysis revealed the pattern of accumulation of vCyclin and vFLIP transcripts in iSLK differs from that previously observed in B cells (Figure 7F and 7G) [59]. To our surprise, we could detect only the tricistronic (LANA-vCyclin-vFLIP), but not the bicistronic (vCyclin-vFLIP) message, in latent iSLK-219 cells (Figure 7F). However, the bicistronic transcript, which has previously been proposed to be the mRNA for these 2 proteins, is abundantly expressed in lytic iSLK-219 cells (Figure 7F) [60]. This observation, in combination with our mRNA-seq and Ribo-seq data, suggests that vCyclin and vFLIP proteins are indeed mainly expressed from the bicistronic message and that their expression is primarily regulated at the RNA level during latency in iSLK-219 cells. It remains possible that translational control governs the latent expression of ORF K1.
As to the lytic cycle, our data support that most gene expression during this phase is controlled through transcriptional regulation, as the vast majority of the newly synthesized mRNAs are protected by ribosomes without delay (Figure 8A and 8B). A notable example of the temporal selection of distinct TSS can be seen in the transcription of the ORF58–62 locus which encodes the EBV-BMRF2 homologue (ORF58), the DNA polymerase processivity factor (ORF59), the small (ORF60) and large (ORF61) subunits of the viral ribonucleotide reductase, and a small capsid protein (ORF62) (Figure 8C) [61], [62]. In this case, our data clearly distinguish at least three independent transcripts with different expression kinetics: the RNA of ORF58–59 is expressed first, followed by the delayed early ORF60–61, and the late ORF62 transcript. The differential expression of these mRNAs correlates with their biological function, as genes required for DNA replication (ORF59, 60 and 61) are expressed before structural proteins (ORF62). Through the detection of these three distinct cistrons, our results unequivocally support the existence of three independent promoters that are integral part of the aforementioned temporally-regulated gene activation. Indeed, our TATA-box analyses, as well as previous reports, have identified at least three temporally-regulated promoters in this region (Table S1), supporting the differential expression control of ORF58–62. The time regulated selection of TSS has also been reported in HCMV, where the expression of transcripts from alternative TSS results in the translation of different protein products at specific times during infection, and represents a conserved mechanism of gene regulation in herpesviruses [14]. Taken together, our results, alongside previous studies, suggest that the transcription of KSHV genes is tightly regulated by promoter availability and the dynamic interplay of host and viral transcription factors in a time-dependent phase switch operating in the transition from latency into the lytic cycle [63], [64].
The wealth of information we generated by combining our DNA-Seq, mRNA-Seq and Ribo-Seq datasets allowed us to build a comprehensive, high-resolution map of the viral genome over the KSHV life cycle. Our approach showcases the great analytical power of next-generation sequencing technologies, as we were able to pan-genomically map coding, non-coding, and regulatory features of the KSHV episome in iSLK-219 cells. Most importantly, this study provides tangible evidence derived from experimental data, as opposed to in silico prediction approaches, of the mechanisms employed by herpesviruses to widen the coding capacity of their genome through the use of diverse strategies including splicing, mRNA recoding, and alternative start codon usage. Furthermore, we demonstrate that the viral genome is not a conventional source of coding sequences as traditionally defined by in silico predictions and homology analyses, but a rather rich collection of diverse coding sequences that include numerous viral sORFs and uORFs. Some of those may exert translational modulation of other viral ORFs, thereby enriching and adding complexity to the viral gene regulatory profile. It is noteworthy that several dozen viral features uncovered by our group in this study were not documented earlier because of the limitations imposed by conventional methods employed to study gene expression.
Our results confirm the striking conservation of eukaryotic cis-regulatory elements in the KSHV genome, including TATA boxes, polyadenylation signals, and splice junctions. The conservation of both, the sequence and position of these features reflects the strict dependence of the virus on host factors and importantly, imposes a major constraint for the pathogen because of the need to compete with host factors for transcription and translation of virally-encoded products. To circumvent this competition, the virus causes the gradual but massive inhibition of host protein expression by increasing mRNA turnover. This mechanism, known as host shutoff, is orchestrated by the concerted action of the viral endonuclease SOX and the cellular exonuclease Xrn1 [65]. The elimination of competing cellular mRNAs renders the translational machinery components available to the virus, relieving the restrictions with respect to the expression of viral gene products. In addition to this upsurge in the availability of translation factors caused by the decreasing amounts of host transcripts, KSHV also hijacks essential cellular pathways that directly impinge on translation. Notable examples are the virus-orchestrated activation of mTOR and MAP kinases, which promote translation during the lytic cycle and result in the expression of viral ORFs required for the progression of infection [66]–[68]. It thus follows that the tug of war between host and virus would result in the commandeering of essential cellular factors by the virus to promote its own replication. Based on our analyses, we posit that the increased accessibility to ribosomes, the creation of an environment that is conducive to high translation, and the accumulation of viral transcripts results in an extremely favorable scenario for pervasive translation of viral encoded ORFs herein annotated, including sORFs and uORFs, and the short ORFs in PAN. Future studies in cells infected with viruses defective for host shutoff will help clarify the contribution of this mechanism to the viral translational output.
Another striking yet puzzling finding derived from our data is the protection of PAN by initiating and elongating ribosomes. This well-characterized and abundant viral RNA is expressed during the lytic cycle and has a predominant nuclear localization [33], [38]. Surprisingly, this nuclear RNA is clearly bound by ribosomes, indicating that either a fraction of PAN is cytoplasmic and available for translation, or (less likely) that ribosomes can access PAN in the nuclear compartment [69], [70]. Even more remarkable is the fact that the relative abundance of ribosome-bound PAN fragments suggests that the encoded peptides, if stable, could be abundant. Irrespective of their abundance, the biological functions of such putative peptides remain to be investigated. Bioinformatics analyses failed to identify discernible domains or particular motifs in 2 out of the 3 PAN-encoded peptides identified in our study. We did however find recognizable motifs within the primary structure of PAN1.1. This peptide harbors a putative signal sequence but lacks a discernible transmembrane domain or an ER retention signal, thus suggesting that PAN1.1 may be a secreted product. Studies aimed at the identification and characterization of PAN1.1 products are currently ongoing.
An interesting observation from our analyses is the detection of mRNA editing in two viral transcripts. This posttranscriptional mechanism of coding and non-coding RNA editing is conserved in eukaryotes. In humans extensive RNA editing has been reported, the majority of the events corresponding to A-to-I transitions mediated by the family of deaminases ADAR. Indeed, alterations in the activity of these enzymes are associated with disease [48], [71]. RNA editing is not restricted to eukaryotic messages and has been observed on viral transcripts in cells infected with RNA and DNA viruses [72]. The effects of RNA editing of viral transcripts antagonize (HCV, MV), or promote (HIV, HDV) viral activity and may affect the function of particular non-coding RNAs (EBV) or viral proteins (KSHV). As in humans, most of the RNA editing events detected on viral transcripts are ADAR dependent and correspond to A-to-I transitions, as the one we observe in Kaposin (position 117809). Interestingly, we see the protein levels of ADAR1 increasing throughout the viral lytic cycle, indicating the message for this enzyme partially escapes the widespread host shutoff caused by SOX. It is not clear however what would be the effects of such up-regulation of ADAR on RNA editing of host messages, given that a large number of transcripts are degraded during KSHV lytic infection. In addition to the nucleotide change in Kaposin, we also detect the non-canonical G-to-A editing in RTA (positions 72795 and 74801). While rare and less frequent than the A-to-I change, the G-to-A recoding has been previously identified in HIV and prostate and colorectal cancer [73], [74]. The enzyme responsible for this editing event has not been characterized to date, and KSHV lytic infection may provide a useful system for its identification.
Remarkably, our results strongly suggest that the peptide coding capacity of KSHV has been previously underestimated. This is illustrated in the overall high representation of sORFs and uORFs in the viral genome, often translated from near canonical start codons. The functional implications of such translation initiation events have been documented in several eukaryotes (yeast and metazoans) where translation initiation from near canonical codons under physiological conditions occurs more often than anticipated [21], [75]–[77]. In the case of KSHV, it is tempting to speculate that the increased use of non-canonical start codons, particularly late during infection (48 and 72 hr), is a probable consequence of the high availability of ribosomes and the translational permissiveness observed during the lytic cycle. A plausible explanation for the non-canonical start codon selection could be the abundance of eukaryotic translation initiation factors (eIFs) such as eIF1 and eIF5. It has been shown that these two proteins have opposite effects in the selection of the start codon; eIF1 increases the stringency of codon selection favoring initiation from AUG initiator codons whereas eIF5 favors translation from non-AUG codons and AUG codons nested within a poor sequence context [78], [79]. The relative abundance and stability of these factors has not been characterized in KSHV-infected cells. One could propose that the ratio of these factors may change during reactivation, favoring the translation from non-AUG codons. Future studies exploring the sequence context of the start codons of viral sORFs, uORFs, and alternative variants of main ORFs, as well as studies aimed at uncovering the interplay between eIF1 and eIF5 during infection will undoubtedly shed light on the mechanisms governing the intricate translation patterns we observed during the KSHV lytic cycle.
Another surprise revealed in our studies was the abundance of uORFs encoded in the 5′ ends of viral transcripts. Like sORFs, uORFs have been found in eukaryotes, where they serve as modulators of gene expression during cellular stresses [80]. Their role has been extensively characterized in the face of amino acid starvation in yeast, where the accumulation of the transcription factor GCN4 depends on regulatory uORFs [81], [82]. This regulatory mechanism is evolutionarily conserved. In metazoans, uORFs regulate the abundance of the stress-responsive transcription factors C/EBPa/b, ATF4 and CHOP as well as the regulatory subunit of protein phosphatase 1, encoded by the GADD34 gene [58], [83], [84]. Although the role of uORFs in viruses remains largely unexplored, these regulatory elements permeate many viral families [14], [85]–[87] suggesting they may also control viral gene expression in instances where cellular stress pathways are engaged. In KSHV, the translational regulatory function of uORFs controlling the expression of ORF35 and ORF36, has been recently described [57]. The existence of a plethora of uORFs throughout the viral genome strongly indicates that this mechanism may be more widely used by KSHV than previously suspected.
Taken together, our results illustrate the dynamics of gene regulation during the different stages of the KSHV life cycle, while they also reveal that the coding capacity of its genome goes well beyond what was anticipated by in silico analyses. The striking conservation of the mechanisms employed by host and virus to regulate transcription, translation, and the diversity of the peptide repertoire, elevates KSHV as a valuable model system to explore the mechanistic underpinnings of the host-virus interactions in herpesviruses at large, as well as those of fundamental cellular processes, including the control of translation initiation in response to cellular stress.
iSLK and iSLK-219 cells (kindly provided by JinJong Myoung) were maintained in DMEM supplemented with 10% fetal bovine serum, L-glutamine (2 mM, Invitrogen), penicillin (100 IU/ml, Gibco) and streptomycin (100 ug/ml, Gibco) at 37°C under a 5% CO2 atmosphere. iSLK-219 cells were grown in the presence of puromycin (10 mg/ml, Invivogen) to maintain selection for the viral episome. LEC-219 cells were maintained in EBM-2 (Lonza cc3156) media supplemented with the EGM2-MV kit (Lonza cc3203) in presence of 0.25 ug/ml puromycin to maintain selection for the viral episome. BCBL-1 cells were maintained in RPMI supplemented with 10% fetal bovine serum, L-glutamine (2 mM, Invitrogen), penicillin (100 IU/ml, Gibco) and streptomycin (100 ug/ml, Gibco) at 37°C under a 5% CO2 atmosphere.
To induce entry to the lytic cycle, iSLK-219 cells were seeded at 1–2.5×105 cells/ml and 24 h after seeding (∼70–80% confluent) cells were induced with doxycycline (1 µg/ml, BD Bioscience cat 631311,). To prevent viral DNA replication in the cells collected at 24 hpi, we induced these cells with Dox in the presence of phosphonoformate (500 uM). All other time points were treated with Dox alone. At the indicated times after induction viral reactivation was evaluated by microscopy detection of the PAN-RFP reporter. To determine the timing of KSHV DNA replication, DNA was isolated at the indicated times following reactivation using the DNeasy blood and tissue kit following manufacturer guidelines (Qiagen cat 69581). The DNA (20 ng) was used for qPCR using primers for the LANA promoter (Fwd: 5′ AGGATGGAGATCGCAGACAC 3′; Rev: 5′ CCAGCAAACCCACTTTAACC 3′) or GAPDH as a normalization control (Fwd: 5′ AGCCACATCGCTCAGACAC 3′; Rev: 5′ TGGAAGATGGTGATGGGATT 3′).
Cells were washed and collected in RIPA buffer (10 mM Tris pH 8; 1 mM EDTA; 150 mM NaCl; 5% glycerol; 0.1% sodium deoxycholate; 0.1%SDS; 1%Triton x-100) at the indicated time points. Cell lysates were clarified and protein concentration was quantified using the Bio-Rad DC protein assay following manufacturer guidelines. 10 ug of protein per sample were fractionated by Bis-Tris PAGE (4–20% gels in MES SDS-running buffer) and transferred to nitrocellulose membranes. Immunoblots were processed, incubated with primary antibody overnight and developed using ECL reagent according to manufacturer recommendations.
The following antibodies were purchased from commercial suppliers: K8.1 (Advanced biotechnologies cat. 13-213-100) Actin (Sigma-Aldrich cat. A2228), ADAR1 (Abcam cat. ab126745), vCyclin (SCBT cat. 19415) and K-bZIP (SCBT cat. F33P1). The LANA rabbit polyclonal antibody was raised against a synthetic peptide from the acidic domain of LANA (Polson and Ganem, unpublished). The Kaposin B and C antibody was raised against the DR1/DR2 regions of the protein (Bisson and Ganem, unpublished). The ORF54 rabbit polyclonal antibody was raised against a synthetic peptide from the C-terminal region of the protein (aa 280–298 EDTNSVRKHTNEDNPVHEP) (Covance).
Ribosome profiling was performed as previously described [13], [21] with some modifications. In brief, iSLK and iSLK-219 cells were left untreated, or treated with cycloheximide (100 ug/ml, 1 minute) or harringtonine (1 ug/ml, 90 seconds) followed by cycloheximide (100 ug/ml, 1 minute). After treatment cells were washed in cold PBS twice and lysed in lysis buffer (20 mM Tris, 1% triton, 220 mM NaCl, 15 mM MgCl2, 1 mM DTT, 8% glycerol) supplemented with Turbo DNAse (Ambion cat. AM2238). The lysate was clarified, treated with RNAse I (Ambion cat. 2295) and overlaid on a 34% sucrose cushion. Monosomes were isolated by centrifugation at 69000 rpm for 4 hr in a TLA 110 rotor. Ribosome protected RNA fragments were isolated from the monosomal fraction by acidic-phenol extraction.
Total RNA was extracted from 1×107 cells using RNABee (AMSBIO cat. Cs-104B) following the manufacturer guidelines. Polyadenylated RNA was isolated from the total fraction using Oligotex mRNA kit (Qiagen cat. 70022). The resulting mRNA was partially fragmented by alkaline hydrolysis with sodium carbonate to ∼150 nt segments on average, and then fragments between 40–100 nt were isolated from gel. The laboratory of Jonathan Weissman has previously documented that this partial fragmentation results in the preferential accumulation of mRNA 5′ terminal fragments for most transcripts with a non-overlapping transcription start site [14]. We determined that for the transcripts where we were able to annotate a transcription start site, there is a 4–5 fold enrichment (average 5-fold, median 4-fold) of the number of reads for the first 10 nt of the transcript compared to three 10 nt windows within the gene body (nt 20–30, 30–40, 50–60 data not shown). The mRNA profiles in all figures show counts of the 5′-most bases of sequencing reads.
Strand specific libraries were generated as in Ingolia et al., 2012, with the modifications described in Stern-Ginossar et al., 2012. Samples were sequenced on either the Illumina Genome Analyzer II or the HiSeq 2000 using the truseq sbs kit v3 50 cycles (Illumina cat. FC-401-3002). Sequence analysis was done as described in Stern-Ginossar et al., 2012. Briefly, linker and polyA sequences were removed from the 3′ end of the reads preceding the alignment. Sequencing reads were aligned with Bowtie2 [88] allowing for 2 mismatches. Sequences aligning to rRNA were discarded and the remaining reads were aligned to KSHV (GQ994935.1) and human (hg19) genomes. For normalization, uniquely mapped reads were used to calculate the mRNA and ribosome footprint reads per kilobase per million (rpkM), and regions containing multi-mapped reads were masked out. The rpkM/gene for mRNA-Seq and Ribo-Seq were visualized using TreeView (1.1.6).
The sequences that did not align to the viral and host genomes were analyzed with TopHat and HMMSplicer for splice junction discovery using default options [22], [23]. We annotated splice junctions present in at least two time points, with an HMMSplicer score>900 and a TopHat score>7. We determined these thresholds based on previously characterized splice junctions. Two previously reported and two novel splice junctions had low or no TopHat score, but were included in the annotations due to their detection in multiple samples and high HMMSplicer score (ORFK1/ORF4, ORF70, ORF46/47, and Kaposin).
While other strong and reliable bioinformatics tools, such as Cufflinks and Scripture, are available for transcript reconstruction, we were not able to use them for viral mRNAs annotation. These tools are optimized for the analysis of transcripts in genomes where transcriptional units are well spaced and well defined. In the case of the compact KSHV genome, most transcripts use common regulatory features, are overlapping, or are very close to each other. Furthermore, and as shown in our study and previous studies by our lab and others, virtually the entire viral genome is transcribed late during the lytic cycle [6], [9]. These conditions make it difficult for these bioinformatics tools to parse out, identify and predict viral transcripts.
The identification of translation initiation sites was done using a machine learning approach as previously described Stern-Ginossar et al., 2012. In total our approach successfully predicted the translation initiation site of 64% (56 of 87) of the previously annotated ORFs (Tables 1, 2 and 3). The sites that were not predicted correspond to regions of low read coverage or overlapping ORFs.
Stretches of 5 or more consecutive adenosines (polyA), allowing one non-A base for every 5, were removed from the 3′ ends of mRNA sequencing reads before alignment. These reads were aligned to the viral genome and the 3′ end was determined as the last nucleotide before the start of the polyA stretch. To prevent false positives, polyA sequences were only used as evidence of 3′ termini when they mismatched the underlying reference genome sequence.
For the validation of the second splicing event within the 3′ end of the ORF57 transcript we prepared cDNA from 1 ug of total RNA using the qScript cDNA-Supermix (Quanta cat. 95048-025) or SuperScript III First Strand Synthesis System for RT-PCR (Invitrogen cat. 18080-051) with a combination of oligo(dT) and random hexamers, following manufacturers recommendations. PCR was done using as a template 1% of the resulting cDNA. The following primers flanking the 5′ and 3′ ends of the second splice junction were used for 30 cycles of amplification (Fwd: 5′ GGCAAAGACGACGAACTCAT 3′ Rev: 5′ GAGAAGAGACCACGCCTGACT 3′). The resulting products were separated in a 1.2% agarose gel and stained with ethidium bromide for visualization.
For the validation of the 3′ end extension of ORF57, PCR was done as described above using the following primers (A-Fwd: 5′ GGGTGGTTTGATGAGAAGGA, B-Fwd: 5′ TGGCAGAGTGTCTCCCGTAT, C-Rev: 5′ GAGAAGAGACCACGCCTGACT, D-Rev: 5′ ATAATGCCGAAGCCGTTATG)
Total RNA was extracted from cells at different times following induction of reactivation by doxycycline treatment, and cDNA was prepared as described above. Gene specific amplification was done using Phusion High-Fidelity DNA Polymerase (NEB cat. M0530S) for 32 cycles following manufacturer guidelines. The following primers were used for amplification: RTA-74281: Fwd-T7 5′ taatacgactcactatagggACGCGCTGTTGTCCAGTATT 3′, Rev-T3 5′ aattaaccctcactaaagggGTACAGTGTGCCGGACTCCT 3′; RTA-72795: Fwd-T7 5′ taatacgactcactatagggCCTCTCGAATGAGGACCAAA 3′, Rev-T3 5′ aattaaccctcactaaagggGTAGACCGGTTGGAAAACCA 3′; Kaposin-117809: Fwd-T7 5′ taatacgactcactatagggGTTGCAACTCGTGTCCTGAA 3′, Rev-T3 5′ aattaaccctcactaaagggAGGCTTAACGGTGTTTGTGG 3′ ; ORF6-6144: Fwd-T7 5′ taatacgactcactatagggGGGATACTTCTCGGGGAGAG 3′, Rev-T3 5′ aattaaccctcactaaagggGGCCCACTGTGCTCAGTAAT 3′; ORF63-102377: Fwd-T7 5′ taatacgactcactatagggGTTGGAAAATATCGCGTGCT 3′, Rev-T3 5′ aattaaccctcactaaagggTTGTGTGTTCGGTCCTGTGT 3′; ORF59-96434 Fwd-T7 5′ taatacgactcactatagggGGACGTGACCCTCCTGTCTA 3′, Rev-T3 5′ aattaaccctcactaaagggTAACGTCTCCACTGCCTTCC3′.
Total RNA was extracted from cells using RNABee (AMSBIO cat. Cs-104B) following manufacturer instructions. Northern blotting was done for 10 ug of RNA, or 100 ng of mRNA per lane using the Ambion NorthernMax system (Invitrogen cat. AM1940). Gene and sense specific riboprobes were synthesized from PCR products using the Ambion MAXIscript T7-T3 Transcription Kit (Ambion cat. AM1326) according to manufacturer guidelines. The primers used for the PCR amplification of gene specific probes are: Probe A (24151–25437): Fwd-T7 5′ taatacgactcactatagggagaCAGTCACAAGCACACAACCC 3′, Rev-T3 5′ aattaaccctcactaaagggagaTTCGGGTGATTAAGCAAAGG 3′; Probe B (25437–25929) : Fwd-T7 taatacgactcactatagggagaCCTTTGCTTAATCACCCGAA, Rev-T3 5′ aattaaccctcactaaagggagaGGTGACCGTACTGCCATACC 3′; Probe C (123441–122984): Fwd-T7 5′ taatacgactcactatagggCGCTAACAGGGGAAACGTTAACCTGC 3′, Rev-T3 5′ aattaaccctcactaaagggCTCATTGCCCGCCTCTATTA 3′
File S2 contains the updated KSHV 2.0 annotations, mRNA profiles, and ribosome occupancy (CHX) plots. The database can be opened using the Mochiview file software [26], free for download at http://johnsonlab.ucsf.edu/sj/mochiview-software/
To open the file, import and activate the database. Restart the program and the new database will be available for viewing. The following files are included: GQ994935 sequence, KSHV2.0 location, mRNA_dox8h_minus, mRNA_dox8h_plus, mRNA_dox24h_minus, mRNA_dox24h_plus, mRNA_dox48h_minus, mRNA_dox48h_plus, mRNA_dox72h_minus, mRNA_dox72h_plus, fp_dox8h_minus, fp_dox8h_plus, fp_dox24h_minus, fp_dox24h_plus, fp_dox48h_minus, fp_dox48h_plus, fp_dox72h_minus, fp_dox72h_plus, fp_harr_dox48h_minus, fp_harr_dox48h_plus, fp_harr_dox72h_minus, fp_harr_dox72h_plus
|
10.1371/journal.pgen.0030206 | The Role of Carcinine in Signaling at the Drosophila Photoreceptor Synapse | The Drosophila melanogaster photoreceptor cell has long served as a model system for researchers focusing on how animal sensory neurons receive information from their surroundings and translate this information into chemical and electrical messages. Electroretinograph (ERG) analysis of Drosophila mutants has helped to elucidate some of the genes involved in the visual transduction pathway downstream of the photoreceptor cell, and it is now clear that photoreceptor cell signaling is dependent upon the proper release and recycling of the neurotransmitter histamine. While the neurotransmitter transporters responsible for clearing histamine, and its metabolite carcinine, from the synaptic cleft have remained unknown, a strong candidate for a transporter of either substrate is the uncharacterized inebriated protein. The inebriated gene (ine) encodes a putative neurotransmitter transporter that has been localized to photoreceptor cells in Drosophila and mutations in ine result in an abnormal ERG phenotype in Drosophila. Loss-of-function mutations in ebony, a gene required for the synthesis of carcinine in Drosophila, suppress components of the mutant ine ERG phenotype, while loss-of-function mutations in tan, a gene necessary for the hydrolysis of carcinine in Drosophila, have no effect on the ERG phenotype in ine mutants. We also show that by feeding wild-type flies carcinine, we can duplicate components of mutant ine ERGs. Finally, we demonstrate that treatment with H3 receptor agonists or inverse agonists rescue several components of the mutant ine ERG phenotype. Here, we provide pharmacological and genetic epistatic evidence that ine encodes a carcinine neurotransmitter transporter. We also speculate that the oscillations observed in mutant ine ERG traces are the result of the aberrant activity of a putative H3 receptor.
| During signaling in the nervous system, individual nerve cells transfer information to one another by a complex process called synaptic transmission. This communication involves the release of a specific neurotransmitter into the synaptic cleft, which then triggers signaling in the downstream neuron by binding to and activating specific cell surface receptors. In order to terminate the neuronal signal, the neurotransmitter must be rapidly removed from the synaptic cleft. This is done by two mechanisms: the neurotransmitter can be degraded or modified, or the transmitter can be taken up by the presynaptic neuron and packaged into vesicles for reuse. In the compound eye of the fruitfly D. melanogaster, the photoreceptor cell responds to light and releases histamine into the synaptic cleft. This signal is terminated by the removal of histamine from the synapse and the enzymatic conversion of histamine to carcinine. We have shown that it is not sufficient just to modify the histamine neurotransmitter, but it is also important to remove carcinine from the photoreceptor synapse. The failure to adequately remove carcinine results in defects in the visual transduction process. Moreover, the work suggests that carcinine itself modulates vision by regulating histamine release into the synapse.
| An exceedingly complex regulation is involved in the synthesis, release, activity, and recycling or degradation of neurotransmitter in the nervous system of animal species. This regulation may involve neurotransmitter degradation within the synapse [1], reuptake of neurotransmitter by presynaptic neurons [2], recycling of neurotransmitter by neighboring cells [3], and/or activation of receptors that trigger positive/negative feedback loops resulting in an increase/decrease of neurotransmitter release in presynaptic neurons [4]. Many of the mechanisms and machinery components associated with this regulation have been well conserved across species, ranging from Caenorhabditis elegans to humans [5]. D. melanogaster is often utilized when studying neurotransmitter dynamics because of its malleable genetics, strong phenotypes in the presence of neurotransmission defects, and its sensitivity to numerous neuropharmacological compounds that have been shown to exert similar effects in humans (for review, see [6]).
The major neurotransmitter released by photoreceptor cells in Drosophila is histamine [7], which is biosynthesized from histidine by the enzyme histidine decarboxylase (Hdc) found within the photoreceptor cell [8]. Upon excitation by light, the photoreceptor cell depolarizes and releases histamine, which then binds to a postsynaptic histamine-gated chloride channel, resulting in the hyperpolarization of the postsynaptic neuron [9,10]. Each synaptic cartridge is surrounded by three glial cells that invaginate into the photoreceptor terminals by means of fingerlike projections known as capitate structures [11]. While the exact site of histamine reuptake is currently unknown, the histamine remaining in the synaptic cleft is thought to be rapidly taken up by glial cells, possibly at the site of these capitate structures [12]. In glial cells this histamine is converted by the enzyme N-β-alanyl dopamine synthase, encoded by the ebony gene in Drosophila, into β-alanyl-histamine, also known as carcinine [3,13]. This carcinine is then released by the glial cell as an “inactive” conjugate of histamine, again possibly at the site of capitate structures, where it is then taken up by the presynaptic neuron. Once in the photoreceptor cell, the carcinine is converted by the enzyme N-β-alanyl dopamine hydrolase, encoded by the tan gene, back into the original neurotransmitter histamine [13,14]. It is proposed that the combination of histamine biosynthesis through histidine decarboxylase and the recycling of histamine by ebony and tan enzymes defines the total pool of histamine available at the photoreceptor cell synapse. The transporters responsible for histamine uptake by glial cells, and for carcinine internalization by photoreceptor cells, are currently unknown.
The ine gene is believed to encode a putative neurotransmitter transporter, and two ine cDNAs have been sequenced and identified [15],[16]. The shorter cDNA, ine-RB, encodes the protein Ine-P2, while the longer cDNA, ine-RA, encodes the protein Ine-P1, which contains an additional ∼300 amino acids at its N terminus. The function of the additional N-terminal region of ine-P1 is currently unknown. Despite efforts to identify the neurotransmitter transported by inebriated in transfected Xenopus laevis oocytes, the substrate of the inebriated protein has remained elusive [17]. Mutations in the ine gene result in an increase in the rate of onset of long-term facilitation at the larval neuromuscular junction [18], as well as an increase in the neuronal excitability associated with mutations in the Shaker gene, which encodes the α-subunit of a potassium channel [19]. Both of these neuronal excitability phenotypes are believed to be caused by the defective reuptake of an unknown neurotransmitter, and thus the overstimulation of postsynaptic neurons. A third, and less understood, phenotype associated with ine mutations is manifested as an aberrant electroretinogram (ERG) [15,20]. ERGs measure the mass retinal response of the eye to a stimulus of light, and the ERG of ine mutants is characterized by several defects, including most noticeably a series of strong oscillations in the presence of light [15,20]. Recently, in an excellent and comprehensive review of histaminergic neuronal signaling in arthropods, it was proposed that the ine gene in Drosophila might encode the carcinine neurotransmitter transporter [21]. Here, we provide genetic and pharmacological evidence linking the mutant ine-associated phenotype with the buildup of carcinine in the photoreceptor synaptic cleft and with the activity of a putative H3 receptor in the Drosophila eye.
The repo-GAL4, “long”-GMR-GAL4, w1118, tan1, tan2, e1, e11, ortPbac fly lines were all obtained from Bloomington Stock Center. HdcP218 and ort5 fly mutants were obtained from W. Pak (Purdue University) while the UAS-ine-RB and ine2 transgenic fly lines were obtained from M. Stern (Rice University). All stocks were maintained in constant darkness at room temperature. Flies carrying two or three mutations/transgenes were generated by standard genetic methodologies. All wild-type flies were of the w1118 background.
Flies were anesthetized by exposure to carbon dioxide and immobilized within a rotating disc using a drop of molten myristic acid (Akros). To record voltage changes within the eye, an electrode filled with signa gel (Parker Labs) was placed on the surface of the eye, while a second gel-filled electrode was gently inserted into the thorax. For light treatments a halogen lamp controlled by a Model T132 Uniblitz shutter was used. All light treatments, unless otherwise stated, were performed using a 580 nm filter and were 4 s in duration. To reduce the effects that exogenous sources of histamine might have on ine2HdcP218 flies during ERG analysis, these animals were starved for 24 h before testing. To induce depolarization spikes in ort5 flies, two 4-s pulses of 480 nm light, followed by two 4-s pulses of 580 nm of light, were delivered to the flies, and a trace was taken during the second 580 nm pulse for analysis. Voltage changes were amplified using a DAM50 amplifier (World Precision Instruments) recorded using Powerlab 4/30 (AD Instruments, Colorado Springs, CO) and analyzed using Chart 5 software (AD Instruments). Oscillation frequency was determined by counting and averaging the number of repolarization spikes observed within 0.2 s of light exposure in either ine2 or carcinine-treated fly ERG recordings.
Thioperamide, immepip, and histamine were obtained from Sigma and carcinine was obtained from Peninsula Laboratories. All compounds were reconstituted in sterilized water for long-term storage. Flies were treated overnight in vials containing Kimwipes soaked with 200 μl of 1% sucrose solution with or without drug compound. Histamine was delivered to flies at a concentration of 10% [22]. Thioperamide and immepip were used at 0.5% and carcinine at either 5% or 10%. Flies were starved for 24 h before drug treatment.
RNA was purified from either embryos or adult fly heads of the w1118 background by employing an RNeasy Mini Kit (Qiagen Sciences). cDNA was generated from purified RNA by utilizing MMLV-Reverse Transcriptase (Fisher Scientific). Forward primers specific to either the ine-RA or ine-RB transcripts and a reverse primer common to both transcripts were obtained from Integrated DNA Technologies. The ine-RA forward primer was ATCGATGGCCACTTCCGGATTACA, the ine-RB forward primer was ATCAGTTGCCACTCCCAGTTTCCA, and the reverse primer used to generate PCR product from both transcripts was TATCCTATGCAGGCCAGGACGAAT. Products were generated and amplified by means of PCR using Taq polymerase, buffers obtained from Invitrogen, and a TECHNE, TC-312 thermocycler (Bartoworld Scientific) using the following parameters for 35 cycles (94 °C for 30 s, 55 °C for 30 s, and 72 °C for 90 s). PCR products were separated in a 1% agarose gel and then stained using ethidium bromide.
An ERG recording from a wild-type fly (Figure 1A) contains a receptor component, or the depolarization response upon exposure to light, and on and off transient spikes that indicate the response downstream of the photoreceptor cell (arrows, Figure 1A). The ERG of ine mutants contains an intact receptor component but with the addition of an initial depolarization spike (unfilled arrowhead, Figure 1B) and prominent oscillations superimposed on the depolarization response (Figure 1B). These oscillations have a wide range of frequencies from 40–90 Hz. These mutants also possess reduced on and off transients (arrows, Figure 1B), indicating impaired photoreceptor synaptic transmission. Finally, these ine mutant ERGs often display a hyperpolarization following a light response (arrowhead, Figure 1B). We observed all of these previously described ERG phenotypes when using either ine2 or ine3 allele flies. The ine3 allele is the result of a deletion of the majority of the ine open reading frame common to both ine-RA and ine-RB [16], while the mutation associated with ine2 was identified as being a nonsense mutation in codon 125 of the ine gene and is believed to only affect the ine-RA-encoded isoform [23]. Because the ine3 allele is associated with reduced viability, and because we could discern no observable difference between the ERG traces of ine2 and ine3 flies, we made use solely of the ine2 fly line for all of our experiments and genetic crosses.
Intracellular voltage recording experiments suggest that the oscillations observed in ine mutants originate in the photoreceptor cell, and that they are not the result of synaptic feedback [24]. However, if ine does encode a neurotransmitter transporter, its expression and localization are not necessarily restricted to photoreceptor cells, as neurotransmitter transporters often function from neighboring glial cells. Indeed, a previous study demonstrated that expression of ine in either neurons or glial cells was sufficient to rescue several mutant ine-associated defects at the neuromuscular junction [23]. In order to confirm that the inebriated protein is needed at the photoreceptor cell synapse, we tested whether the ine2 mutant phenotype could be rescued by expressing the ineRB transcript in photoreceptor and glial cells. The UAS-ine-RB transgenic fly line [23,25] contains the ine-RB cDNA under the control of the upstream activator sequence of the yeast Gal4 transcription factor. These UAS-ine-RB flies will only express Ine-P2 when crossed with a second line of flies expressing Gal4. The Gal4 lines utilized were “long”-GMR-GAL4, which expresses Gal4 protein specifically in photoreceptor cells [26], and repo-GAL4, which expresses Gal4 in glial cells [27]. A strong rescue of the ine2 ERG phenotype was observed when ine-RB was expressed in either photoreceptor or glial cells (Figure 1C and 1D) compared to non-rescued ine2 or wild-type control (w1118) ERGs (Figure 1A and 1B). As expected, the UAS-ine-RB transgene failed to rescue the ine2 phenotype if neither GAL4 transgene was present (unpublished data). The oscillations observed in ERG traces from these transgenic rescued animals were either absent or greatly reduced, and the hyperpolarization response was also significantly diminished. Finally, the rescued ERG traces contained larger on and off transients than the ine2 non-rescued controls (arrows, Figure 1C and 1D). Expression of ine-RB in glial cells appears to give a stronger and more consistent rescue of the ine2 ERG phenotype than when expressed in photoreceptor cells. This may be due to stronger expression of the Gal4 transcription factor in repo-GAL4 flies than in GMR-GAL4 animals, or it may be due to the need for full-length ine-RA, rather than ine-RB, expression in photoreceptor cells. It is also surprising that ine-RB expression has the ability to rescue the ine2 ERG phenotype, as the ine-RB transcript was previously thought to remain intact in ine2 mutant flies [25]. These results suggest that ine-RB is normally expressed at only low levels compared to ine-RA, and that overexpression of ine-RB is sufficient in compensating for the loss of ine-RA associated with ine2 mutants. Reverse transcriptase PCR experiments confirm these suspicions; ine-RB is expressed at low levels in adult wild-type heads compared to robust expression of ine-RB in the developing embryo (Figure 1E). The ine-RA transcript was found at high levels in both wild-type embryos and adult heads (Figure 1E).
The ability to rescue the ine2 ERG response by expressing inebriated protein in photoreceptor and glial cells suggests that inebriated functions primarily at the site of the photoreceptor cell in the eye. Histamine is believed to be the predominant neurotransmitter that signals between photoreceptor cells and second order laminar neurons in the optic lobe [28], and it is possible that inebriated serves as a histamine transporter. However, previous studies showed that when inebriated protein from the tobacco hornworm Manduca sexta, which has significant homology to the inebriated protein from Drosophila, was expressed in Xenopus laevis eggs, it was unable to transport histamine across the cell membrane [17]. However, these authors do propose that a second unknown protein may be required to assist inebriated in proper neurotransmitter transport function, or that inebriated may possess different substrates in Manduca compared to Drosophila. Thus, histamine could still be the substrate of inebriated in Drosophila. Histamine is generated by the activity of histidine decarboxylase, encoded by the Hdc gene in Drosophila (Figure 2A). Mutations in the Hdc gene, such as in the case of the HdcP218 allele, result in flies possessing disrupted photoreceptor synaptic transmission, as demonstrated by the lack of on and off transients in their ERGs ([8], and Figure 2C). Approximately 80% of flies that were homozygous for both the HdcP218 and ine2 alleles displayed ERGs with no oscillations (Figure 2D) when compared to ine2 controls (Figure 2B). There was a small percentage (∼20%) of HdcP218 ine2 flies that displayed weak or delayed oscillations. However, the ERGs from these flies that displayed this weak rescue also possessed on and off transients, suggesting that the HdcP218 allele was either not fully penetrant in these double mutants, or that their food provided an outside source of histamine. This is not surprising, as Drosophila photoreceptors are known to regain some function from exogenous histamine taken up at minute levels in their food [22]. The ERGs from HdcP218 ine2 also often lack the hyperpolarization response characteristic of ine2 flies (Figure 2D). These findings suggest that histamine production or signaling plays a strong role in the oscillation and hyperpolarization phenotype observed in ine2 traces.
The postsynaptic receptor for histamine in Drosophila is a histamine-gated chloride channel (Figure 2A), and a subunit of this channel is encoded by the ora transientless (ort) gene [10]. The ort5 and ortPbac alleles both result in reduced activity of this histamine receptor in Drosophila, as shown by the lack of on and off transients in their ERG traces ([10] and Figure 2E). If ine encodes a histamine neurotransmitter transporter, then reduced function of this protein may result in an excess of histamine in the synaptic cleft, and this excess of neurotransmitter may be acting upon this postsynaptic histamine receptor to somehow generate the observed oscillations. If this were the case, then ine2;ort double mutants should have reduced oscillations. However, neither ine2;ort5 (Figure 2F) nor the ine2;ortPbac (unpublished data) double mutants exhibited rescue of the oscillation or hyperpolarization components of the ine2 ERG recordings, indicating that the oscillations do not arise from histamine signaling in downstream neurons. Moreover, since ort mutations block signaling in laminar neurons, these data are consistent with the oscillations being generated in the photoreceptor cells. Surprisingly, the ort5 allele, which is the result of a frameshift mutation and therefore likely serves as a null allele for this gene [10], often displays strong depolarization spikes of its own in the receptor component of its ERG trace (Figure 2G). Therefore, mutations in Hdc, which block the formation of histamine, rescue ine2 whereas mutations in ort, which still allow for the synthesis of histamine, fail to rescue.
The ablation of the ine mutant ERG phenotype upon the introduction of Hdc, but not ort, mutations, suggests that histamine is involved in generating the ine2 ERG phenotype, but that histamine's downstream signaling in the optic lobe is not. The recycling pathway of histamine in the eye has been well elucidated [13,14,29,30]. It has been shown that following release into the synaptic cleft, histamine is rapidly taken up by neighboring glial cells and is converted by the β-alanyl-dopamine synthase, encoded by the gene ebony, into β-alanyl-histamine, also known as carcinine (Figure 3A). This carcinine is then transported into the presynaptic photoreceptor cell and is converted back into histamine by β-alanyl-dopamine-hydrolase, encoded by the gene tan, for use as a recycled source of neurotransmitter (Figure 3A). Both tan and ebony mutations in Drosophila are associated with significant reductions in size of the on and off transients in ERG traces (Figure 3C and 3E), due to the loss of this recycled pool of histamine in the eye. Introduction of either tan2 (unpublished data) or tan1 mutations into an ine2 background failed to have any effect in reducing the size of the oscillations or hyperpolarization response when compared to ine2 mutants alone (Compare Figure 3B with Figure 3D). However, ine2;ebony1 (unpublished data) or ine2;ebony11 double mutants displayed complete rescue of the oscillation phenotype in all flies tested (Compare Figure 3B with Figure 3F). These data, combined with the fact that histamine synthesis is necessary for the presentation of a mutant ERG phenotype in ine2 flies, provide genetic evidence that carcinine is involved in generating ine2-associated oscillations.
If ine encodes a carcinine neurotransmitter transporter, as the genetic evidence above suggests, than a potential cause of the aberrant ERG phenotypes seen in ine mutants could be the buildup of carcinine within the photoreceptor synaptic cleft. In order to test whether or not carcinine is able to induce an ine2-like ERG phenotype in wild-type animals, w1118 flies were treated with 5% carcinine overnight and then subjected to ERG analysis. Approximately 35% of the w1118 flies treated with carcinine displayed occasional weak oscillations or brief depolarization/repolarization spikes in the photoreceptor response of their ERG traces (Figure 4B and 4C, compare to Figure 4A). While these spikes exhibit no consistent frequency, unlike the oscillations seen in ine2 recordings, the carcinine-induced ERG disturbances were never observed in untreated starved flies. If carcinine was delivered to ebony11 flies, which lack the ability to synthesize carcinine from histamine, they surprisingly displayed an abnormal ERG trace. All ebony11 flies treated with carcinine manifested phenotypes reminiscent of those seen in ine2 ERG traces, including sharp depolarization spikes in response to a light response, weak oscillations, and a hyperpolarization peak upon the termination of light (Figure 4E and 4F, compare to Figure 4D). The oscillations observed in carcinine-treated ebony11 mutants, while only appearing briefly during the initiation of light exposure, were seen at a similar frequency as those found in ine2 ERG recordings (63 spikes/s). As discussed below, a possible mechanism underlying these carcinine-induced ERG disturbances may involve the sensitization of a putative histamine/carcinine receptor.
If inebriated does serve as a carcinine transporter, and if carcinine is indeed building up within the synaptic cleft in ine2 mutants, then this uncleared carcinine appears to somehow be acting on some synaptic receptor to initiate this aberrant oscillation phenotype. The ine2;ort5 experiments suggest that this receptor is not the post-synaptic histamine-gated chloride channel. In mammals and various other vertebrate systems, presynaptic histaminergic neurons often contain their own histamine receptors, known as H3 receptors. The H3 receptor is a G-protein–coupled receptor that was first identified in 1983 by Arrang et al. [4] and is now known to act as a presynaptic autoreceptor that inhibits histamine release from histaminergic neurons in the brain (for review, see [31]). Thus, H3 receptors serve to negatively regulate histamine release and synthesis in the presence of high histamine levels in the synaptic cleft. While no H3 receptor has been identified yet in Drosophila, there are several candidate genes that may encode such a putative receptor. There are numerous well-characterized pharmaceutical compounds that act as agonists, antagonists, or inverse agonists of the H3 receptor in vivo in mammals, and recently carcinine was identified as being an inverse agonist of this receptor in mice [32]. It was shown that, rather than reduce histamine release, as occurs in the case of histamine binding to an H3 receptor, carcinine had the opposite effect and induced both histamine synthesis and release from presynaptic histaminergic neurons in vivo.
A possible scenario to explain the oscillations seen in ine2 ERGs is that histamine and uncleared carcinine are competing for binding to putative H3 receptors, resulting in opposing signaling cascade responses in the photoreceptor cell. If this is the case, disrupting this balance of histamine and carcinine binding to the putative H3 receptor in ine2 fly eyes should result in a rescue of ERG oscillations. Indeed, treatment of ine2 flies with 10% carcinine resulted in a rescue of oscillations in 35% of flies (unpublished data), and treatment of ine2 flies with 0.5% thioperamide, another well-characterized and potent inverse agonist of the H3 receptor in mammals, resulted in the consistent and complete ablation of oscillations in ERG traces in all flies tested (Compare Figure 5A with Figure 5C). In addition, treatment of ort5 flies with 0.5% thioperamide resulted in a loss of ort5-associated depolarization spikes (unpublished data). Surprisingly, treatment of wild-type control flies with 0.5% thioperamide resulted in the loss of on and off transients in their ERGs (compare Figure 5B with Figure 5D).
It should also be possible to disrupt the hypothetical balance of histamine and carcinine binding to a photoreceptor cell-specific H3 receptor in ine2 by introducing an H3 receptor agonist, such as histamine itself. Indeed, treatment of ine2 flies with 10% histamine (unpublished data) or 0.5% immepip (Figure 5E), another potent H3 receptor agonist, resulted in a strong rescue of oscillations in >50% of flies tested. Occasionally, weak oscillations and depolarization spikes were still observed in immepip- or histamine-treated ine2 flies (Figure 5E). Neither histamine nor immepip treatment had a strong or consistent effect on the on and off transients seen in wild-type control ERGs. Since immepip and thioperamide are known to be specific agonists and inverse agonists of the mammalian H3 receptor, these pharmacological experiments suggest that an H3 receptor may exist in Drosophila and that abnormal stimulation of this H3 receptor is occurring in the eyes of ine2 Drosophila mutants.
Our findings indicate that the presumed neurotransmitter transporter encoded by the ine gene in Drosophila transports the histamine metabolite carcinine. We show using genetic epistasis that the oscillations observed in mutant ine ERGs require histidine decarboxylase activity and the carcinine-synthesizing enzyme ebony, but not the carcinine-hydrolyzing enzyme tan. We also reveal that treating wild-type flies with carcinine can phenocopy components of the mutant ine ERG phenotype. Finally, by rescuing the ine2-associated phenotype with drugs that target the mammalian H3 receptor, we provide pharmacological evidence for the presence of a putative H3 receptor in Drosophila that may be responsible for the ERG oscillations observed in flies carrying mutations in the ine gene.
Previous studies involving intracellular voltage recordings of ine mutants led the authors to conclude that the oscillations observed in ine mutant ERGs were the result of a defect occurring within the photoreceptor cell [24]. We were able to support these conclusions by expressing ine specifically in photoreceptor cells and demonstrating a rescue of the ine2-associated oscillations. Neurotransmitter transporters are often able to function from either the presynaptic neuron or from neighboring glial cells, as shown at the neuromuscular junction in ine mutants [23]. We found that glial cell–specific expression of the ine gene in ine2 flies resulted in a complete rescue of the ine mutant ERG phenotype. It was somewhat unexpected that ine expression in glial cells rescued the ine2 phenotypes, as glial cells have been shown to lack tan protein and thus would be unable to convert carcinine back to a recycled pool of histamine [30]. However, it is possible that glial cells do express trace amounts of the enzyme tan to hydrolyze carcinine and generate a renewable source of histamine for photoreceptor cells, and it is also possible that the inebriated protein is expressed in a non-autonomous manner and can be transported from glial cells to photoreceptors in the fly eye.
The finding that an ERG recording can exhibit oscillations is somewhat surprising. An ERG does not record the electrical response of a single photoreceptor, but rather is a collective measure of the retinal photoresponse. Thus, if the mutant ine-associated ERG defects are indeed localized to the photoreceptor synapse, as our data and that of previous labs suggest, then one would expect that different photoreceptors would be excited/inhibited at different timepoints, ultimately resulting in the oscillations simply canceling themselves out. The fact that oscillations are indeed observed, and appear to be due to a defect occurring at the photoreceptor synapse, implies the existence of an uncharacterized and complex synchronization of photoreceptor cell de-/repolarization.
The lack of rescue of ine2-associated oscillations in flies carrying additional mutations in the postsynaptic histamine receptor gene ort, the finding that mutant ine oscillations were detected within single photoreceptor cells [24], and our observations that the mutant ine phenotype can be rescued when ine is expressed in photoreceptors, all combine to strongly suggest that the oscillation phenotype is likely a result of a defect occurring within the photoreceptor itself. In addition, by crossing ine2 animals with HdcP218 flies, we demonstrated that the ine2-associated oscillations are dependent upon histamine synthesis. All of these results indicate that histamine is somehow contributing to the aberrant ERG witnessed in ine2 flies, and that histamine appears to be acting on the presynaptic photoreceptor cell to induce this oscillation phenotype. Further epistatic analyses also revealed that ebony, but not tan, activity is required for the generation of oscillations in ine2 ERGs (Figure 6A). These genetic experiments are consistent with ine encoding either a carcinine importer found in the photoreceptor cell or a carcinine exporter found in glial cells. The homology of inebriated with other known Na+/Cl− neurotransmitter transporters (which import neurotransmitter into cells) [16] suggests that inebriated protein is transporting carcinine into the photoreceptor, and not out of glial cells.
While ebony is known to act on multiple substrates, such as dopamine to generate β-alanyl-dopamine [13], the requirement of histamine synthesis for the maintenance of ine2-associated oscillations suggests that it is β-alanyl-histamine, or carcinine, that is somehow responsible for the oscillations observed in ine2 ERGs. It should be noted, however, that ebony mutations were not sufficient in rescuing the hyperpolarization response observed in mutant ine ERG traces (Figure 3F). The origins of this hyperpolarization response are still unclear and further research will be required to elucidate its exact meaning. In tan mutants, one would predict that there would be a buildup of carcinine. However, this buildup does not give rise to an ERG recording similar to that of ine2. This is due most likely to the presence of functional inebriated protein in tan mutant flies, which should effectively clear the carcinine from the synaptic cleft for degradation within the photoreceptor cell.
By treating wild-type and ebony11 flies with carcinine and subsequently inducing components of the ine2-ERG phenotype, we provide further evidence suggesting that the sharp depolarization spike, the oscillations, and the hyperpolarization response all seen in ine2-ERGs are due to a buildup of carcinine within the photoreceptor synaptic cleft. While the oscillations observed in carcinine-treated wild-type flies do not mimic exactly the oscillations seen in ine2 ERG recordings, it is presumably difficult to replicate the carcinine and histamine balance occurring in the eyes of ine2 animals. Indeed, treatment of wild-type flies with higher (10%) or lower (1%) concentrations of carcinine were less effective in inducing the oscillations than the described 5% carcinine dose (unpublished data).
It is possible that carcinine is being degraded or modified by the fly before the compound is able to exert its effects at the photoreceptor cell. In order to eliminate the activity of one enzyme known to be involved in carcinine metabolism, tan1 flies were treated with 5% carcinine overnight. Surprisingly, none of the tan1 flies treated with carcinine showed an aberrant ERG phenotype (unpublished data). It was surprising that carcinine treatment had a strong effect in flies of the ebony11, but not the tan1, background. While the results of these tan1 and ebony11 carcinine-treatment experiments are unexpected, one possible explanation may involve the regulation of carcinine clearance/degradation. The tan1 flies presumably suffer from a perpetual excess of carcinine even before exogenous carcinine treatment, and these flies, in order to reduce their sensitivity to this compound, may consequently decrease the levels of a putative carcinine receptor, increase their rate of carcinine degradation, or increase the levels of inebriated protein for carcinine clearance. However, ebony11 flies are relatively “naïve” to the effects of carcinine, as their ability to synthesize this compound has been greatly diminished, and as a result these animals may have an increased level of the supposed carcinine receptor, a decrease in inebriated receptor levels or a decrease in carcinine degradation, ultimately making them more sensitive to the effects of carcinine treatment.
It remains to be seen whether or not all of the mutant ine-associated phenotypes, including increased neuronal excitability [19],[23] and increased sensitivity to osmotic stress [25], are due to the inability of these flies to transport carcinine. It is possible that the inebriated protein transports other compounds that perhaps share the common feature of β-alanine conjugation. This might help explain why none of the more common neurotransmitters were taken up by ine-transfected Xenopus oocytes [17]. In order to assist in confirming that inebriated is indeed a carcinine neurotransmitter transporter, in vitro experiments, such as neurotransmitter uptake assays, will need to be performed. In addition, the ability of inebriated protein to take up other β-alanyl-neurotransmitters/osmolytes also should be examined.
The oscillations present within the photoreceptor response of ine2 ERGs appear as sharp depolarization/repolarization spikes, and this oscillation phenotype is dependent upon both histamine synthesis and ebony activity (Figure 6A), and is sensitive to drugs that target mammalian H3 receptors. It is perplexing that the synthesis of a single metabolite, carcinine, could be responsible for both the depolarization and repolarization spikes observed within ine mutant ERGs. We speculate that these oscillations are the result of aberrant signaling involving both carcinine and histamine at a putative H3 receptor in Drosophila (Figure 6B and 6C). H3 receptors are an unusual example of the G-protein coupled receptor family, in that they have partial constitutive activity, resulting in a constant small percentage of stimulated G-proteins [33] that trigger a reduction of histamine synthesis and release [34] as well as a decrease in extracellular calcium inflow [35,36,37]. The presence of an H3 receptor agonist, such as histamine, causes an increase in activity of the associated G-protein and therefore a stronger inhibition of both histamine release and calcium inflow. Thus, synaptic histamine serves as a negative regulator for its own release and induces a slight repolarization of a stimulated presynaptic histaminergic neuron by inhibiting presynaptic calcium channels. An H3 receptor inverse agonist is believed to act by blocking the constitutive activity of the H3 receptor, resulting in the liberation from a histamine release checkpoint as well as the release of restrictions on calcium inflow [38]. Recently, it has been shown that carcinine has the ability to act as an inverse agonist of presynaptic H3 receptors in mice [32]. While significant further research is required to confirm this hypothesis, we surmise that histamine and carcinine are exerting opposing effects on the polarization state of the histaminergic photoreceptor cell by activating or inhibiting presynaptic calcium channels via a putative Drosophila H3 receptor. While a recent search of the Drosophila genome did not uncover any direct homologs to vertebrate metabotropic histamine receptors [39], the CG7918 gene was listed as a possible candidate for encoding such a receptor, and this gene bears strong homology to genes encoding H3 receptors in mammals. In addition, the ine2-associated oscillations display sensitivity to mammalian H3 receptor agonists and inverse agonists, strengthening the possibility that an H3 receptor does exist in Drosophila. It is still unclear what the origins of the thioperamide-sensitive depolarization spikes are that are observed in ort5 ERGs. The presence of these thioperamide-sensitive spikes in ort5 ERG recordings implies the requirement of some postsynaptic retrograde signal for ERG stability, and this ort-dependent signal may be involved in the sensitization of the putative H3 receptor.
It was unexpected that thioperamide treatment of wild-type flies resulted in the loss of on and off transients within their ERG traces. It is possible that histamine release was so extreme in the presence of the potent thioperamide that histamine levels were nearly depleted in the eye, resulting in the disruption of downstream signaling events. Indeed, treatment of mice with high concentrations of carcinine, which acts as an inverse agonist of H3 receptors similar to thioperamide, was shown to result in significantly lower overall levels of histamine within the brains of treated mice [32]. This model of indirect histamine depletion has also been postulated to occur in ebony mutant flies. The absence of on and off transients in ebony mutant ERG recordings is attributed to the normal release of histamine by photoreceptor cells, but this histamine subsequently lacks the ability to be “trapped” by β-alanine conjugation, ultimately resulting in histamine diffusing away from the eye [13]. Interestingly, expression of pertussis toxin in photoreceptor and laminar neurons in Drosophila results in a similar loss of on and off transients in ERG traces, and this is believed to be the result of inactivation of an unknown G-protein coupled receptor found in photoreceptor cells that is unlikely to be rhodopsin [40]. It is possible that pertussis toxin was acting within photoreceptor cells upon the putative H3 receptor in this study, resulting in a lack of negative feedback on histamine synthesis/release, eventually causing the exhaustion/depletion of histamine pools. Further research will be required to confirm or dismiss the presence of a histamine/carcinine-sensitive H3 receptor in Drosophila photoreceptor cells.
The National Center for Biotechnology Information (NCBI) Entrez (http://www.ncbi.nlm.nih.gov/sites/gquery) accession numbers for the genes discussed in this paper are ebony, 42521; Hdc, 36076; ine, 33659; ort, 54910; and tan, 4478. |
10.1371/journal.pntd.0006811 | SIV/SHIV-Zika co-infection does not alter disease pathogenesis in adult non-pregnant rhesus macaque model | Due to the large geographical overlap of populations exposed to Zika virus (ZIKV) and human immunodeficiency virus (HIV), understanding the disease pathogenesis of co-infection is urgently needed. This warrants the development of an animal model for HIV-ZIKV co-infection. In this study, we used adult non-pregnant macaques that were chronically infected with simian immunodeficiency virus/chimeric simian human immunodeficiency virus (SIV/SHIV) and then inoculated with ZIKV. Plasma viral loads of both SIV/SHIV and ZIKV co-infected animals revealed no significant changes as compared to animals that were infected with ZIKV alone or as compared to SIV/SHIV infected animals prior to ZIKV inoculation. ZIKV tissue clearance of co-infected animals was similar to animals that were infected with ZIKV alone. Furthermore, in co-infected macaques, there was no statistically significant difference in plasma cytokines/chemokines levels as compared to prior to ZIKV inoculation. Collectively, these findings suggest that co-infection may not alter disease pathogenesis, thus warranting larger HIV-ZIKV epidemiological studies in order to validate these findings.
| The co-infection incidence of human immunodeficiency virus (HIV) infection and neglected tropical infectious diseases such as Zika virus (ZIKV) is on the rise due to the large geographical overlap of populations exposed to both of these viruses. Thus, research on such co-infection is of particular importance. In this study, we investigated SIV/SHIV-ZIKV co-infection dynamics in adult non-pregnant Rhesus Macaques (RMs) chronically infected with simian immunodeficiency virus (SIV)—or chimeric simian human immunodeficiency virus (SHIV). We found that post ZIKV inoculation, ZIKV plasma viral loads in co-infected macaques were similar to ZIKV alone-infected animals, and minimal changes were observed in cytokines/chemokines levels. Viral levels of SIV and SHIV also did not change as compared to pre-ZIKV inoculation levels. These findings thus suggest that co-infection may not alter disease pathogenesis of either HIV or ZIKV infections.
| Experimental and theoretical attention has been devoted to the interactions between human immunodeficiency virus (HIV) infection and various neglected tropical infectious diseases such as Zika virus (ZIKV), Dengue virus (DENV). These interactions could potentially lead to either pathogen altering the epidemiology, pathogenesis, immunology, and response to therapy of the other, sometimes even resulting in entirely new ailments that neither pathogen would have instigated alone [1]. In the last six decades since its discovery, Zika virus (ZIKV) has been considered a relatively mild human pathogen. Recently however, it has emerged as a threat to global health, demonstrating increased virulence, rapid spread and an association with grave neurological complications [2–4]. The two main types of clinical complications from ZIKV infection are microcephaly of newborns from women infected during early pregnancy [5], and a variety of neurological conditions in adults, including Guillain-Barré syndrome [1, 4] Serological tests cross-react DENV, and there are no specific antivirals or vaccines that are yet approved by Food and Drug Administration. Currently, the most effective tool for combating ZIKV is the prevention of mosquito bites, through measures such as repellents, protective nets, and insecticides [1].
The association between HIV infection and endemic diseases has been described in tropical regions with varying levels of complications. The first case of HIV-ZIKV co-infection was reported in Brazil without major health complications [6]. However, as the geographical range of ZIKV infection expands, exposed HIV immunosuppressed individuals may unveil new and more severe clinical manifestations, which must be anticipated. To the end, close surveillance of HIV-positive individuals to mirror such co-infections is of particular importance [1]. In this study, we investigated SIV/SHIV-ZIKV co-infection dynamics in a biologically relevant nonhuman adult non-pregnant primate model, with the objective of determining if and how ZIKV infection in HIV positive individuals may result in any potentially altered pathogenesis.
As described in Fig 1A, total of 6 adult female Indian-origin rhesus macaque (Macaca mulatta; age range 4.5 to 5 yrs) were chronically infected with SIVmac239 (n = 4) or SHIV3618MTF (n = 2) (newly developed clade C, T/F SHIV) [7] over a period of 6–7 months (Table 1). These animals were inoculated subcutaneously with 104 plaque forming unit (PFU) of ZIKV strain PRVABC59 (Table 1) and monitored for post ZIKV infection by viral loads and also evaluated for any clinical manifestation caused by ZIKV. All animal studies were conducted in accordance with UNMC IACUC approved protocols. Animal maintenance and procedures were carried out at the Department of Comparative Medicine, University of Nebraska Medical Center (UNMC) in accordance with the rules and regulations of the Committee on the Care and Use of Laboratory Animal Resources”. All protocols and procedures were performed under approval of the UNMC Institutional Animal Care and Use Committee according to the National Institute of Health guidelines.
The SIV plasma viral loads of all the chronically infected SIV macaques were stable (105−107 copies/ml) and did not change even after ZIKV inoculation (Fig 1B). Additionally, ZIKV plasma viral loads were found to peak at 105 copies/ml on 7 dpi (Fig 1C). ZIKV was detected in plasma samples up to 9 dpi in two of SIV-ZIKV co-infected animals; < 104 copies/ml for RMo15R and > 102 copies/ml for REd15R on 9 dpi. RGm15R that had the highest viral load, > 105 copies/ml on 7 dpi, and was found negative to ZIKV on 9 dpi (Fig 1C). Interestingly, although REd15R had the lowest peak of ZIKV viral load of > 102 copies/ml, ZIKV was still detected up to 15 dpi (Fig 1C). Furthermore, the SIV viral load status of these four SIV-ZIKV co-infected RMs was similar and quite stable during 51 dpi with ZIKV (Fig 1B).
In SHIV-ZIKV co-infected RMs, SHIV plasma viral loads were also found to be stable at 102−103 copies/ml during 20 dpi with ZIKV (Fig 1D). Additionally, ZIKV plasma viral loads were found to peak on 3 dpi to 103 copies/ml for R21612R and 104 copies/ml for RZi15R (Fig 1E). On 5 dpi, ZIKV was only detected at lower levels, < 103 copies/ml, in RZi15R. Following this, ZIKV was never again detected in either RMs R21612R or RZi15R (Fig 1E).
The viral loads of ZIKV in all six RMs were found to be negative after 20 days onwards. However, this delay of self-recovery of viremia appeared to be longer in co-infected animals as compared to animals that were infected with ZIKV alone (S1 and S2 Figs). However, the statistical analysis of ANOVA single (p value = 0.48), two-way without replication (p value = 0.42) and two-way with replication (p value = 0.51) of viral load of ZIKV did not reveal any significant differences between co-infected and exclusively ZIKV-infected RMs inoculated with 104 PFU of ZIKV PRVABC59. Thus, the low sample size of our study does not provide enough evidence to confirm the significance of our observed delay of self-recovery from ZIKV viremia (S1B Fig), which may require a larger group of animals in future studies. Clinical investigation of co-infected RMs did not reveal any severe symptoms and/or clinical signs of ZIKV infection with permanent sequela after acute phase (≤ 9 dpi). Additionally, at necropsy 6–7 months post-inoculation with ZIKV, various tissues including the brain stem, hippocampus, caudate, cerebellum, frontal cortex, spleen, mesenteric lymph node, uterus, liver, lung, and kidney, were collected and tested for ZIKV detection using a highly sensitive Droplet Digital PCR (ddPCR). We found that no detectable ZIKV viral RNA was present in any of the sampled tissues of three SIV-ZIKV co-infected RMs (RMo15R, RGm15R, and REd15R) and two of the SHIV infected RMs, suggesting that ZIKV infection had been cleared in these animals. This finding indicates the clearance of the ZIKV infection from the SIV-ZIKV co-infected adult RMs as similar to previously reported studies on RM that were infected with ZIKV alone [12, 13, 14].
Using Luminex methodology [11], cytokines and chemokines were measured from the plasma samples of three SIV-ZIKV co-infected RMs (RMo15R, RGm15R, and REd15R) on 0, 4, 7, 9, and 26 dpi with ZIKV. The results reveal that with the exception of MIF, IL-8, and SDF-1α, significant elevations of the plasma interleukin concentrations are evident in the acute phase of ZIKV especially at the peak on 7 dpi. Many of the cytokines and chemokines that were elevated in the acute phase of ZIKV displayed a tendency to return to normal levels in the later recovery phase (26 dpi) of ZIKV. Others however, such as BLC/CXCL13, Eotaxin, IP-10, MCP-1, and IFN-α, remained elevated during the peak of viremia (D4-D7) and also in recovery phase (D26). RANTES, I-TAC and, at lower levels, IL-1β, IL-6, IL-7, IFN-γ, SDF-1α, IL-1Rα, and GRO-α were elevated during acute phase and suppressed in recovery phase. Elevation of IL-1β and IFN-α in acute phase was noted. MIF and, at lower levels, IL-8 were suppressed during acute phase and elevated in recovery phase. The changes, either in the acute or in the recovery phase, were minor for MCP-1, IL-5, and IL-7 (Fig 2, S1 Table).
The main objective of this study was to examine the dynamics of HIV-ZIKV co-infection in order to evaluate how one pathogen may affect the pathogenesis of the other. We used rhesus macaques that were chronically infected with either SIVmac239 or SHIV3618MTF. These macaques were later inoculated with 104 PFU of ZIKV (PRVABC59) subcutaneously. In SIV-ZIKV co-infected RMs, Zika viral loads in plasma were found to be very similar to ZIKV infected animals drawn from both literature [10], and our own data [14]. Plasma viral loads of SIV and SHIV did not change as compared prior to ZIKV inoculation. These levels of viral load status of SIV and SHIV were also found to be very similar to chronically infected RMs drawn from both literature and our own data [15, 16].
Generally, mosquito-borne flaviviruses are initially detected in blood and lymphoid tissue of infected animals and subsequently invade peripheral organs and the central nervous system via the hematogenous route [2, 10]. In chronically infected SIV/SHIV RMs, ZIKV (PRVABC59) exhibited a similar pattern of viral kinetics as previously described in the ZIKV infected animals [10, 13]. The SIV/SHIV viremia kinetics in co-infected RMs were similar to those in the SIV/SHIV infected RMs from prior literature as well [15, 16]. Importantly, the rapid control of acute viremia of ZIKV infection that was observed in SIV/SHIV chronically infected RMs suggests that their peripheral immune system protects the host from peripheral ZIKV infection and that chronically infected SIV/SHIV RMs are similarly protected as non-immunocompromised RMs from infection by ZIKV.
All necropsies were performed 6–7 months after ZIKV infection, and ZIKV viral loads were measured in various tissues and organs using a highly sensitive Droplet Digital PCR and noted undetectable ZIKV viral loads. Prior studies had also revealed similar levels of clearance from tissue organs and body fluids for ZIKV (PRVABC59) infection in both humans [12] and RMs [13,14]. Human and animal model studies have demonstrated that ZIKV infection can result in persistence of infectious virus and viral nucleic acid in several body fluids (e.g., semen, saliva, tears, and urine) and target organs, including immune-privileged sites (e.g., eyes, brain, and testes) and the female genital tract [13, 17]. The ZIKV persistent or occult neurologic and lymphoid disease may occur following clearance of peripheral virus in ZIKV-infected individuals [12]. It has previously been demonstrated in infected RMs that ZIKV can persist in cerebrospinal fluid and lymph nodes for weeks after the virus has been cleared from peripheral blood, urine, and mucosal secretions [13]. The present adult RMs model confirmed that ZIKV persistent infection would also be cleared in immunocompromised RMs chronically infected with SIV. However, in this study we were unable to document the rate of any ZIKV clearance that occurred earlier than 6–7 months. ZIKV infection of rhesus and cynomolgus monkeys has been shown to recapitulate many key clinical findings, including rapid control of acute viremia, early invasion of the central nervous system, and prolonged viral shedding in adult animals [10, 13, 17]. Collectively, these findings enumerate that ZIKV infection in HIV infected individuals would not cause significant changes in pathogenesis and treatment plans.
Furthermore, besides the exceptions of MIF, IL-8, and SDF-1α, the cytokine/chemokine patterns of our RM were elevated during acute phase of ZIKV infection. Interestingly, this pattern is in accordance with previously described clinical findings for ZIKV-infected individuals [18, 19]. Major elevation was observed for IL-1β and IFN-α in acute phase ZIKV infection was observed in SIV-ZIKV co-infected RMs. The cytokine IL-1β is a key mediator of the inflammatory response and is essential for both host-response and resistance to pathogens. IFN-α is also mainly involved in innate immune responses against viral infection in acute phase of infection [11]. Together, these data confirm that ZIKV replication in the acute phase triggered rapid innate immune responses in peripheral blood. Several other chemoattractant chemokines (BLC/CXCL13, IP-10, MCP-1, RANTES, I-TAC, SDF-1α, and GRO-α) were also elevated during the ZIKV infection in chronically SIV/SHIV infected RMs described in this study, which is in accordance with previously described findings for ZIKV-infected individuals [19].
In this study IP-10 (CXCL10), has shown to be involved in ZIKV-related fetal neuron apoptosis or Guillain-Barré syndrome, and is suggested as potential biomarker of acute infection [19]. Specifically, these chemokines induce protective immunity against various viral infections including influenza, herpes simplex virus, Coxsackie virus, respiratory syncytial virus, and flaviviruses such as Dengue and West Nile viruses [18]. Elucidating the function of these chemokines in ZIKV infection, such as trafficking of lymphocytes into the various tissues, may reveal new mechanisms of immunological protection or immunopathology in SIV/SHIV-ZIKV co-infection. Additionally, although both B cell and T-cell recruitment to the sites of infection were highly triggered in SIV-ZIKV co-infected RMs in order to quickly control the infection, there were no significant changes between any of the cytokine levels measured in the acute versus the recovery phase.
Recent studies have revealed a significant increase in the number of HIV-ZIKV co-infected individuals reported in endemic areas in America [20–23] and Africa [24]. Additionally, the National Institute of Health (NIH) initiated an international, multisite clinical study, “Prospective Cohort Study of HIV and ZIKV in Infants and Pregnancy” (HIV ZIP; ClinicalTrials.gov, # NCT03263195) to register 2,000 pregnant women for the purpose of investigating ZIKV/HIV co-infection in the United States, Brazil, and Puerto Rico. Further studies suggest that although a high prevalence (72%) of ZIKV among 219 HIV-infected pregnant women was reported [20, 23]; their clinical presentation suggested a mild disease with a rapid and complete recovery. However, the fetuses of these women were often born with significant abnormalities, similar to those described previously in the children of women without HIV infection [25] and fetal demise would often occur [20]. In our current study, the cytokines/chemokines profile pattern following ZIKV infection SIV/SHIV models revealed a high level of similarity to the one described for HIV infected individuals [19]. Additionally, the notable similarity of ZIKV viral load pattern and clearance of ZIKV between SIV/SHIV models in this study and HIV infected individuals reported in these recent studies [20–25] highlights the utility of this co-infection model to understand disease pathogenesis.
In summary, we demonstrated that ZIKV viremia pattern in chronically infected SIV/SHIV RMs did not change significantly when compared to RMs that were infected with ZIKV alone, as well as to recent human epidemiological studies of HIV-ZIKV co-infection. These data suggest that ZIKV infection in chronically infected HIV individuals may not significantly alter the pathogenesis and disease progression of HIV or ZIKV, thus warranting larger epidemiological studies to validate these findings.
|
10.1371/journal.pgen.1002079 | Association of Genetic Variants in Complement Factor H and Factor H-Related Genes with Systemic Lupus Erythematosus Susceptibility | Systemic lupus erythematosus (SLE), a complex polygenic autoimmune disease, is associated with increased complement activation. Variants of genes encoding complement regulator factor H (CFH) and five CFH-related proteins (CFHR1-CFHR5) within the chromosome 1q32 locus linked to SLE, have been associated with multiple human diseases and may contribute to dysregulated complement activation predisposing to SLE. We assessed 60 SNPs covering the CFH-CFHRs region for association with SLE in 15,864 case-control subjects derived from four ethnic groups. Significant allelic associations with SLE were detected in European Americans (EA) and African Americans (AA), which could be attributed to an intronic CFH SNP (rs6677604, in intron 11, Pmeta = 6.6×10−8, OR = 1.18) and an intergenic SNP between CFHR1 and CFHR4 (rs16840639, Pmeta = 2.9×10−7, OR = 1.17) rather than to previously identified disease-associated CFH exonic SNPs, including I62V, Y402H, A474A, and D936E. In addition, allelic association of rs6677604 with SLE was subsequently confirmed in Asians (AS). Haplotype analysis revealed that the underlying causal variant, tagged by rs6677604 and rs16840639, was localized to a ∼146 kb block extending from intron 9 of CFH to downstream of CFHR1. Within this block, the deletion of CFHR3 and CFHR1 (CFHR3-1Δ), a likely causal variant measured using multiplex ligation-dependent probe amplification, was tagged by rs6677604 in EA and AS and rs16840639 in AA, respectively. Deduced from genotypic associations of tag SNPs in EA, AA, and AS, homozygous deletion of CFHR3-1Δ (Pmeta = 3.2×10−7, OR = 1.47) conferred a higher risk of SLE than heterozygous deletion (Pmeta = 3.5×10−4, OR = 1.14). These results suggested that the CFHR3-1Δ deletion within the SLE-associated block, but not the previously described exonic SNPs of CFH, might contribute to the development of SLE in EA, AA, and AS, providing new insights into the role of complement regulators in the pathogenesis of SLE.
| Systemic lupus erythematosus (SLE) is a complex autoimmune disease, associated with increased complement activation. Previous studies have provided evidence for the presence of SLE susceptibility gene(s) in the chromosome 1q31-32 locus. Within 1q32, genes encoding complement regulator factor H (CFH) and five CFH-related proteins (CFHR1-CFHR5) may contribute to the development of SLE, because genetic variants of these genes impair complement regulation and predispose to various human diseases. In this study, we tested association of genetic variants in the region containing CFH and CFHRs with SLE. We identified genetic variants predisposing to SLE in European American, African American, and Asian populations, which might be attributed to the deletion of CFHR3 and CFHR1 genes but not previously identified disease-associated exonic variants of CFH. This study provides the first evidence for consistent association between CFH/CFHRs and SLE across multi-ancestral SLE datasets, providing new insights into the role of complement regulators in the pathogenesis of SLE.
| SLE (OMIM 152700) is a debilitating autoimmune disease with strong genetic and environmental components, characterized by the production of autoantibodies resulting in tissue injury of multiple organs [1]. In SLE patients, aberrant complement activation leads to inflammatory injury [2], and fluctuation of serum C3 is a commonly used clinical biomarker of SLE disease activity [3]. In addition, a hereditary deficiency of C1q, C1r, C1s, C4 or C2 of the classical complement pathway impairs the clearance of immune complexes and debris from apoptotic cells, which strongly predisposes to SLE susceptibility [2]. Common variants of C3 and C4 have also been associated with risk of SLE [4], [5], [6]. Collectively, these findings indicate the important role of complement in the development of SLE.
Complement factor H (CFH), a key regulator of the alternative complement pathway, modulates the innate immune responses to microorganisms, controls C3 activation and prevents inflammatory injury to self tissue [7], [8]. CFH inhibits complement activation by preventing the formation and accelerating the decay of C3 convertase and acting as a cofactor for factor I-mediated degradation of C3b, both in plasma and on cell surfaces. Structurally, CFH contains 20 short consensus repeats (SCRs). SCR1-4 in the N-terminus mediate the cofactor/decay accelerating activity and SCR19-20 in the C-terminus are essential for cell surface regulation of CFH. In addition, CFH contains specific binding sites for polyanion (heparin or sialic acid), C-reactive protein (CRP) and microorganisms. CFH has five related proteins (CFHR1-5), all of which are also composed of SCRs [9]. SCRs in the N-terminus and C-terminus of CFHRs are highly homologous to SCR6-9 and SCR19-20 of CFH, respectively, suggesting that CFHRs and CFH may compete for binding to ligands. CFHRs lack SCRs homologous to SCR1-4 of CFH, and consequently do not exhibit cofactor/decay accelerating activity. Distinct from CFH, CFHR1 can inhibit C5 convertase activity and the formation of terminal membrane attack complex (MAC) [10]. A recent study has shown that CFH deficiency accelerates the development of lupus nephritis in lupus-prone mice MRL-lpr [11]. However, the role of CFHRs in the pathogenesis of SLE is still unknown.
CFH, CFHR3, CFHR1, CFHR4, CFHR2 and CFHR5, that present in tandem as a gene cluster located in human chromosome 1q32, are positional candidate genes within the 1q31-32 genomic region linked to SLE [12], [13]. In recent years, multiple exonic SNPs in CFH, such as I62V, Y402H, D936E and A473A, have been specifically associated with various human diseases including age-related macular degeneration (AMD) [14], [15], atypical hemolytic uremic syndrome (aHUS) [16] and membranoproliferative glomerulonephritis type II (MPGN II) [16], [17] as well as host susceptibility to meningococcal disease [18]. In addition, a common deletion of CFHR3 and CFHR1 (CFHR3-1Δ) has been associated with increased risk of aHUS [19] and decreased risk of AMD [20]. Taken together, these data prompted us to test whether genetic variants in CFH and CFHRs predisposed to SLE susceptibility.
Although recent genome wide association studies (GWAS) have b`n successfully used to identify SLE susceptibility genes [21], they still may be underpowered for specific genomic regions due to many factors such as sample size, marker density, ethnicity of subjects and over-stringent significance threshold. In these cases, a well-designed candidate gene-based association study can be used as a complementary approach to GWAS to identify genetic variants with modest effect size.
In this study, we fine mapped the CFH-CFHRs region using 60 SNPs and assessed their association with SLE susceptibility in a collection of 15,864 subjects (8,372 cases vs. 7,492 controls) from four ethnic groups. In addition, we assessed the association of CFHR3-1Δ with SLE by using tag SNPs.
To assess the association of CFH and CFHRs genes with SLE, we genotyped 60 tag SNPs covering the ∼360 kb CFH-CFHRs region in unrelated case-control subjects derived from four ethnic groups including European Americans (EA), African Americans (AA), Asians (AS), and Hispanics enriched for the Amerindian-European admixture (HS) (Figure 1A) (Table S1). According to the latest Hapmap CEU dataset (release 28), within the CFH-CFHRs region, 203 of 224 (90%) common SNPs (frequency>5%) could be captured by SNPs used in this study with r2>0.70. Within the most-studied gene CFH, previously identified disease-associated exonic SNPs including I62V (rs800292, typed), Y402H (tagged by rs7529589), D936E (tagged by rs10489456) and A474A (tagged by rs1410996) were evaluated for the association with SLE.
In the largest dataset (3,936 EA cases vs. 3,491 EA controls), after removing those failing the Hardy-Weinberg equilibrium (HWE) testing or showing low genotyping quality, fourteen SNPs were significantly associated with SLE (allelic P<0.05) (Table 1), of which rs6677604, located in intron 11 of CFH, exhibited the strongest association signal (minor allele frequency [MAF]: 23.0% in case vs. 20.1% in control, P = 2.4×10−5, OR[95%CI] = 1.19[1.10–1.28]). In the second largest dataset (1,679 AA cases vs. 1,934 AA controls), four SNPs were significantly associated with SLE (Table 1), all of which confirmed the association detected in EA, with rs16840639, located in the intergenic region between CFHR1 and CFHR4, showing the strongest association signal with a similar effect size (MAF: 37.5% vs. 33.7%, P = 6.6×10−4, OR[95%CI] = 1.18[1.07–1.31]). After Bonferroni correction for multiple comparisons, the association of rs6677604 and rs16840639 with SLE remained significant in both EA and AA (Table 1). However, in the two smaller datasets (1,265 AS cases vs. 1,260 AS controls and 1,492 HS cases vs. 807 HS controls), we failed to detect significant association of these SNPs with SLE (Table S1).
Of note, we did not detect significant association of I62V, Y402H and D936E with SLE in any of the four datasets (Table S1). A474A was associated with risk of SLE in EA (P = 0.015 before correction, OR[95%CI] = 1.09[1.02–1.16]), but it was not confirmed in the other three ethnic groups (Table S1).
To localize the underlying causal variant, we compared all SLE-associated SNPs (P<0.05) identified in EA and AA and carried out linkage equilibrium (LD) analysis. Fourteen SNPs, spanning from intron 6 of CFH to the 3′ region downstream of CFHR5, were associated with SLE in EA. However, only 4 of 14 SNPs, spanning from intron 6 of CFH to the 5′ region upstream of CFHR4, showed consistent association with SLE in AA, suggesting a smaller SLE risk region. Of interest, within the risk region, rs6677604 and rs16840639 exhibited the strongest association with SLE in EA and AA, respectively. We found that rs6677604 and rs16840639 were in strong LD with each other in both EA (r2 = 0.96) and AA (r2 = 0.77). Haplotype analysis showed that rs6677604 and rs16840639 could be defined into a ∼171 kb block in EA and a smaller ∼146 kb block in AA, respectively (Figure 1B). The minor allele of rs6677604 or rs16840639 perfectly tagged two SLE risk haplotypes in EA (H1: 16.1% vs. 14.1%, P = 0.0010; H2: 6.7% vs. 5.7%, P = 0.014), and the minor allele of rs16840639 perfectly tagged the only risk haplotype in AA (H1: 35.5% vs. 32.2%, P = 0.0028) (Figure 2).
Using the conditional haplotype-based association test, we showed that after conditioning on rs6677604 or rs16840639 significant associations of all other SNPs were eliminated in both EA and AA (Table 1), which suggested that rs6677604 and rs16840639 could account for all association signals in the CFH-CFHRs region. Due to the strong LD between rs6677604 and rs16840639, the conditional test could not be applied to further distinguish their association signals.
To compare between rs6677604 and rs16840639, we combined their ORs detected in EA and AA to generate a meta-analysis P value. The combined P value of rs6677604 (Pmeta = 6.6×10−8, OR[95%CI] = 1.18[1.11–1.26]) was stronger than that of rs16840639 (Pmeta = 2.9×10−7, OR[95%CI] = 1.17[1.10–1.2]).
Taken together, these data suggested that the underlying causal variant of SLE was captured by two strongly SLE-associated SNPs rs6677604 and rs16840639 in this study, which might reside in a ∼146 kb block. Neither rs6677604 nor rs16840639 are located in genomic regions with known biological function, which prompted us to seek other likely causal variants within the SLE-associated block.
CFHR3-1Δ is a likely functional variant within the ∼146 kb SLE-associated block (as shown in Figure 1A and 1B), which results in the deletion of CFHR3 and CFHR1 and has been associated with AMD and aHUS [19], [20]. Because co-segregation of the CFHR3-1Δ deletion with the minor allele of rs6677604 in subjects with European Ancestry was observed in a previous study of AMD [20], we hypothesized that the association of CFHR3-1Δ with SLE was captured by SNPs in this study. Using multiplex ligation-dependent probe amplification (MLPA) (location of MLPA markers were shown in Figure 1A), we genotyped CFHR3-1Δ in 275 EA, 106 AA, 282 AS and 196 HS subjects, and then measured its LD with rs6677604. We found that CFHR3-1Δ and rs6677604 were in complete LD in EA (r2 = 1.00) and AS (r2 = 1.00), strong LD in HS (r2 = 0.85) and moderate LD in AA subjects (r2 = 0.60) (Table 2). In a subset of 58 unrelated AA subjects who were genotyped at both rs6677604 and rs16840639, we found that CFHR3-1Δ was in stronger LD with rs16840639 (r2 = 0.70) than with rs6677604 (r2 = 0.60). These results indicated that the association of the CFHR3-1Δ deletion with risk of SLE was tagged by the minor allele of rs6677604 in EA and rs16840639 in AA, respectively, suggesting that CFHR3-1Δ might be a risk variant for SLE.
We showed that rs6677604 and CFHR3-1Δ were in the same block in AS (Figure 1B), and the minor allele of rs6677604 could perfectly tag the CFHR3-1Δ deletion (r2 = 1.00). Thus, the lack of significant association of rs6677604 with SLE in our previous AS dataset might be due to insufficient statistical power. To increase power, we further genotyped 787 Chinese SLE cases and 1065 Chinese controls and then assessed the association of rs6677604 with SLE in an enlarged AS dataset (2052 cases vs. 2325 controls). In the enlarged AS dataset, we detected the significant association of rs6677604 with SLE (MAF: 7.1% vs. 6.1%, P = 0.0485, OR[95%CI] = 1.19[1.00–1.40]), supporting the hypothesis that CFHR3-1Δ might also be a risk variant for SLE in the AS population.
To test whether homozygous deletion of CFHR3-1Δ might confer a higher risk of SLE than heterozygous deletion, we compared the genotypic frequency of homozygous and heterozygous deletion to that of no deletion, respectively. In EA, using rs6677604 as a tag SNP, we found that the homozygous deletion of CFHR3-1Δ conferred a significantly increased risk of SLE (P = 7.5×10−4, OR[95%CI] = 1.47[1.17–1.84]) compared to no deletion, which was stronger than that of the heterozygous deletion (P = 0.0018, OR[95%CI] = 1.17[1.06–1.29]) (Table 3), suggesting a dosage dependent risk effect of the CFHR3-1Δ deletion. To confirm, we compared genotypic associations of CFHR3-1Δ in AS and AA using rs6677604 and rs16840639 as tag SNPs, respectively. In these two ethnic groups, we found that only homozygous deletion of CFHR3-1Δ conferred a significantly increased risk of SLE compared to no deletion (AS: P = 0.0021, OR[95%CI] = 3.30[1.47–7.41]; AA: P = 0.0011, OR[95%CI] = 1.40[1.14–1.71]) (Table 3), supporting the hypothesis that homozygous deletion of CFHR3-1Δ conferred a higher risk of SLE than heterozygous deletion. In a meta-analysis combining ORs of EA, AA and AS, we confirmed that the homozygous deletion of CFHR3-1Δ (Pmeta = 3.2×10−7, OR[95%CI] = 1.47[1.27–1.71]) had a stronger association with risk of SLE than the heterozygous deletion (Pmeta = 3.5×10−4, OR[95%CI] = 1.14[1.06–1.23]).
SLE is a complex disease with heterogeneous sub-phenotypes. To determine whether CFHR3-1Δ had a stronger association with specific clinical manifestations of SLE, we compared its frequency in SLE cases stratified by the presence or absence of each of the eleven ACR classification criteria (malar rash, discoid rash, photosensitivity, oral ulcers, arthritis, serositis, renal disorder, neurologic disorder, hematologic disorder, immunologic disorder and antinuclear antibody) and five autoantibodies (anti-dsDNA, anti-Sm, anti-RNP, anti-SSA/Ro and anti-SSB/La). In EA, we found that tag SNP rs6677604 of CFHR3-1Δ was associated with the absence of neurologic disorder (Table S2). However, in AA, we found that the corresponding tag SNP rs16840639 was associated with the absence of anti-dsDNA and the presence of serositis (Table S2), the latter of which was found not to be significant after Bonferroni correction for multiple comparisons. Insufficient clinical information for the majority of AS SLE patients precluded us from conducting these analyses. Taken together, these data did not provide evidence for a stronger association of CFHR3-1Δ with specific clinical manifestations of SLE.
In this study, we identified SLE-associated SNPs in the CFH-CFHRs region in three ethnic groups consisting of EA, AA and AS. In addition, we showed that the underlying causal variant was captured by rs6677104 and rs16840639 and could be localized to a ∼146 kb block extending from intron 9 of CFH to the 5′ region upstream of CFHR4. We demonstrated that the CFHR3-1Δ deletion, which has been associated with AMD and aHUS, could be tagged by the minor risk alleles of rs6677604 (r2 = 1.00 in EA and AS) and rs16840639 (r2 = 0.70 in AA) and showed dosage-dependent association with risk of SLE. These data strongly suggested that CFHR3-1Δ, which leads to reduced levels of CFHR3 and CFHR1 proteins, was the causal variant for increased risk of SLE within the SLE-associated block.
Multiple CFH exonic SNPs have been associated with various human diseases, but none of them were associated with SLE in this study. Y402H (rs1061170) is the most studied non-synonymous SNP of CFH. Y402H is located in SCR7 and affects the binding of CFH with glycosaminoglycans and CRP [22], [23], [24]. Y402H has been strongly associated with risk of AMD and MGPN2 but not associated with aHUS [16]. In this study, we genotyped a tag SNP of Y402H (rs7529589, r2 = 0.75 with Y402H according to HapMap CEU data) and detected no statistically significant association with SLE (Table S1). In a previous study, we had genotyped Y402H directly in 2033 EA cases and 2824 EA controls, and observed a similar result (37.4% vs. 37.7%, P = 0.81, OR = 0.99). I62V (rs800292) located in the N-terminal SCR2 is another well-studied non-synonymous SNP of CFH. Although I62V may result in increased binding of CFH with C3b and enhanced CFH co-factor activity and has been associated with decreased risk of AMD, MPGN II and aHUS [16], [25], it was not associated with SLE in this study (Table S1). D936E (rs1065489 in SCR16) was associated with lower host susceptibility to meningococcal disease in a recent GWAS [18]. We genotyped a perfect tag SNP (rs10489456) of D936E and failed to detect an association with SLE (Table S1). A synonymous SNP A474A (rs2274700 in SCR8) and its tag SNP rs1410996 were strongly associated with risk of AMD independent of Y402H [26], [27], but we detected only a marginal association between rs1410996 and risk of SLE in EA (Table 1), which was eliminated after conditioning on rs6677604 or rs16840639. In addition, two synonymous SNPs A307A (rs1061147 in SCR5) and Q672Q (rs3753396 in SCR13) that are in strong LD with Y402H and D936E, respectively, were not associated with SLE in our study. These data suggest that the previously described disease-associated CFH exonic SNPs do not contribute to the development of SLE.
Compared with SNP genotyping assays, genotyping assays for copy number variation are more labor-intensive and costly. Consequently, CFHR3-1Δ was not specifically genotyped in this study to assess its association with SLE. Instead, we evaluated the effect of the CFHR3-1Δ deletion on SLE development indirectly using tag SNPs that were in strong LD with it. We first confirmed that CFHR3-1Δ was in strong LD with rs6677604 in EA, similar to previous studies of AMD [20], [28]. Furthermore, we showed that CFHR3-1Δ was also in strong LD with rs6677604 in AS and HS. In addition, we found that CFHR3-1Δ was in stronger LD with rs16840639 than with rs6677604 in AA. Of note, in AA, the most significant association with SLE was detected at rs16840639 rather than rs6677604, and the risk haplotype H1 in AA was perfectly tagged by the minor allele of rs16840639 rather than rs6677604 (Figure 2), suggesting that rs16840639 captured the underlying causal variant CFHR3-1Δ in AA. Using these tag SNPs, we deduced that homozygous CFHR3-1Δ deletion conferred higher risk of SLE than heterozygous deletion, which suggested a change in gene dosage of the encoded proteins CFHR3 and CFHR1 might account for the increased SLE risk.
The CFHR3-1Δ deletion was associated with the general phenotype of SLE but did not consistently exhibit stronger signals to a specific clinical manifestation in EA and AA, and was not specifically associated with the presence of renal disorder. This is in contrast to the effect of CFH deficiency, which results in the development of glomerulonephritis in CFH knockout mice due to uncontrolled C3 activation [11], [29]. In addition, the absence of CFH in plasma causes human MPGN II [30], but an association of the CFHR3-1Δ deletion with MPGN II has not been reported. The absence of an association of CFHR3-1Δ with renal disorder in lupus suggests that CFHR3 and CFHR1 play a different role from CFH in the pathogenesis of lupus, although further studies are required to validate the lack of association between the CFHR3-1Δ deletion and renal disorder in SLE.
The CFHR3-1Δ deletion has opposite effects in different diseases [9], and the underlying mechanism is poorly understood. Activated complement pathways converge to generate C5 convertase, which cleaves C5 into C5a and C5b. C5a is a potent chemoattractant. C5b initiates the formation of the terminal MAC. CFHR1 acts as a complement regulator to inhibit C5 convertase activity and terminal MAC formation [10], and CFHR3 displays anti-inflammatory effects by blocking C5a generation and C5a-mediated chemoattraction of neutrophils [31]. Increased neutrophils lead to inflammatory injuries in many non-infectious human diseases [32]. It has been shown that immune complex-induced inflammatory injuries are largely mediated by C5a receptor and blocking C5a receptor reduces manifestation of lupus nephritis in mice [33], [34]. In addition, increased apoptotic neutrophils contribute to autoantigen excess and have been associated with increased disease activity in SLE [35]. The CFHR3-1Δ deletion results in decreased CFHR3 and CFHR1 levels and may therefore lead to uncontrolled production of chemoattractant C5a predisposing to SLE. Of interest, the CFHR3-1Δ deletion also has a risk effect in aHUS and the CFHR3 and CFHR1 deficiency in plasma has been associated with the presence of anti-CFH autoantibodies, which bind to the C-terminus of CFH and block CFH binding to cell surfaces [36], [37]. It is also possible that CFHR3-1Δ is also associated with the presence of anti-CFH autoantibodies in SLE and thus leads to impaired CFH cell surface regulation.
Both CFHR3 and CFHR1, lacking the CFH N-terminus regulatory activity, were reported to compete with CFH for binding to C3b, and thus CFHR3 and CFHR1 deficiency may lead to enhanced CFH regulation [31], which may explain the protective effect of the CFHR3-1Δ deletion in AMD. Of interest, as mentioned before, the non-synonymous SNP I62V in the CFH regulatory domain may also increase CFH regulation. I62V confers a protective effect in AMD, aHUS and MPGN II [16], but it was not associated with SLE in this study.
Statistical under-powering might account for the failure to detect a significant association in HS dataset. First, rs6677604 and CFHR3-1Δ were in strong LD and could be defined into a block in HS, which excluded the possibility that the CFHR3-1Δ deletion was not tagged in the HS dataset. In addition, there was no genetic heterogeneity of rs6677604 in the four ethnic groups (P = 0.76), in which the risk minor allele showed consistently higher frequency in cases than in controls. Finally, based on rs6677604, post hoc analysis indicated a much lower power of 51% in HS to detect association with SLE (P<0.05) than the power of 98% in EA and 92% in AA. Thus, the association of CFHR3-1Δ with SLE in HS needs to be further evaluated in a larger dataset.
One limitation of this study is that we have not addressed whether rare variants in the CFH-CFHRs region may contribute to the development of SLE. Pathogenic rare variants clustering in CFH C-terminus affect CFH cell surface binding, but they were only found in aHUS patients, not in AMD, MPGN II patients and healthy controls [16]. Deep sequencing of exons in CFH C-terminus in patients with SLE may elucidate whether these rare variants are associated with SLE.
To our knowledge, this study is the first to show that genetic variants in the CFH-CFHRs region are associated with SLE susceptibility. Our consistent observations of dose-dependent association between CFHR3-1Δ and SLE across three distinct ancestral populations and no association in CFH exonic SNPs suggest a novel role for CFHR3 and CFHR1 in the pathogenesis of SLE. Further functional studies are required to elucidate the underlying mechanism of CFHR3-1Δ.
The study was approved by the Human Subject Institutional Review Boards or the ethnic committees of each institution. All subjects were enrolled after informed consent had been obtained.
To test the association of CFH and CFHRs with SLE, we used a large collection of samples from case-control subjects from multiple ethnic groups. These samples were from the collaborative Large Lupus Association Study 2 (LLAS2) and were contributed by participating institutions in the United States, Asia and Europe. According to genetic ancestry, subjects were grouped into four ethnic groups including European American (3,936 cases vs. 3,491 controls), African American (1,679 cases vs. 1,934 controls), Asian (1,265 cases vs. 1,260 controls) and Hispanic enriched for the Amerindian-European admixture (1,492 cases vs. 807 controls). Asians were comprised of Koreans (884 cases vs. 994 controls), Chinese (200 cases vs. 205 controls) and subjects from other East Asian countries such as Japan and Singapore (181 cases vs. 61 controls). African Americans included 275 Gullahs (152 cases vs. 123 controls), who are subjects with African Ancestry.
To test LD between CFHR3-1Δ and SLE-associated SNPs, we used 275 unrelated European Americans (187 cases vs. 88 controls), 106 African Americans (88 unrelated subjects [58 cases vs. 30 controls] and 18 subjects from 6 SLE trios families), 282 unrelated Chinese (218 cases vs. 64 controls) and 196 Hispanics (157 unrelated subjects [91 cases vs. 66 controls] and 39 subjects from 13 SLE trios families). All of these subjects were enrolled from UCLA.
To enlarge the sample size of Asians for association test, we used 1,852 Chinese case-control subjects (787 vs. 1065) recruited from Shanghai Renji Hospital, Shanghai Jiao Tong University School of Medicine.
All SLE patients met the American College of Rheumatology (ACR) criteria for the classification of SLE [38].
LLAS2 samples were processed at the Lupus Genetics Studies Unit of the Oklahoma Medical Research Foundation (OMRF). SNP genotyping was carried out on the Illumina iSelect platform. Subjects with individual genotyping call rate <0.90 were removed because of low data quality. Subjects that were duplicated or first degree related were also removed. Both principal component analysis and global ancestry estimation based on 347 ancestry informative markers were used to detect population stratification and admixture, as described in another LLAS2 report [39]. After removing genetic outliers, a final dataset of 15,864 unrelated subjects (8,372 cases vs. 7,492 controls) was obtained.
Taqman SNP genotyping assay (Applied Biosystems, California, USA) was used to genotype rs6677604 for subjects who were not recruited into LLAS2.
MLPA kit “SALSA MLPA KIT P236-A1 ARMD mix-1” was used to genotype the CFH-CFHRs region according to the manufacture's instruction (MRC-Holland, Amsterdam, The Netherlands). ABI 3730 Genetic Analyzer (Applied Biosystems) was used to run gel electrophoresis. Software Peak Scanner v1.0 (Applied Biosystems) was used to extract peaks generated in electrophoresis. Coffalyser v9.4 (MRC-Holland) was used to readout copy number of target region.
The HWE test threshold was set at P>0.01 for controls and P>0.0001 for cases. SNPs failing the HWE test were excluded from association test. SNPs showing genotyping missing rate>5% or showing significantly different genotyping missing rate between cases and controls (missing rate>2% and Pmissing<0.05) were also excluded from association test. In allelic association test (Pearson's χ2–test), the significance level was set at P<0.05. Haploview 4.2 was used to estimate pairwise LD values between SNPs, define haplotypes blocks and calculate haplotypic association with SLE. Haplotype-based conditional association analysis was carried out by Plink v1.07. Mantel-Haenszel analysis was performed to generate the meta-analysis P value. CaTS was used to calculate statistical power.
|
10.1371/journal.pntd.0006668 | Scabies and risk of skin sores in remote Australian Aboriginal communities: A self-controlled case series study | Skin sores caused by Group A streptococcus (GAS) infection are a major public health problem in remote Aboriginal communities. Skin sores are often associated with scabies, which is evident in scabies intervention programs where a significant reduction of skin sores is seen after focusing solely on scabies control. Our study quantifies the strength of association between skin sores and scabies among Aboriginal children from the East Arnhem region in the Northern Territory.
Pre-existing datasets from three published studies, which were conducted as part of the East Arnhem Healthy Skin Project (EAHSP), were analysed. Aboriginal children were followed from birth up to 4.5 years of age. Self-controlled case series design was used to determine the risks, within individuals, of developing skin sores when infected with scabies versus when there was no scabies infection. Participants were 11.9 times more likely to develop skin sores when infected with scabies compared with times when no scabies infection was evident (Incidence Rate Ratio (IRR) 11.9; 95% CI 10.3–13.7; p<0.001), and this was similar across the five Aboriginal communities. Children had lower risk of developing skin sores at age ≤1 year compared to at age >1 year (IRR 0.8; 95% CI 0.7–0.9).
The association between scabies and skin sores is highly significant and indicates a causal relationship. The public health importance of scabies in northern Australia is underappreciated and a concerted approach is required to recognise and eliminate scabies as an important precursor of skin sores.
| Skin sores, also known as impetigo, are highly contagious bacterial skin infections, which are found commonly in school children and occasionally in adults. Skin sores are prevalent in disadvantaged or resource-poor settings. In Australia, about two thirds of Aboriginal children suffer from skin sores by their first birthday. If untreated or treated poorly, skin sores can eventually cause heart and kidney problems. It is also believed that scabies, another common skin infection in Aboriginal children, can increase the risk of developing skin sores by allowing the bacteria to enter the skin more easily through breaks in the skin. Our research explored the following: if scabies is a risk factor for skin sores then what is the strength of the association between the two conditions.
| Remote Aboriginal communities in northern Australia have the world’s highest prevalence of skin sores with more than 80% of children affected by their first birthday[1]. Skin sores, also known as impetigo, are commonly caused by Group A Streptococcus (GAS) infections in these populations and can have serious sequelae such as invasive bacterial infection and post-streptococcal glomerulonephritis, which in turn increases the risk of chronic renal disease [2].
Acquisition of skin sores is influenced by other skin infections, particularly scabies. Scabies is endemic in remote northern Australia and found in up to 35% of children and 25% of adults in the region [3]. Scabies infection often leads to a secondary GAS infection of the skin and scabies control is considered a priority in measures aimed at reducing skin sores in the Aboriginal population [4–7].
Kearns et al. found that skin sores were seven times more likely to be concurrently diagnosed with scabies than when there was no scabies diagnosis[1]. The risk ratio was calculated based on the diagnosis of skin sore infection at a presentation of scabies compared to no diagnosis of scabies at the same presentation. However, their study focused on the concurrent risk between scabies and skin sores and further research is needed to understand the temporal relationship between the two conditions.
Our study employed a self-controlled case series method to quantify the risk of scabies on skin sores using historical data from observational studies in Northern Territory. The study findings will contribute to the ongoing mathematical modelling research on understanding of GAS transmission and inform policy makers in prevention and control of skin sores and scabies in remote Aboriginal communities.
We analysed pre-existing datasets from three published studies, which were conducted as part of the East Arnhem Healthy Skin Project (EAHSP) overseen by the Menzies School of Health Research [1, 8, 9]. The EAHSP was a regional collaboration to reduce skin infections among children in five remote Aboriginal communities in the East Arnhem region in the Northern Territory. The three studies, namely Kearns et al.[1], McMeniman et al.[9], and Clucas et al.[8], were separate but overlapping cohorts in the communities where EAHSP was conducted. These studies employed the same methodology to retrospectively review medical records from community health clinics in the region.
Despite their similar research methodology, the three studies varied in inclusion criteria and duration of follow-up (Table 1). In Kearns et al., eligible participants needed to have at least one clinic presentation in each quarter of their first year of life [1]. The duration of follow-up differed among the studies: participants were followed from birth to 1 year in Kearns et al., to 2-years in McMeniman et al., and up to 5-years of age in Clucas et al. One caveat of the Clucas et al. study is that although children born after 1 January 2001 were included, presentations were followed only from 1 January 2002, and therefore not all participants were followed from birth (28%) [8]. Complete descriptions of the research methods and study population for these studies are available elsewhere [1, 8, 9].
Data collected included date of presentation, child’s height and weight, any reason for presentation, antibiotic use, and referrals to hospital. Scabies were recorded as reasons for presentation if they were either noted specifically or with reference to scabies treatment given, while skin sores were recorded if there was any mention of skin sores or other presumed bacterial infections of the skin including boils, carbuncles, abscesses, ulcers and pustules (Clucas). Multiple presentations on the same day were recorded as a single event. We merged the datasets from three published studies into a single composite line-listed dataset. All data analysed were de-identified. Ethics approval was obtained from the Human Research Ethics Committee of the Northern Territory Department of Health and Families and the Menzies School of Health Research (Ethics approval 2015–2516).
Data on infection-free children i.e., who developed neither scabies nor skin sores were not available, in the three pooled studies, to determine the difference between the exposed and non-exposed groups using cohort methods. Therefore, we used self-controlled case series (SCCS) method as an alternative to cohort studies by comparing the risks during different time periods within individuals[10].
The SCCS method compares the risks of developing events in the periods following exposure versus non-exposed periods within individuals. This method controls for fixed (i.e. time-independent) confounders such as gender and ethnicity, and only requires cases for analysis [11, 12].
Our study included children from five Aboriginal communities who attended the community health clinics from January 2001 to February 2007. We applied the following exclusions sequentially to our study cohort:
The observation period ran from birth to the earliest of (i) the first time the observation period went for more than 120 days between clinic visits; (ii) the end of the study follow-up period; or (iii) the maximum age for their respective study (Table 1). The observation period was further divided in two age groups: ≤1 year and >1 year (up to 4 years old) to account the effect of age on both exposure and outcome. Scabies and skin sores episodes were identified from the clinic records. We assessed, within individuals, the risk of developing skin sores following scabies infection (exposed period) versus the risk of skin sores infection when there is no scabies (non-exposed period).
The period with the risk of developing skin sores was further subdivided into two segments: pre-exposure period and exposed period. The pre-exposure period is the period before the diagnosis of scabies (the time taken from onset of scabies symptoms to the diagnosis of scabies by the clinician) while the exposed period is the day of the diagnosis of scabies plus a specified period following the scabies diagnosis (Fig 1). The pre-exposure period is included to account for the presence of scabies prior to clinic attendance and any delay in the diagnosis of scabies.
Our baseline assumption was that the pre-exposure period was seven days and exposed period was 14 days. These periods were defined based on the first diagnosis of scabies and any subsequent episodes that fell within the same risk period, were counted as one infection. Skin sores episodes were then mapped in relation to different time periods. A seven-day duration was chosen for the pre-exposure period based on the shortest time period reported in the literature [13, 14] and the very frequent presentation of our participants for medical attendance (on average every two weeks) [1, 8, 9].
The natural history of skin sores is that an uncomplicated impetigo heals spontaneously within two weeks [15, 16]. Detailed studies of the natural history of impetigo primarily caused by S. pyogenes found that the mean time to spontaneous healing was 12.6 days, with a range of 6–31 days [17]. These data and assumptions may not be generalisable to impetigo in non-endemic settings where Staphylococcus aureus is the primary pathogen. In our study, we assumed that repeat clinic presentations, positive for skin sores within a two-week period, were the same infectious episode and only the first episode was included in the analysis.
We transformed the data into a time series format and applied deterministic imputation to substitute missing data as follows. The imputation was performed in a stepwise approach. Firstly, we found the first event (skin sores or scabies) in an individual and carried the value forward for the following 14 days. This step was repeated for any recurring events within the same individual. After that, any missing data were assumed as non-events. This was to ensure that skin sores and scabies events within 14 days were counted as one episode.
Our primary analysis involved estimating the relative incidence rate of skin sores during exposed periods compared to the non-exposed period. Conditional Poisson regression was used to calculate incidence rate ratios (IRRs) and 95% confidence intervals (CIs). Analyses were undertaken using Stata/IC version 14 (StataCorp. College Station, TX, US). We converted the time periods from years to days for higher granularity and better accuracy. We undertook sensitivity analyses by increasing the exposed period from 14 to 21 and 28 days while keeping the pre-exposure interval consistent. We further analysed the incidence rate ratios within each community.
Among the initial study population of 417 children, 55.9% (n = 233) were male. The number of clinic attendances ranged from 1 to 117 with median attendance of 19 times per individual during the six-year follow-up period from January 2001 to February 2007. Approximately 75% of children had their first episode of skin sores before their first birthday.
After a series of exclusions, we arrived at a study population of 307 children (Fig 2). Censoring the observation period for each child once they went more than 120 days between clinic visits dropped skin sores episodes beyond the censored date, resulting in more children without a skin sore episode. Therefore, an additional 16 children were excluded resulting in a total of 291 children in the final analysis (Fig 2).
Compared to the non-exposed periods, the overall rate of skin sores was increased in both pre-exposure and exposed periods, with a significant increase in the exposed periods. Although the pre-exposure periods were associated with a higher prevalence of skin sores than baseline non-exposed periods (IRR 1.3; 95% CI 0.8–1.9), this finding was not statistically significant (p = 0.296). However, children were 11.9 times more likely to develop skin sores during the exposed periods compared with the baseline non-exposed periods (IRR 11.9; 95% CI 10.3–13.7; p<0.001) (Table 2). When exposed to scabies, children had lower risk of developing skin sores at age ≤1 year compared to at age >1 year (IRR 0.8; 95% CI 0.7–0.9).
Sensitivity analysis on the duration of the exposure period (14 days, 21 days and 28 days) showed that increasing the length of the exposed period resulted in a progressive decrease in the incidence rate ratio for skin sores associated with scabies, compared with the baseline (Table 2). However, the study outcomes for pre-exposure periods and age groups did not change significantly (Table 2). Community stratified rates in five communities (using the primary analysis assumption of 14-day exposure period) showed similar results to the overall rates in both pre-exposure and exposed periods, with overlapping confidence intervals (Table 3).
Our study is the first to use the self-controlled case series method to investigate the association between scabies and skin sores. We found that, when infected with scabies, children were 12 times more likely to develop skin sores than in the absence of scabies infestation. Our findings reiterate the extreme burden of skin disease in remote Aboriginal and Torres Strait Islander communities, with up to 75% of children in these communities having a skin sore by their first birthday. This force of infection is compatible with data from low and middle-income regions such as India and those in the South Pacific where the prevalence of skin sores among preschool children and adolescents ranges from 42% to 70% [13].
The risk estimate from our study is substantially higher than previously reported relative risks of scabies and skin sores co-infections, which ranged from 2.4 to 7.0 [1, 3, 13, 18]. There are two possible explanations. Firstly, earlier studies analysed risk based on concurrent presentation with scabies and skin sores—our study extended the hypothesised association to a temporal relationship between diagnosed scabies and skin sores, within a defined risk window. Secondly, the SCCS method has the advantage of controlling implicitly for fixed confounders, which can affect case-control and cohort studies, and would be anticipated to report a more faithful estimate of true risk [19].
Whilst our study cannot establish causality definitively, it has a number of attributes supporting a causal relationship between scabies and skin sores. These include the significant association and the temporality between the two conditions and consistency with earlier studies. These results are complemented by the sensitivity analyses on the duration of the post diagnosis scabies exposure period. We found that lengthening the exposure window from 14 to 21 days and then to 28 days was associated with a decline in the incidence rate ratio. This strengthens inference on the purported causal relationship between scabies and skin sores, i.e., if scabies truly increases the risk of skin sores, it is logical that the risk of this outcome will occur within a time frame more narrowly associated with the diagnosis (i.e. 14 days rather than 28 days).
One of the strengths of our study is inclusion of pre-exposure period in the analysis so as to reduce the inflation of relative risks in the exposed period. We observed that the risk of developing skin sores in pre-exposure periods is much lower than that in exposed periods. However, the results for pre-exposure periods are not statistically significant, and furthermore, these findings are probably due to clinic nonattendance rather than actual absence of infections.
We found that children were less likely to develop skin sores in their infanthood than when they were between one and four years of age. This is consistent across different durations of scabies infection. One possibility is that infants are less likely to scratch and develop excoriation of the skin when infected with scabies than older children, thereby reducing the incidence of skin sores. However, it is important to note that the overall association between skin sores and scabies remains significant regardless of age.
Our finding is of a link at the individual level between scabies and skin sore risk. Studies of skin sores in the East Arnhem region have identified GAS as the dominant pathogen in around 80% of cases [2], and clearance of GAS has been identified as the only independent predictor of complete resolution of skin sores [20]. It is therefore likely that the observed association between scabies and skin sores reflects a heightened risk of streptococcal infection in the presence of scabies, but we can neither confirm nor refute a similar association with staphylococcal infection on the basis of the evidence here presented. The findings are unlikely to be generalisable to non-endemic settings for impetigo where S. aureus is usually the primary pathogen.
Our study limitations include that, in the SCCS design, occurrence of an event should not affect subsequent exposures. Although research is limited in the area of scabies and skin sores, to date, there is no evidence suggesting the impact of skin sores on the risk of developing scabies. Therefore, we assumed that a skin sore infection was unlikely to affect subsequent scabies exposure. However, our exposures and events are defined by diagnosed cases and may be influenced by clinic attendances. For example, if a diagnosis of skin sores prompted a participant to return to the clinic for clearance check-up, these follow-up visits may increase the chance of scabies being detected incidentally.
Secondly, our study has likely underestimated the incidence of scabies as we only included the children who presented to the clinics. Furthermore, we did not control for any potential time varying confounders, which can cause delay in diagnosis such as stigma associated with scabies, restricted access to health care and chronic steroid use causing masked presentation[21–23].
Thirdly, our study relies on the documentation and clinical diagnosis of the clinicians. While there is imprecision involved in relying on syndromic reporting of skin sores and clinically observed scabies identified from an historical clinic record review, this level of diagnostic uncertainty reflects the reality of clinical practice in a remote setting. We anticipate that the high prevalence of both conditions in the communities studied should increase clinical experience and the consistency and reliability of observer reporting. Indeed, the involved communities and associated healthcare clinics have had a long standing interest in skin health related research, and thus have well engaged clinicians around the diagnosis and management of impetigo and scabies. Moreover, our use of the self-controlled case series method accounts for any variability between participants.
Lastly, whilst the SCCS method provides a robust study design, we may not be able to generalise the findings conclusively. Our findings are based on a specific subgroup of Aboriginal children and may not be representative of the wider Australian population or older age groups nor perhaps of Aboriginal children living in other remote communities. Furthermore, SCCS method was first developed for rare events and historically used in vaccine trials, yet we assessed two common conditions in Aboriginal communities. However, others have utilised the SCCS method for common events such as drug safety and have produced findings consistent with those of previous studies, which suggests the SCCS method was appropriate for our study[24, 25].
The self-controlled case series method we have used is potentially transferrable to other settings where data have been collected that do not invalidate the necessary conditions. By investigating associations within an individual, the self-controlled case series method implicitly controls for contextual factors that may influence the risk of either infection, making findings more comparable across situations and settings. As regards to public health implications, our findings robustly support existing inference on the contribution of scabies to skin sore risk, particularly in the very young [13, 14]. Accordingly, we endorse holistic healthy skin strategies that raise awareness of the need for prevention, early presentation and multi-pronged treatment strategies to reduce the overall burden and the long-term sequelae of skin disease.
In conclusion, our study demonstrated that the association between scabies and skin sores is significant, more so than previously reported in standard cohort and case- controlled studies. A concerted approach is needed in implementing scabies elimination programs to prevent skin sores, particularly due to GAS, and their devastating complications in remote Aboriginal communities.
|
10.1371/journal.ppat.1003773 | Type I and Type III Interferons Drive Redundant Amplification Loops to Induce a Transcriptional Signature in Influenza-Infected Airway Epithelia | Interferons (IFNs) are a group of cytokines with a well-established antiviral function. They can be induced by viral infection, are secreted and bind to specific receptors on the same or neighbouring cells to activate the expression of hundreds of IFN stimulated genes (ISGs) with antiviral function. Type I IFN has been known for more than half a century. However, more recently, type III IFN (IFNλ, IL-28/29) was shown to play a similar role and to be particularly important at epithelial surfaces. Here we show that airway epithelia, the primary target of influenza A virus, produce both IFN I and III upon infection, and that induction of both depends on the RIG-I/MAVS pathway. While IRF3 is generally regarded as the transcription factor required for initiation of IFN transcription and the so-called “priming loop”, we find that IRF3 deficiency has little impact on IFN expression. In contrast, lack of IRF7 reduced IFN production significantly, and only IRF3−/−IRF7−/− double deficiency completely abolished it. The transcriptional response to influenza infection was largely dependent on IFNs, as it was reduced to a few upregulated genes in epithelia lacking receptors for both type I and III IFN (IFNAR1−/−IL-28Rα−/−). Wild-type epithelia and epithelia deficient in either the type I IFN receptor or the type III IFN receptor exhibit similar transcriptional profiles in response to virus, indicating that none of the induced genes depends selectively on only one IFN system. In chimeric mice, the lack of both IFN I and III signalling in the stromal compartment alone significantly increased the susceptibility to influenza infection. In conclusion, virus infection of airway epithelia induces, via a RIG-I/MAVS/IRF7 dependent pathway, both type I and III IFNs which drive two completely overlapping and redundant amplification loops to upregulate ISGs and protect from influenza infection.
| The response of cells to virus infection depends on Interferons (IFNs), a group of cytokines which activate the expression of hundreds of genes that help control viral replication inside infected cells. While type I IFN was discovered in 1957, type III IFN (IFNλ, IL-28/29) was characterized recently and is known for its role in the response to hepatitis C virus. Airway epithelia are the primary target of influenza virus, and we studied how infection induces IFNs and which IFN is most important for the epithelial anti-influenza response. We found that infected epithelia detect virus through the cytoplasmic RIG-I/MAVS recognition system, leading to activation of the transcription factor IRF7 and subsequent induction of both type I and III IFNs. All ensuing cellular responses to infection are dependent on the production and secretion of IFNs, as responses are lost in epithelia lacking receptors for both type I and III IFNs. Finally, gene induction is indistinguishable in single receptor-deficient and wild-type cells, indicating that the two IFN systems are completely redundant in epithelia. Thus, influenza infection of airway epithelia induces, via a RIG-I/MAVS/IRF7 dependent pathway, both type I and III IFNs which drive two overlapping and redundant amplification loops to upregulate antiviral genes.
| The type I interferon family is a group of cytokines encoded by a single IFNβ gene and a tandem cluster of multiple IFNα genes that were first characterized for their ability to interfere with influenza virus replication [1] and are now recognized as powerful inducers of the host response to viral infections.
IFN induction by influenza A virus (IAV) depends on recognition of viral components by either cytoplasmic receptors or the Toll-like receptor (TLR) system, depending on the infected cell type. While plasmacytoid dendritic cells (pDC) use TLR7 to recognize influenza virus, in fibroblasts and conventional DCs IFNβ induction requires recognition of RNA viral genomes by the cytoplasmic RNA helicase retinoic acid-induced gene I (RIG-I) [2], [3]. Upon RNA binding, RIG-I interacts with the mitochondrial adaptor protein MAVS (also known as IPS-1, CARDIF and VISA) and initiates a signalling cascade that culminates in the activation of the transcriptional factors AP-1, NF-κB and IRF3, and the expression of IFNβ and IFNα4 in mouse (or IFNα1 in humans) [4]–[7].
Once secreted, IFNβ and IFNα4 acts in both a paracrine and autocrine way through binding to the ubiquitously expressed heterodimeric IFNα/β receptor (IFNAR1/2) to induce activation of the receptor–associated tyrosine kinases JAK1 and Tyk2 and subsequent phosphorylation of the transcriptional factors STAT1 and STAT2 [8]. Activated STATs then form transcription factor complexes, including STAT1 homodimers and a STAT1/STAT2/IRF9 heterotrimer known as ISGF3 [9], and mediate the induction of hundreds of IFN-stimulated genes (ISGs), whose expression determines the establishment of an antiviral state inside the cell.
Recently, a novel group of IFNs was described and named type III IFNs. This new IFN family has three members: IFNλ1 (a pseudogene in mouse), IFNλ2 and IFNλ3, alternatively named IL-29, IL-28A and IL-28B respectively [10], [11].
IFNλ induction depends on the same triggers and signalling pathways that regulate type I IFN expression [12], [13], with the RIG-I/MAVS/TBK1/IRF3 axis being particularly relevant in mouse embryonic fibroblasts (MEFs) upon viral infection [14].
IL-28/29 act through a distinct receptor complex consisting of IL-28Rα, specific for type III IFNs, and the IL-10Rβ chain, which is also part of the receptors for IL-10, IL-22 and IL-26. Despite activating different receptors, both type I and III IFNs activate the JAK-STAT signalling pathway, that leads to the formation of the ISGF3 complex [15] and the induction of ISGs.
The most important determinant of the different biological activities of IFNα/β and IFNλ is the distribution of their receptors. While the receptor for type I IFNs is expressed on all cells, IL-28Rα is found primarily on epithelial cells of both the respiratory and the gastrointestinal tract [16]–[18]. This finding suggests that type III IFNs may act in a cell-type restricted manner and may selectively contribute to the innate immunity of mucosal surfaces, potential entry sites for many pathogenic viruses.
The airway epithelium is a pseudostratified, columnar epithelium consisting of ciliated, basal and secretory goblet cells, that lies at the interface between the host and the environment and provides the first line of defence against inhaled microorganisms. Airway epithelial cells represent the target of many respiratory viruses, including Influenza virus, Adenovirus, Rhinovirus and RSV [19]. As epithelial cells express both cell surface and endosomal pattern recognition receptors (PRRs) and intracellular viral sensors, they can promptly detect invading microbes and react by producing cytokines, chemokines and antimicrobial peptides, thus initiating inflammatory and immune responses [20], [21].
While the PRRs and downstream signals required for influenza ISG induction have been mapped out in detail in other cell types, it is less clear which recognition systems are in action in airway epithelia. Moreover, while the importance of type III IFNs in epithelial responses has been documented, it is unclear whether the signatures induced by interferon type I or III overlap and which, if any, ISG subsets are selectively induced by one or the other.
To address these issues, we established cultures of primary differentiated murine tracheal epithelial cells (MTEC) and, using a genetic approach together with microarray analysis, investigated the mechanisms that lead to IFN induction in response to influenza A infection. We also assessed the relative contribution of type I and III IFNs to the establishment of an antiviral state by comparing the pattern of influenza-induced gene expression in the absence of either IFNα/β signalling, IFNλ signalling or both. These studies help define the biology and nature of the antiviral state induced by different IFNs in primary cells and determine whether and which genes are still induced by influenza infection in the complete absence of both IFN type I and III signalling.
Primary MTEC were grown to confluence and exposed to air for 14 days, leading to formation of a fully differentiated, polarized epithelium containing ciliated and secretory goblet and Clara cells (Fig. 1A), a cellular composition which closely matches that of airway epithelia in vivo. These differentiated cultures were then infected with IAV, at a multiplicity of infection (moi) of 0.3. Intracellular FACS analysis of the infected cultures showed that approximately 10% of MTEC were expressing viral nucleoprotein (NP) 24 hours post infection (hpi) (Fig. S1). Total RNA was isolated from five replicate cultures 24 hpi and analyzed using microarrays. We first investigated which type of IFN was induced in response to the infection. To this purpose, after normalizing the signal intensity of each probe in each sample to the median intensity of that probe in the control group, and filtering for genes that were expressed above background level, we searched for the probe sets which represent IFNs in the data set. As shown in Fig. 1B, both type I (IFNβ, α4 and α5) and type III IFNs (IL-28A/B) were significantly induced in the infected samples, with IL-28A/B being the most strongly upregulated. The induction of all other IFNα genes did not reach statistical significance. Type II IFN (IFNγ) did not pass the initial filtering for genes expressed above background and was not induced upon infection (not shown). To identify the pattern recognition receptors responsible for IFN induction in this experimental model, we infected MTEC cultures from wild-type and knock-out mice and analyzed the expression of both IFNβ and IL-28 by quantitative RT-PCR and ELISA. While in some immune cells TLR7 is the major PRR mediating IFN induction by influenza virus [22], we found that the absence of TLR7 or its downstream adaptor MyD88 had no impact on the induction of IFNs in influenza-infected epithelia. Similarly, the adaptor TRIF used by TLR3 and TLR4 is not required for IFN induction in epithelia. In contrast, MTEC deficient in MAVS were unable to produce IFNs in response to influenza infection, suggesting involvement of the RIG-I pathway (Fig. 1C, 1D). We next sought to determine whether virus replication is required for the induction of IFNs. Treatment of PR8 at 65°C inactivates the viral polymerase and prevents viral replication, without abrogating virus attachment to the cells (Fig. S2A, S2B). This treatment abolished the virus' ability to induce IFNs (Fig. 1C, 1D) indicating that IAV virus access to the cytoplasm and subsequent replication are required to initiate an IFN response in MTEC.
To identify the transcripts that were differentially expressed in infected MTEC cultures at 24 hpi, we performed a supervised analysis under stringent conditions (≥4-fold change relatively to mock infected samples, t-test unpaired p value <0.01, Benjamini-Hochberg multiple statistical correction). The differentially expressed genes were then partitioned by K-means clustering (Fig. 2A) into 2 groups (177 downregulated genes; 234 upregulated genes), and the 234 upregulated genes were hierarchically clustered to generate the heat map in Fig. 2B. Ingenuity Pathway Analysis of the list of upregulated genes confirmed “activation of interferon regulated factors by cytosolic PRR” and “interferon signalling” as two overrepresented pathways in the infected samples (Fig. 2C). Moreover, 80 (35%) of the upregulated genes were recognized as interferon stimulated gene by the INTERFEROME database.
A major level of control of IFN production depends on transcriptional regulation. The general paradigm for IFNβ induction involving recruitment of the transcription factors IRF3 and p50/RelA NF-κB has recently been shown to apply also to type III IFN induction, at least in MEFs [14]. Unexpectedly, IRF3 deficient MTEC were not impaired in their ability to upregulate both IL-28 and IFNβ in response to infection (Fig. 3A). In contrast, IRF7−/− epithelia showed a marked reduction in the amount of IFN induced; however, the induction of both IL-28 and IFNβ was completely abolished only in doubly deficient IRF3−/−IRF7−/− MTECs, as assessed by qPCR for gene expression and by ELISA for protein secretion (Fig. 3B, C).
Previous studies have demonstrated that the entry of enveloped viruses like HSV and VSV into fibroblast cells can lead to the induction of a subset of ISGs in an IFN-independent manner and that either IRF3 [23], [24], IRF7 [25] or IRF1 [26] may have functions that are redundant to that of ISGF3 and therefore induce an IFN-like transcriptome in the absence of IFN signalling. For these reasons, we sought to determine whether in airway epithelia, ISGs could be induced in the absence of IFNs. The data shown so far point to two situations, i.e. the MAVS−/− and the IRF3−/−IRF7−/− epithelia, in which no IFN production could be detected, both at the protein and at the RNA level (Fig. 1C, 1D, 3B, 4A). Surprisingly, in both conditions, infection with influenza virus led to the induction of most of the ISGs tested, albeit at lower levels than in wild-type epithelia (Fig. 4B and S3). This correlated with virus control as infected wild- type, MAVS and IRF3/7 deficient epithelia had similar virus titers over the course of infection (Figure 4C).
These results can be interpreted in two ways; first, following IAV infection, a subset of ISGs may be induced through a MAVS and/or IRF3/IRF7 independent pathway that does not require interferons. Alternatively, in MAVS−/− and IRF3−/−IRF7−/− cells, minute, steady state amounts of IFNs could still be produced and be sufficient to induce ISGs in a context of viral infection.
To test these alternative hypotheses, we infected epithelia deficient for either type I, type III or both IFN receptors and analysed their transcriptional response to influenza A virus by microarray analysis. RNA from five replicate samples were first normalized to the median of mock infected samples and then filtered on expression (20–100th percentile in at least 50% of samples). A supervised analysis under stringent conditions (≥4-fold change versus wild-type mock in at least one infected group, 2-way ANOVA, p value of <0.01, Benjamini-Hochberg multiple statistical correction) and k-means clustering led to a list of 136 upregulated genes and a list of 50 downregulated genes (not shown). The induced genes were then hierarchically clustered to generate the heat map in Fig. 5A.
Influenza infection of IFNAR1−/−IL-28Rα−/− double knock-out epithelia induced the expression of IFNβ and IL-28A/B at levels comparable to the wild-type controls even at later time points during an infection (Fig. 5B), indicating that these genes are most likely upregulated directly downstream of the RIG-I/MAVS pathway and do not require IFN-driven positive feed-back on themselves. In contrast, IFNAR1−/−IL-28Rα−/− cells have lost the ability to upregulate many of the genes that were induced in the wild-type control (Fig. 5A), including known ISGs such as Rsad2, Oasl2 and others (Fig. 5B).
To analyse more globally the 136 up-regulated genes shown in Fig. 5A, they were further partitioned by K-means clustering into those that were not induced in infected IFNAR1−/−IL-28Rα−/− cells (110 “IFN-dependent” genes, Fig. S4C) and those that were still induced (26 “IFN-independent” genes, Fig. S4A, B): analysis by the INTERFEROME database scored 58 (53%) of the “IFN-dependent” genes as ISGs. The “IFN-independent” genes comprise a smaller group of genes, including many chemokines (Fig. S4B). Although 5 (23%) of these genes (CXCL1, CXCL3, CSF-2, CXCL5 and CD274) were identified as ISGs by the INTERFEROME database, it has been described elsewhere that their expression can be also induced independently of IFNs, most likely through regulation by transcription factors like NF-kB, PPARγ and GATA-1 [27]. The mechanism by which they are induced by IAV in our system is currently under investigation.
The transcriptional signatures obtained for the single IFNAR1 and IL-28Rα knock-outs were very similar to the one for cells of wild-type origin (Fig. 5A). To address directly whether induction of some ISGs specifically depends on one type of IFN, we filtered the list of 136 infection-induced genes in Fig. 5A for genes that differ between either the wild-type and single knock-outs or between the two single knock-outs (Fig. S5). Indeed, only 11 genes were found, and in most cases, genes were induced in all three genotypes although at lower intensity in the single knock-outs.
Overall, our results indicate that in airway epithelia, the induction of an antiviral state depends on either type I or type III IFN signalling. Both types of IFN independently drive parallel, completely redundant amplification loops, each leading to the induction of the same set of genes. In the absence of both receptors, the IFN signature disappears almost entirely, indicating that no other mechanism can replace the IFN loop for the induction of ISGs.
Importantly, the lack of ISG induction in IFNAR1/IL28Rα deficient epithelia has biologic consequences as it leads to significantly higher virus titers at later points during infection (Figure 5C). No significant differences in viral titers were seen between wild-type and either IFNAR1 or IL-28Rα single knock-out.
To test in vivo whether lack of IFN responsiveness in lung epithelia impacts the disease course of influenza infection, we generated bone marrow chimeras where B6 wild-type bone marrow was grafted into either wild-type or IFNAR1/IL28Rα deficient hosts. These two groups have both fully functional immune cells but differ in the ability of radioresistant cells including lung epithelia to respond to type I and III IFNs. We confirmed successful immune cell reconstitution by staining blood cells for IFNαβR (Fig. S6) and by testing Sca-1 upregulation on blood cells in response to IFNβ (not shown). Both experiments confirmed that >85% of immune cells in these chimeras have wt phenotype. When these chimeras were infected with the PR8 strain, high susceptibility and mortality was found only in the group lacking IFN receptors on stromal cells and this correlated with higher viral titers (Fig. 6A, B). These results suggest that the ability to respond to IFNs in infected airway epithelia is crucial for successful elimination of the virus from infected animals.
Infected MAVS−/− and IRF3−/−IRF7−/− epithelia show ISG induction in the absence of any detectable IFN upregulation, while IFNAR1−/−IL-28Rα−/− cells, which produce but cannot respond to IFNs, did not express ISGs in response to infection. Although type I IFN genes are tightly regulated in response to viral infection, many tissues constitutively secrete low amounts of type I IFN even in the absence of infection (reviewed in [28]). It has been proposed that these constitutive levels of IFNs are required to maintain basal expression of IFN-inducible signalling intermediates (STAT1/2, IRF7/9/5) and to modulate the relative expression of STAT proteins, therefore “priming” cells for future responses.
For these reasons, we sought to determine the basal level of different IFN-signalling intermediates at steady state and upon infection in IFNAR1−/−IL-28Rα−/− cells. Wild-type and double knock-out cells were infected with influenza A and the level of different STAT and IRF molecules analyzed by qPCR. Some of these molecules (IRF7, IRF9, STAT1, STAT2) are known ISGs and were upregulated in wild-type but not in IFNAR1−/−IL-28Rα−/− epithelia at 24 hours post infection. However, the levels of all these transcripts measured relatively to HPRT at steady state were comparable in the two genotypes (Fig. 7A, 7B) and in MAVS−/− and IRF3−/−IRF7−/− cells (not shown).
To test directly whether residual IFN production is responsible for ISG induction in MAVS and IRF3/7 deficient epithelia, we compared ISG induction in infected MAVS deficient epithelia in the presence or absence of an antibody cocktail blocking both IFNαβ and IFNλ signalling. As shown in Figure 8A, antibody-treated epithelium had further reduced ISG expression compared to the untreated one. To confirm independently the requirement of autocrine signalling by soluble factors for ISG induction, we infected wild-type epithelia in the presence or absence of brefeldin A (BFA). BFA treatment left IFN gene induction unaffected but abolished ISG induction, indicating that soluble factors which include IFNs are required for ISG induction (Fig. 8B).
Collectively, these results indicate that the different responses to infection in MAVS−/− and IRF3−/−IRF7−/− compared to IFNAR1−/−IL-28Rα−/− cells can not be ascribed to a lack of priming in the latter, but are most likely due to a residual production of IFN in MAVS−/− and IRF3−/−IRF7−/− cells, that, in the context of infection, is sufficient to ensure ISG induction.
Here, we delineate the influenza-triggered pathways leading to the induction of an antiviral state in primary airway epithelia, the first and most important target tissue of the virus in an infected organism. We show that TLR7 or other TLRs relying on the adaptor molecules MyD88 and TRIF are not involved in the induction of interferons, while the RIG-I/MAVS pathway is crucial for this process. We also show that between the two transcription factors IRF3 and IRF7 implied in IFN induction, IRF3 is of less importance than IRF7, but complete abolition of influenza-triggered IFN expression is seen only in the absence of both molecules. Most importantly, we show that, upon influenza infection, IFN type I and III independently mediate parallel amplification loops leading to the induction of a completely overlapping set of ISGs, and that this induction is abolished only when none of the two amplification loops are active.
The general paradigm for type I IFN induction involves recruitment of transcription factors that are activated by phosphorylation in response to signalling cascades stimulated during viral infection. The IFNβ promoter contains four positive regulatory domains (PRDI-IV), which are occupied by different transcription factors. PRDI and III are binding sites for IRF3 (early during infection, due to its constitutive expression) and IRF7 (with delayed kinetics, due to its inducible expression through an IFN-dependent positive feedback loop), while PRDIV and PRDII bind the ATF-2/c-Jun AP-1 and the p50/RelA NF-κB complexes respectively (reviewed in [8]).
This initial model of a positive feedback loop, in which IRF3 is primarily responsible for the early induction of IFNβ while IRF7 is required later in the response [29], was subsequently modified when a study performed on IRF7−/− MEFs revealed that both the early and the late production of type I IFN induced by VSV or EMCV is abolished in the absence of IRF7 [30]. Our results are in line with these observations and identify IRF7 as the major regulator of both type I and type III IFN responses in epithelial cells. Indeed, our data in fig. 6A suggests that the expression level of IRF7 was higher than that of IRF3 at steady state, which would support the notion that IRF7 could act directly downstream of viral recognition to induce IFNs at the earliest stage of infection. Moreover, our data indicates that, even later during an infection, the expression of IFNβ1 and IL-28A/B can be sustained independently of the IFN-driven positive amplification loop (Fig. 5B).
IFNλ is preferentially induced by influenza A virus, both in vivo and in vitro [31]. We extend these findings to primary airway epithelia and demonstrate that epithelial-derived IFN type I or type III are sufficient to fuel their respective amplification loop and that no extrinsic IFNs from other cells, for instance immune cells, is required to induce an epithelial IFN signature.
The importance of type III IFNs in epithelial responses has been well documented. Several studies have shown that IFNλ protects the epithelium of lung, intestine and vagina from viral infections and that IFNAR1−/−IL-28Rα−/− double deficient mice are more susceptible to viral infection than each single knock-out strain [17], [18], [32]. Using chimeric mice with a wild type immune system and either wild type or IFNAR1−/−IL-28Rα−/− double deficient stroma, we show here that IFN unresponsiveness in the stromal cell compartment is sufficient to render mice more susceptible to influenza infection.
Previous studies have shown that the promoters of many ISGs have a simple structure and can be easily turned on directly by IRF proteins independently of interferon. In these studies, alternative pathways of direct ISG induction were suggested that rely on either IRF3 [23], IRF7 [25], IRF1 [26] or peroxisomal MAVS/IRF1/IRF3 [33]. More recently, the cytosolic exonuclease Trex1 has also been identified as a negative regulator of a novel pathway involving STING, TBK1, IRF3 and IRF7 that can lead to interferon-independent activation of ISGs [34].
The finding that a subset of ISGs could still be induced in both MAVS- and IRF3/IRF7-deficient epithelia suggested to us that IFN-independent ISG induction may take place here. However, the nearly complete absence of ISG induction in IFNAR1−/−IL-28Rα−/− epithelia led us to conclude that, at least in our experimental model, no other mechanism can efficiently replace the IFN loop for the induction of ISGs. Constitutive low-level signalling of IFNβ (IFN “priming”) has been suggested to help preserve IFN responsiveness but also to allow IFN-independent ISG induction [28], by maintaining the expression of STATs and other signalling intermediates. It could be argued that, unlike IFNAR1−/−IL-28Rα−/− epithelia, MAVS−/− and IRF3−/−IRF7−/− cells still possess this sub-threshold signalling which helps maintain STATs and IRFs at sufficient levels to allow for direct ISG induction upon viral trigger, even in the absence of IFNs. We did however not detect steady-state differences in the expression of a range of IRF and STAT molecules between genotypes and therefore have no evidence that differences in IFN priming contribute to the phenomenon described here.
Previous studies that assessed IFN independent ISG induction have mostly relied on IFNAR deficient cells to confirm IFN independence. Here we show that care must be taken to evaluate IFN independence. As IFN type III can stand in for IFN type I in inducing an IFN signature, the analysis of each single receptor knock-out epithelium would have wrongly suggested complete independence of ISG induction from the IFN system. Moreover, while the IFN signature was completely abolished in influenza infected IFNAR1−/−IL-28Rα−/− cells, the addition of neutralizing antibodies against secreted type I and type III IFNs, used in combination on wild-type epithelia or for the complementary IFN on IFN receptor single knock-out epithelia, had little effect on ISG induction (not shown), indicating that even minute concentrations of IFN were still able to induce ISG expression in responsive cells.
In vivo studies with single-knock-out mice clearly showed that ISG induction by type III IFN translates into less powerful protection against influenza A virus than ISG induction by type I IFN [32]. At present it is unclear whether slight differences in the kinetics of virus-triggered induction of type I and type III IFN may account for this observation. An alternative explanation is that lung macrophages which are productively infected by most influenza A virus strains and which do not respond well to type III IFN may quickly amplify the incoming virus in the respiratory tract of IFNAR1-deficient mice and thus overwhelm the type III-mediated protection of epithelial cells in such mice.
Through the induction of ISGs, the IFN system has potent effects not only to directly combat virus, but also on cell physiology and survival and on the immune response. Therefore, it is considered a very tightly controlled system to avoid excessive inflammation, cell death and tissue damage. On this background, we were surprised to find that in MAVS and IRF3/7 deficient epithelia, IFNs at levels below ELISA or qPCR detection threshold still lead to only slightly reduced ISG upregulation. ISG induction disappeared in MAVS deficient epithelia when IFN type I and III signalling was blocked by an antibody cocktail, indicating that a residual IFN production drives ISG induction in these cells.
In apparent contrast to these results, when exogenous IFN was titrated onto epithelial cultures, the minimum amount of IFN required to induce ISG expression was clearly detectable by the ELISA assay (not shown). Possible reasons for this discrepancy could be differences in the bioactivity of endogenous versus recombinant IFNs and differential ability by the ELISAs to detect endogenous or recombinant IFNs. Moreover, biological explanations include local concentrations of endogenous IFNs that may be much higher than those measured in the total supernatant, and the possibility that autocrine IFNs bind to their receptor already in intracellular vesicles and are therefore not measured by ELISA.
Overall these data suggest that presence or absence, rather than absolute amounts of IFN, determine the response. The biological sense of such a binary switch could be to respond robustly even to small perturbations of the steady state, which may gain the host precious time when infected by fast-replicating viruses. In the uninfected state, absolute IFN shut-down would be required to avoid chronic “Flu-like symptoms”, a scenario that is in contradiction to the proposed priming effect of sub-threshold IFN levels.
One hypothesis for how the vast differences in IFN levels are translated into a largely unaltered IFN signature in MAVS−/− and IRF3−/−IRF7−/− epithelia is that Influenza A infection may render cells much more IFN-sensitive by unknown pathways. To test this hypothesis, we titrated exogenous IFN on uninfected or infected wt or MAVS-deficient epithelia and measured ISG induction. No increased IFN responsiveness was found in infected versus uninfected epithelia, suggesting that there is no synergy between these two signals (not shown). An alternative hypothesis is that IFN protein that is prestored [35] and therefore not measurable by qPCR (not transcriptionally controlled) or by ELISA (too low sensitivity) could be released locally and mediates the observed ISG induction.
In conclusion, we show here that airway epithelia rely on two parallel, redundant amplification loops to induce an IFN signature in response to influenza A infection. Only a small fraction of genes, mostly non ISGs, are induced by the virus in the absence of both IFN systems, and no ISG appears to rely specifically on one IFN system only. This complete redundancy may guarantee induction of antiviral responses even if one or the other IFN system is blocked, for instance by specific virally encoded antagonists. In contrast to the two redundant IFN loops in epithelia, the majority of immune cells respond only to IFN type I, thus potentially allowing for differential control of the epithelial antiviral state and the induction of immune responses: while high levels of IFN type I would activate both epithelia and immune cells, high levels of type III would specifically activate epithelial responses but leave immune responses unaffected, which may help limit immune-mediated pathology in the lung and at other mucosal surfaces.
All animal breeding was approved by the local ethical committee of the NIMR and is part of a project approved by the UK Home Office (licence number 80/2236). Breeding was conducted according to local guidelines and UK Home Office regulations under the Animals Scientific Procedures Act 1986 (ASPA).
Influenza A virus strain A/PR/8/34 (H1N1) was grown in day 10 embryonated chicken eggs, and titrated on MDCK by 50% tissue culture infective dose (TCID50), according to the Spearman-Karber method.
All cells used in this study were derived from mice on the C57BL/6 background. MAVS−/− mice [36] and tracheae from TLR7−/− mice [37] were kindly provided by Dr. C. Reis e Sousa; MyD88−/− [38], IFNAR1−/− [39] and TRIF−/− mice [40] were kindly provided by Dr. A. O'Garra. These and C57BL/6 wt mice were bred in-house under SPF conditions. Tracheae from Irf3−/− [29]; Irf7−/− [29]; Irf3−/−Irf7−/− cells were also used. IL-28Rα−/− [12], IFNAR1−/− [39] and IL-28Rα−/−IFNAR1−/− cells were obtained from mice on a congenic B6.A2G-Mx1 background carrying an intact Mx1 gene [41].
To generate chimeric mice, naïve B6.A2G-Mx1 and B6.A2G-Mx1 IL-28Rα−/−IFNAR1−/− recipient mice were lethally irradiated with 1000 rad and reconstituted with donor B6.A2G-Mx1 BM cells (7×106) by intravenous injection. Chimeric mice were maintained for 7 weeks and chimerism assessed by IFNAR1 (MAR1-5A3 antibody) expression on Gr1+, CD19+, CD4+ and CD8+ cells in the blood (Fig. S6). Chimeric mice were then infected intranasally with 105 TCID50 of Influenza A/PR/8/34 in 30 µl PBS after anesthesia.
Isolation and culture of primary MTEC were performed as previously described [42]. Briefly, cells isolated by enzymatic treatment were seeded onto 0.4 µm pore size clear polyester membrane (Corning) coated with a collagen solution. At confluence, media was removed from the upper chamber to establish an air- liquid interface (ALI). Fully differentiated, 10–14 days-old post ALI cultures were routinely used for experiments.
In some experiments, cultures were infected in the presence of brefeldin A (2.5 µg/ml) or neutralizing anti-IL28A/B, anti-IL28B (R&D Systems) and blocking anti-IFNAR1 (MAR1-5A3) antibodies, at a final concentration of 10 µg/ml each.
Differentiated, ALI day 14 cultures were fixed in 4% paraformaldehyde and permeabilized with 0.1% Triton X-100. Cells were then incubated with the indicated primary antibodies for 1 hour at room temperature, washed, incubated with fluorochrome-conjugated secondary antibody, washed and finally mounted. Image acquisition and processing information: (i) microscope: Olympus IX70; (II) magnification: 20×; (III) imaging medium: Vectashield with DAPI (Vector labs); (IV) fluorochromes: Alexa Fluor 488, Alexa Fluor 568. (V) acquisition software: Softworx. Images were processed with Image J.
The apical surface of MTEC cultures was washed extensively to remove accumulated mucins before inoculation with IAV (moi = 0.3). After incubation at 37°C for 1 h, the virus inoculum was removed and the cultures were incubated in complete growth medium for 24 hours. Aliquots of the supernatants were collected at different time points and titrated by ELISA as described below. Cells were then lysed to extract RNA or for detection of viral protein by Western blotting.
RNA was isolated from MTEC cultures by directly lysing the cells in the transwells, using the Qiagen RNeasy mini kit, according to the manufacturer's instructions. One microgram total RNA was reverse transcribed using the ThermoScript RT-PCR System kit (Invitrogen). The cDNA served as template for the amplification of genes of interest and the housekeeping gene (Hprt1) by real-time PCR, using TaqMan Gene Expression Assays (Applied Biosystems), universal PCR Master Mix (Applied Biosystems) and the ABI-PRISM 7900 sequence detection system (Applied Biosystems). The fold increase in mRNA expression was determined using the ΔΔCt method relatively to the values in mock treated samples, after normalization to Hprt1 gene expression.
Total RNA harvested from MTEC cultures was hybridized using Affymetrix Mouse Genome 430 2.0 microarrays. The raw intensities values for each entity were preprocessed by RMA normalization against the median intensity in mock infected samples. Using GeneSpring 11.5, all transcripts were filtered based on signal values, to select the ones whose level of expression was in the 100–20th percentile, in at least 50% of samples. Student's t test (infected versus mock infected) or 2-way ANOVA (parameters: treatment and genotype) were performed to identify gene significantly differentially expressed relative to controls (≥4-fold change; p<0.01, Benjamini-Hochberg multiple test correction).
Ingenuity Pathway Analysis (IPA) was used to select, annotate and visualize gene by function and pathway. ISGs were identified with the Interferome database (www.interferome.org/).
Microarray data can be accessed at GEO under accession number GSE43710 for the superseries.
Cell culture supernatants were harvested from the apical compartments of mock or IAV infected samples. IL-28A/B was measured using the IL-28A/B ELISA Duo kit (R&D Systems), IFNβ with the Verikine IFNβ ELISA kit (PBL Interferon Source).
|
10.1371/journal.ppat.1004460 | Kaposi's Sarcoma-Associated Herpesvirus Induces Nrf2 during De Novo Infection of Endothelial Cells to Create a Microenvironment Conducive to Infection | Kaposi's sarcoma-associated herpesvirus (KSHV) is the etiological agent of Kaposi's sarcoma (KS) and primary effusion B-cell lymphoma. KSHV induces reactive oxygen species (ROS) early during infection of human dermal microvascular endothelial (HMVEC-d) cells that are critical for virus entry. One of the downstream targets of ROS is nuclear factor E2-related factor 2 (Nrf2), a transcription factor with important anti-oxidative functions. Here, we show that KS skin lesions have high Nrf2 activity compared to healthy skin tissue. Within 30 minutes of de novo KSHV infection of HMVEC-d cells, we observed Nrf2 activation through ROS-mediated dissociation from its inhibitor Keap1, Ser-40 phosphorylation, and subsequent nuclear translocation. KSHV binding and consequent signaling through Src, PI3-K and PKC-ζ were also important for Nrf2 stability, phosphorylation and transcriptional activity. Although Nrf2 was dispensable for ROS homeostasis, it was essential for the induction of COX-2, VEGF-A, VEGF-D, Bcl-2, NQO1, GCS, HO1, TKT, TALDO and G6PD gene expression in KSHV-infected HMVEC-d cells. The COX-2 product PGE2 induced Nrf2 activity through paracrine and autocrine signaling, creating a feed-forward loop between COX-2 and Nrf2. vFLIP, a product of KSHV latent gene ORF71, induced Nrf2 and its target genes NQO1 and HO1. Activated Nrf2 colocalized with the KSHV genome as well as with the latency protein LANA-1. Nrf2 knockdown enhanced ORF73 expression while reducing ORF50 and other lytic gene expression without affecting KSHV entry or genome nuclear delivery. Collectively, these studies for the first time demonstrate that during de novo infection, KSHV induces Nrf2 through intricate mechanisms involving multiple signal molecules, which is important for its ability to manipulate host and viral genes, creating a microenvironment conducive to KSHV infection. Thus, Nrf2 is a potential attractive target to intervene in KSHV infection and the associated maladies.
| KSHV infection of endothelial cells in vivo causes Kaposi's sarcoma and understanding the steps involved in de novo KSHV infection of these cells and the consequences is important to develop therapies to counter KSHV pathogenesis. Infection of endothelial cells in vitro is preceded by the induction of a network of host signaling agents that are necessary for virus entry, gene expression and establishment of latency. Our previous studies have implicated reactive oxygen species (ROS) as part of this network. In the current study, we show that ROS activate Nrf2, a master transcriptional regulator of genes involved in ROS homeostasis, apoptosis, glucose metabolism and angiogenesis. Besides ROS, KSHV utilizes additional aspects of host signaling to induce Nrf2 activity. We also observed that infection of endothelial cells deficient in Nrf2 resulted in downregulation of multiple genes important in KSHV pathogenesis, such as COX-2 and VEGF, and affected proper expression of two hallmark KSHV genes, lytic ORF50 and latent ORF73. Taken together, this study is the first to demonstrate the importance of Nrf2 during de novo KSHV infection of endothelial cells, and establishes Nrf2 as an attractive therapeutic target to control KSHV infection, establishment of latency and the associated cancers.
| Kaposi's sarcoma-associated herpesvirus (KSHV) or human herpesvirus 8 (HHV-8), a γ-2 lymphotropic herpesvirus with a double-stranded DNA genome of ∼160 kb in length, is the etiological agent of hyper-proliferative disorders such as Kaposi's sarcoma (KS), primary effusion B-cell lymphoma (PEL), and plasmablastic multicentric Castleman's disease (MCD) [1]–[3]. KS lesions exhibit a heterogeneous environment of hyperplastic, endothelium-derived spindle cells, neovascular structures and inflammatory cells [4]. Like all herpesviruses, the KSHV life-cycle alternates between lytic and latent phases, and KSHV is predominantly in the latent state in KS endothelial cells [5]. KSHV genome and transcripts are also detected in the KS lesion fibroblasts, monocytes, and cells of epithelial origin and the expression of multiple latent and lytic genes in the infected cells, aided by the concomitant action of pro-inflammatory cytokines released by these cells, drives the excessive proliferation and hyperplasia of endothelial cells that lead to their spindle-shaped morphology [5].
Investigation of KSHV infection of endothelial cells is frequently carried out in vitro in primary endothelial cell types such as human dermal microvascular endothelial cells (HMVEC-d), human umbilical vein endothelial cells (HUVEC) and lymphatic endothelial cells (LEC), or in immortalized endothelial cell-lines such as TIVE/TIVE-LTC and epithelial SLK/iSLK cells. HMVEC-d cells provide an excellent in vitro model for studying the early events that follow de novo infection of endothelial cells because i) they are naïve, primary cells permissive to KSHV infection, ii) they are derived from the same cells that eventually transform into the characteristic spindle-shaped morphology in KS lesions, and iii) are not transformed, hence, exhibit signaling cascades that closely resemble early events during in vivo infection [6]. As primary cells, HMVEC-d cells have a limited life-span and culturing them is labor intensive, with about ∼50–70% infection efficiency if ample virus is used, and exhibit progressive viral episome loss with each cellular division [5], [6].
The KSHV-binding receptor on HMVEC-d cells is heparan sulfate (HS), a negatively-charged plasma membrane macromolecule that uses electrostatic forces to attract KSHV envelope glycoproteins to the cell surface [7]–[11]. Once on the surface of the cells, KSHV envelope glycoproteins interact with entry receptors such as integrins (α3β1, αVβ3 and αVβ5), xCT/CD98, and the receptor tyrosine kinase EphA2 to induce important signaling pathways that result in the phosphorylation and activation of many additional kinases and transcription factors [8], [10]–[14]. Specifically, KSHV infection sequentially induces activation of FAK, Src, PI3-K/Akt, ROS, EphA2, c-Cbl and CIB1 to mediate macropinocytosis and virus entry [14]–[21]. Subsequently, infection induces activation of PKC-ζ, COX-2, MAPKs (MEK and ERK1/2) and NF-κB, which collectively create a microenvironment conducive to establishment of viral gene expression and latency [10], [22]–[25]. Unlike alpha- and beta-herpesviruses, whose in vitro infection results in robust lytic replication with high progeny virus formation and cytopathological changes in the cell, KSHV establishes latency within 24 hr post-infection (p.i.), which is clearly mirrored by a steady rise in major latency regulatory ORF73 (LANA-1) gene expression and with no progeny virus formation [5], [26]. Another interesting feature of de novo infection is that before the establishment of latency, a quick burst of lytic genes with important anti-apoptotic and immune-evasive roles peaks between 2–8 hr p.i., which subsequently declines by 24 hr p.i., likely due to i) increase in the antagonizing LANA-1 expression, ii) chromatin modifications of the KSHV genome, and iii) potentially other unidentified mechanisms [26].
We and others have also shown that stress-associated agents, such as reactive oxygen species (ROS), play important roles in KSHV pathogenesis. ROS have been shown to induce lytic reactivation of KSHV in latently-infected endothelial cells [27]–[29]. Our studies for the first time demonstrated that de novo KSHV infection of HMVEC-d cells induces ROS by ∼2-fold as early as 30 min p.i., which was sustained throughout the course of infection and during latency [16]. This induction played an important role in mediating phosphorylation of a signal pathway involved in macropinocytosis, virus entry and establishment of KSHV infection [16]. However, these studies did not examine the downstream effects of ROS activation.
Nuclear factor E2-related factor 2 (Nrf2) is a ROS-responsive, ∼65 kDa, master transcription factor involved in the transcriptional activation of hundreds of human genes [30]. Nrf2 belongs to the basic leucine zipper (bZIP) subset of the cap ‘n’ collar family of transcription factors and consists of six highly conserved Nrf2-ECH homology domains labeled Neh1-6 [30]. Neh1 contains the DNA-binding domain; Neh2 contains the ETGE and DLG motifs that bind to its inhibitor, Keap1 (Fig. S1); Neh3 is important for activity of the transactivation domains Neh4 and Neh5; Neh6 binds to GSK-3β and β-TrCP [31]–[40]. When activated, Nrf2 binds to DNA promoters that contain the anti-oxidant response element (ARE - TGAnnnnGC) and induces expression of genes such as NQO1, GCS, HO1, and GST etc. involved in the stress response [34], [41]–[45]. In addition, recent studies have determined that Nrf2 induces transcription of genes involved in drug clearance (Mrp1 and Mrp2) [46], [47], glucose and glutamine metabolism (G6PD, TKT, TALDO, PGD, ME1 etc.) [48], [49], apoptosis (Bcl-2, Bcl-xL) [50], [51], angiogenesis (HIF-1α and VEGF) [52]–[54] and cell invasion (MMP9) [55].
Nrf2 is the major regulator of ROS homeostasis in multiple cell types [56]. Because of the importance of responding to elevated ROS rapidly before irreversible cell damage ensues, the cells have evolutionarily developed a quickly-inducible system of Nrf2 activation. In a steady state, new Nrf2 molecules are constantly translated in abundance, but are quickly degraded by its inhibitor, kelch-like ECH-associated protein 1 (Keap1), which acts as a scaffold for the E3 ubiquitin ligase Cul3 (Fig. S1) [57]–[59]. Keap1 consists of three domains, i) the BTB domain (Broad Complex, Tramtack, and Bric-a-Brac), ii) the linker region heavy in reactive cysteine residues, and iii) the Kelch domain [60]. The N-terminal BTB domain homodimerizes with another Keap1 molecule and recruits Cul3 (Fig. S1A, i) [57]. In the closed conformation, each of the C-termini Kelch domains of a Keap1 homodimer binds to the ETGE or DLG domains of one Nrf2 molecule (Fig. S1A, ii) [61], [62]. The N-terminus-bound Cul3 then mediates ubiquitination of 7 lysine residues on the Neh2 domain of Nrf2 (Fig. S1A, ii), an event that opens the complex and allows 26S proteasomal degradation of Nrf2 and recycling of Keap1-Cul3 (Fig. S1A, iii) [58], [59], [63], [64]. Therefore, in a steady state, this system cycles between a closed conformation important for Nrf2 ubiquitination, and an open conformation important for Nrf2 degradation and Keap1-Cul3 recycling (Fig. S1A, overarching arrow) [61], [62]. The recycled Keap1-Cul3 complex then targets a newly synthesized Nrf2, restarting the cycle and keeping Nrf2 levels low.
When ROS levels are elevated, free radicals attack multiple cysteine residues on Keap1, leading to conformational and functional changes that disrupt the normal activity of the Keap1-Cul3 system (Fig. S1B) [60], [65]. Specifically, Nrf2 inducers “lock” the Keap1-Nrf2 interaction in the closed conformation, with both Kelch domains tightly bound to ETGE and DLG, irrespective of the ubiquitination status of Nrf2. As new Nrf2 is translated, the Keap1-Cul3 system is quickly saturated (Fig. S1B, ii and iii) and the newly synthesized Nrf2 accumulates in the cell, free of inhibition from the Keap1-Cul3 ubiquitination machinery (Fig. S1B, i) [61], [62]. The Nrf2 inducers tBHQ and sulforaphane have been shown to affect this pathway [59]. Once stabilized, Nrf2 is phosphorylated on its Ser-40 residue, which results in its nuclear translocation and transcriptional activation of multiple ARE-responsive genes [66]. Several kinases have been reported to either directly or indirectly affect Nrf2 activity, including ERK1/2, casein kinase 2 (CK-2), multiple PKCs and PI3-K/Akt (Fig. S1B, i) [67]–[72].
Several stimuli can induce Nrf2 activation, and virus infection is one of them. The Influenza A virus induced ROS, and as a consequence Nrf2 in alveolar epithelial cells, an event that mitigated the Influenza-induced alveolar toxicity [73]–[75]. Human cytomegalovirus (HCMV) also induced Nrf2 activity during infection of human foreskin fibroblasts, which enhanced the productivity of the infection by decreasing the cytopathic effects of the infection [76]. Hepatitis C virus (HCV), through its core, E1, E2, NS4B and NS5A proteins, induced Nrf2 activity via ROS-dependent and ROS-independent mechanisms involving multiple kinases [77]–[80]. Recently, the Marburg virus structural protein VP-24 has been shown to inhibit Keap1 activity and enhance Nrf2-mediated anti-inflammatory responses [81], [82].
Oxidative stress is essential in development of all four types of KS [83] and analysis of KS tissue sections in the present study detected enhanced Nrf2 activity. We hypothesized that KSHV infection of endothelial cells induces ROS levels to manipulate Nrf2. Studies conducted to test this hypothesis demonstrate for the first time that KSHV induces Nrf2 activity during de novo infection of HMVEC-d cells. KSHV binding, signaling, gene expression and cytokine induction play important roles in inducing Nrf2 activity by mediating its stability, Ser-40 phosphorylation and nuclear translocation. This induction required ROS upregulation as well the activity of a series of host kinases induced by KSHV. Nrf2 induction was essential for the transcriptional activation of multiple host and viral genes that play important functions in KSHV infection. Collectively, these data suggest that KSHV has developed multiple mechanisms to induce Nrf2 activity during de novo infection, establishing Nrf2 as an important agent in KSHV biology and KS pathogenesis.
Kaposi's sarcoma (KS) lesions consist of a heterogeneous environment of multiple cell types that include spindle-shaped endothelial cells, fibroblasts, monocytes, and neovascular structures [4]. At the molecular level, biochemical studies by us and others have demonstrated that KS lesions exhibit substantially elevated levels of pro-inflammatory and stress-associated agents like NF-κB, COX-2 and PGE2 [10], [22]–[25], [84], [85]. A hallmark of the stress response is the activation of the master regulator Nrf2 [56]. To determine whether KSHV-positive tissues exhibit enhanced Nrf2 activity, we performed an immunofluorescence assay (IFA). KS tissue samples exhibited substantially higher levels of Nrf2 compared to normal tissue (Fig. 1A). Moreover, while Nrf2 in healthy tissue localized mostly in the cytoplasm (Fig. 1A, top enlarged box, white arrows), we observed that Nrf2 localized predominantly in the nuclei of KS tissue cells (Fig. 1A, bottom enlarged box, red arrows) in addition to the cytoplasmic distribution (Fig. 1A, bottom enlarged box, white arrows).
Ser-40 phosphorylation of Nrf2 is an essential step in its activation that is required for its nuclear translocation and transcriptional activation, resulting in the augmentation of Nrf2-dependent gene expression [59]. To determine if the increased nuclear Nrf2 observed in KS tissues is of the phosphorylated, active form, we reacted the tissues with an antibody specific to Ser-40-phosphorylated Nrf2 (pNrf2). As expected, the level of pNrf2 in healthy tissue cells was undetectable in the nucleus or the cytoplasm (Fig. 1B, top two rows), indicating low Nrf2 activity. In contrast, KS tissue exhibited elevated levels of pNrf2, which predominantly colocalized with cells expressing latency-associated LANA-1, a common marker of KSHV infection (Fig. 1B, bottom row).
To determine if Nrf2 activation was linked with KSHV infection, we quantified the association between KSHV infection of a cell and its pNrf2 levels by counting all DAPI-staining cells in a particular field (red boxes on the bottom row of Fig. 1B) and considered that as the total number of cells within that field. Cells with only LANA-1 staining overlapping with DAPI were categorized as KSHV+/pNrf2−, while cells that displayed only detectable pNrf2 staining were categorized as KSHV−/pNrf2+. Cells with no staining other than DAPI were categorized as KSHV−/pNrf2−, while those that displayed a triple colocalization of the two proteins and DAPI were categorized as KSHV+/pNrf2+. By this method, >65% of the cells were positive for both LANA-1 and high levels of pNrf2 (Fig. 1C, KSHV+/pNrf2+). Fewer than 14% of the cells were LANA-1-deficient and exhibited high levels of pNrf2 expression (Fig. 1C, KSHV−/pNrf2+), while only 10% of the cells expressed LANA-1 but showed undetectable levels of pNrf2 (Fig. 1C, KSHV+/pNrf2−).
Because KS tissue may contain uninfected or non-endothelial bystander cells that may confound the observed increased Nrf2 activity, we compared the level of pNrf2 in KS areas abundant in spindle-shaped cells (Fig. 1D, area enclosed by dashed white line) to areas of tissue scarce in spindle-shaped cells (Fig. 1D, surrounding area). The spindle-shaped cells, in addition to their typical elongated nuclei, exhibited a high level of LANA-1 as well as pNrf2 staining in contrast to the cells in the surrounding areas (Fig. 1D). Moreover, mostly white staining, representing the triple-colocalization of pNrf2, LANA-1 and DAPI (Fig. 1E), was observed in the enlarged image of the merged panel (Fig. 1D, red arrows), which clearly suggested a predominantly nuclear localization of pNrf2 and LANA-1 in KS tissue.
Taken together, these results demonstrated that KS skin tissue exhibits elevated levels of active, nuclear Nrf2, and that this effect correlates strongly with KSHV infection.
Since KSHV-infected tissues had increased Nrf2 activity, we next determined whether de novo KSHV infection induces Nrf2 activity in endothelial cells. KSHV infection of HMVEC-d cells induces a signaling cascade that promotes virus entry, nuclear delivery and viral gene expression [8], [10]–[14]. To determine and quantitate virus binding and entry-mediated signaling pathways, we serum-starved HMVEC-d cells for 8 hr to reduce the signals induced by the serum (growth factors) in the culture medium. We then infected the cells for the indicated time points, and the levels of total Nrf2 (tNrf2) and pNrf2 were assessed by immunoblot analysis. We observed a 1.8-fold induction of tNrf2 as early as 30 min p.i., which steadily increased to ∼5-fold at 2 hr p.i. and was sustained during the monitored period of 8 hr p.i. (Fig. 2A, second panel). We also observed a robust ∼4-fold induction of pNrf2 as early as 30 min p.i., which peaked to 8-fold at 2 hr p.i. and was sustained during 8 hr p.i. (Fig. 2A, first panel). PKC-ζ phosphorylation was used as an infection marker, as it has been previously shown to be induced during de novo KSHV infection (Fig. S2A) [24]. To determine whether ROS can activate Nrf2 in HMVEC-d cells, we treated the cells with H2O2, a member of the ROS family, and observed an induction in tNrf2 levels to a similar magnitude, indicating a possible role for ROS in tNrf2 accumulation during KSHV infection (Fig. 2B, middle panel). H2O2 also induced pNrf2 by a similar magnitude as tNrf2 (∼1.9 vs. 2.0-fold at 15 min) (Fig. 2B), suggesting that HMVEC-d cells have a constitutive pathway that readily phosphorylates Nrf2 into pNrf2.
To determine whether the increase in pNrf2 observed during KSHV infection is an indirect result of constitutive tNrf2 phosphorylation, or due to an increase in activity of KSHV-induced kinases during early infection, the ratio of fold induction of pNrf2 to tNrf2 was plotted for each time point (Fig. 2C). If the induction of Nrf2 phosphorylation was constitutive, as with H2O2, one would expect the ratio to be ∼1.0 (Fig. 2C, dashed horizontal line). However, the actual ratio of pNrf2/tNrf2 fold induction was significantly higher than 1.0 throughout the observed course of early infection, 2.3-fold higher at 30 min p.i., leveling to ∼1.6-fold higher at 2 hr p.i. and thereafter (Fig. 2C, red line). These results suggested that Nrf2 phosphorylation during KSHV infection involves signaling kinase(s) induced by de novo KSHV infection. The kinetics of Nrf2 phosphorylation are consistent with KSHV signaling kinetics, where signaling occurs early during KSHV entry, followed by a series of cytokine releases that mediate a second wave of signaling [23]. The biology of de novo KSHV infection differs widely between early infection, where events are driven by virus-host surface receptor interactions, and late infection, where events are driven by the expression of a few KSHV latent genes. To assess whether latent KSHV infection also induces Nrf2 activation, we performed a Western blot analysis on HMVEC-d cells infected for 18 and 24 hr. We observed a 4.6 and 2.2-fold induction of tNrf2 and a 4.4 and 2.1-fold induction of pNrf2 at 18 and 24 hr p.i., respectively (Fig. 2D).
Collectively, these results demonstrated that de novo KSHV infection of HMVEC-d cells induced tNrf2 accumulation and its Ser-40 phosphorylation during the early and late stages of infection.
Phosphorylation of Nrf2 is an important step that leads to its nuclear localization and increased transcriptional activity [59]. IFA of uninfected cells and cells infected with KSHV for 2 and 24 hr showed that tNrf2 was elevated throughout the cell (cytoplasm and nucleus) during the infection (Fig. 3A, red and white arrows). pNrf2, on the other hand, accumulated predominantly in the nuclei of infected cells both at 2 and at 24 hr p.i. (Fig. 3B, red arrows), while no cytoplasmic accumulation was observed (Fig. 3B, white arrows). Western blotting of fractionated cytoplasmic and nuclear protein from uninfected and infected cells corroborated the IFA data. Specifically, the levels of tNrf2 were induced by 1.7, 6.0 and 2.6-fold in the cytoplasm at 0.5, 2 and 24 hr p.i., respectively, while pNrf2 was induced by 2.7, 4.3 and 3.0-fold at the same time points (Fig. 3C, lanes 1–4). The nuclear protein fraction also exhibited a robust increase in tNrf2 by 4.2, 4.8 and 3.3-fold and pNrf2 by 4.3, 5.0 and 4.1-fold at 0.5, 2 and 24 hr p.i., respectively (Fig. 3C, lanes 5–8). Interestingly, although IFA at 2 hr p.i. showed predominantly nuclear pNrf2, the western blot assay demonstrated both nuclear and cytoplasmic pNrf2 accumulation at the same time point. The explanation for this discrepancy remains elusive, and is likely due to the different sensitivity of the pNrf2 antibody used in these different techniques. Taken together, these results demonstrated that Nrf2 induction during de novo KSHV infection of HMVEC-d cells leads to nuclear accumulation of pNrf2.
Because of the necessity to respond to stress in a rapid manner, cells constitutively translate new Nrf2 protein. However, under unstressed conditions, these cells maintain low Nrf2 protein levels by shunting it towards the proteasome through the ROS-dependent Keap1-Cul3 ubiquitination axis [59]. In this complex, Keap1 acts as a scaffolding protein for the Cul3 E3 ubiquitin ligase, which mediates lysine-48 ubiquitination of Nrf2 [57]–[59]. Once a targeted Nrf2 has been amply ubiquitinated, this complex releases it for proteasomal degradation, and scavenges for another newly synthesized Nrf2 [33], [58], [63], [64]. When ROS levels are elevated, conformational changes in Keap1 disrupt the axis, Cul3 cannot ubiquitinate newly synthesized Nrf2, which quickly accumulates inside the cell instead of being degraded by the proteasome [60]–[62], [65]. This leads to an appropriate anti-oxidative response achieved through a prompt Nrf2 increase and transcriptional upregulation of its target genes.
A KSHV-mediated increase in Nrf2 protein levels, as shown in figures 2 and 3, could be either due to upregulation of Nrf2 transcription, or due to modulation of Nrf2 protein stability through destabilization of the Keap1-Cul3 axis. To determine whether KSHV induces total Nrf2 levels by increasing its transcriptional levels, we performed real-time RT-PCR analysis using Nrf2-specific primers. We observed a 1.2, 1.8 and 1.4-fold induction of Nrf2 mRNA at 2, 8 and 24 hr p.i. (Fig. 4A). While statistically significant, the fold induction of the transcript is substantially lower than the fold induction of Nrf2 protein observed in figure 2A, explaining only in part the observed increase in total Nrf2 protein. Moreover, because the induction of Nrf2 protein preceded the induction of its transcript, the increase in Nrf2 transcript is likely a consequence and not a significant cause of increased Nrf2 protein. This is consistent with previous studies that have established the Nrf2 gene promoter as a target of Nrf2 transcriptional activity [86].
Since transcriptional upregulation may not be the only reason for the observed Nrf2 protein increase, we focused on the main post-translational modifier of Nrf2, the Keap1-Cul3 ubiquitination axis. Keap1 is highly abundant in cysteine residues, an amino acid that contains highly reactive thiol groups (-SH). When ROS levels increase, neighboring thiol groups of Keap1 get oxidized by the highly nucleophilic electrons present on oxygen radicals, and interact with one-another to form disulfide bonds (R-S-O-S-R), leading to conformational and functional changes in Keap1 [87]. It is for this reason that the Keap1-Nrf2 interaction is especially sensitive to ROS, which we have shown to be induced during early as well as later stages of de novo KSHV infection of HMVEC-d cells [16].
To determine if ROS are responsible for the induction of tNrf2 by KSHV, we utilized N-Acetylcysteine (NAC) and pyrrolidine dithiocarbamate (PDTC), two well-characterized anti-oxidants. At 2 hr p.i., cells pretreated with NAC exhibited decreased pNrf2 and tNrf2 induction by 82% and 77%, respectively (Fig. 4B). Similarly, PDTC pretreatment inhibited pNrf2 and tNrf2 induction by 87% and 96%, respectively (Fig. 4B). Because NAC and PDTC can affect cells in ROS-independent ways, as in the case of their inhibition of NF-κB, we used an additional ROS inhibitor, Diphenyleneiodonium (DPI), which has been shown to inhibit ROS in macrophages and endothelial cells [88], [89]. Similarly to NAC and PDTC, DPI treatment of uninfected HMVEC-d cells reduced, albeit weakly, the levels of pNrf2 (Fig. 4C, lanes 1–6), likely by decreasing basal ROS activity. More importantly, DPI pretreatment significantly reduced KSHV-mediated tNrf2 and pNrf2 induction in a dose-dependent manner (Fig. 4C, lanes 7–12), while not affecting KSHV-mediated NF-κB activation (Fig. S2B). These results further confirmed the importance of ROS induction in Nrf2 activation.
At the later time points (24 hr p.i.), cells pretreated with NAC exhibited diminished pNrf2 and tNrf2 induction by 83% and 77%, respectively (Fig. S2C, lane 3). PDTC pretreatment inhibited pNrf2 and tNrf2 induction by 93% and 88%, respectively (Fig. S2C, lane 4). Both drugs decreased tNrf2 levels below basal levels when compared to uninfected/untreated cells, which is most likely due to their anti-oxidative properties decreasing ROS levels of the cells below basal, an effect which increases Keap1 activity and Nrf2 degradation (Fig. 2SC, compare lanes 3 and 4 to 1). Such findings were corroborated by IFA analysis, which showed that the KSHV-mediated pNrf2 nuclear localization at 2 and 24 hr p.i. was substantially abrogated by pretreatment with NAC and PDTC (Fig. S2D). However, we have shown previously that pretreatment of the cells with anti-oxidants drastically reduces the entry and infectivity of KSHV [16]. Therefore, the inhibition of Nrf2 induction observed at 24 hr p.i. by NAC or PDTC (Fig. S2C, left) may be a result of reduced KSHV infection and not necessarily due to the dependence of this induction on ROS. To address this potential confounding variable, we performed a similar experiment where NAC or PDTC were not added to the cells until 16 hr p.i., a time when latency has already been established. Eight hours after NAC or PDTC addition (24 hr p.i.), the cells were immunoblotted for Nrf2 activity. NAC treatment of the latently-infected cells inhibited tNrf2 and pNrf2 induction by 62% and 46%, respectively, while PDTC inhibited their induction by 23% and 56%, respectively (Fig. S2C, lanes 5–8). These results clearly demonstrated the importance of ROS elevation in Nrf2 activation during early KSHV latency (Fig. S2C, compare lanes 7 and 8 to 5). It is important to note that there was still some induction of pNrf2 and tNrf2 above the basal levels in uninfected cells, which suggested additional, ROS-independent mechanisms of induction at this time point.
Taken together, these results demonstrated that KSHV-mediated ROS induction is essential for Nrf2 upregulation during the early stages of infection, and although ROS are important for latent Nrf2 induction, additional pathways are likely involved.
Since ROS were involved in KSHV-mediated Nrf2 upregulation, we next determined the effect of KSHV infection on Keap1, the centerpiece of the ROS-dependent Nrf2-ubiquitination machinery. As expected, KSHV infection induced tNrf2 and pNrf2 (Fig. 4D, top two panels). Simultaneously, we observed a decrease in Keap1 levels to 0.8 and 0.3-fold at 0.5 and 2 hr p.i. when compared to uninfected cells (Fig. 4D, third panel, lanes 1–3). The levels of Keap1 were restored to that of uninfected cells at 24 hr p.i. (Fig. 4D, third panel, lanes 4–5) and real-time RT-PCR analysis suggested that this could be due to increased levels of Keap1 mRNA (Fig. 4E). This positive regulatory feedback loop that exists between Nrf2 and the Keap1 gene promoter has been previously described [90].
Although a decrease in absolute Keap1 levels is a strong indicator of its decreased inhibitory activity on Nrf2, we wanted to determine how de novo KSHV infection affects Nrf2-Keap1 interaction per se. According to the “open-closed conformation” model (Fig. S1), if KSHV disrupts Keap1-Cul3 cycling, we would not expect an increase in the interaction levels between Nrf2 and Keap1 despite an increase in total Nrf2 levels, because Keap1 molecules are complexed with already ubiquitinated Nrf2. To assess this, we performed co-immunoprecipitation (co-IP) studies by pulling down with anti-Keap1 antibody and Western blotted for Nrf2. When normalizing to whole cell lysate (WCL) β-tubulin, we observed a decrease in co-IPed Nrf2 by 13, 40 and 2% at 0.5, 2 and 24 hr p.i., (Fig. 4F, first and second panel, and quantification in Fig. 4G, top graph-black bars). Moreover, when accounting for robust WCL Nrf2 induction, the relative interaction between Keap1-Nrf2 decreased 65%, 84% and 43% at 0.5, 2 and 24 hr p.i., respectively (Fig. 4G, top graph-white bars). Such findings demonstrated that new Nrf2 is being generated and is not being bound and targeted by the Keap1-Cul3 inhibitory axis.
To further demonstrate that these events lead to decreased Nrf2 degradation, we determined the levels of lysine-48 (K-48) ubiquitination of Nrf2, a marker for its proteasomal degradation, and compared to the overall induction of Nrf2 in the WCL. As expected, when normalized to β-tubulin, the ubiquitination levels of Nrf2 increased slightly by 30, 24 and 10% at 0.5, 2 and 24 hr p.i., respectively (Fig. 4F, third panel, and Fig. 4G, bottom graph-black bars), likely due to the increased amount of Nrf2 pulled down (Fig. 4F, fourth panel). As a consequence, when we normalized to the total Nrf2 pulled down, relative K-48 ubiquitination levels decreased by 48, 68 and 39% at the same time points (Fig. 4G, bottom graph-white bars), indicating that although KSHV infection results in a slight increase in Nrf2-Keap1 pulled down, a lesser fraction of it is in the ubiquitinated form.
Collectively, these Nrf2-Keap1 pulldown experiments suggested that KSHV infection likely strengthens the already existent, ubiquitinated Nrf2 interaction with Keap1, resulting in Keap1 saturation and possible degradation, while concomitantly allowing the newly-translated Nrf2 to rapidly accumulate in the non-ubiquitinated and active form.
To determine if KSHV interaction with its cell-surface binding and entry receptors is necessary for Nrf2 activation, we incubated KSHV with heparin, which interacts with the virus envelope glycoproteins and prevents their interaction with heparan sulfate, thereby blocking KSHV-mediated binding and associated signaling [7]–[9], [11]. While infection with KSHV for 2 hr induced tNrf2 and pNrf2 by 1.9 and 2.6-fold, respectively, infection with heparin-treated KSHV or heparin alone did not induce any significant changes in Nrf2 (Fig. 5A). In addition, although pNrf2 accumulated in the nuclei of HMVEC-d cells during KSHV infection, heparin-treated KSHV and heparin alone did not induce nuclear accumulation of pNrf2 when observed by IFA (Fig. S3). These results demonstrated that KSHV interaction with the cell-surface receptors is essential in inducing the observed Nrf2 activity early during de novo infection, and provided additional verification of the specificity of KSHV-dependent Nrf2 induction.
De novo KSHV infection initiates a signaling cascade that induces many kinases [6]. Induction of FAK, Src, PI3-K/Akt, Rho-GTPases, EphA2 and ROS is important for KSHV macropinocytosis and entry, whereas induction of PKC-ζ, ERK1/2 and NF-κB is required for proper establishment of viral gene expression [14]–[21]. Because we observed increased Ser-40 phosphorylation of Nrf2 beyond the constitutive phosphorylation induced by H2O2 (Fig. 2), we hypothesized that one or more of these kinases induced by KSHV could induce this phosphorylation. Interestingly, atypical PKCs, PI3-K, ERK1/2 and Src have also been reported to induce Nrf2 phosphorylation in various cellular models, so we decided to assess the effect of their inhibition on Nrf2 phosphorylation during KSHV infection [67]–[72], [91]–[93]. To determine this, we pretreated cells with specific kinase inhibitors for 1 hr prior to infection with KSHV for 30 min, at which point the cells were Western blotted for tNrf2 and pNrf2 levels.
To assess if the observed KSHV-mediated Nrf2 accumulation, phosphorylation, and nuclear accumulation was transcriptionally active, we first assessed the DNA-binding activity of Nrf2. To assess this, we isolated nuclear protein from HMVEC-d cells infected for 8 and 24 hr and performed an Nrf2 ELISA assay. The Nrf2 ELISA assay measures the binding activity of transcriptionally active Nrf2 to the ARE sequence located on specific oligonucleotides immobilized at the bottom of each well. Upon addition of equal amounts of nuclear lysate per condition, calorimetric measurements provide a relative measure of the binding activity of Nrf2 in each sample, which increases linearly with increased Nrf2 binding affinity. As expected, infection with KSHV induced Nrf2 DNA-binding activity by ∼3-fold at 8 and 24 hr p.i. when compared to uninfected cells (Fig. 6A, top panel). A Western blot of the lysates is shown for quality control (Fig. 6A, lower panel).
Nrf2 transcriptionally activates genes involved in a variety of functions such as ROS homeostasis, apoptosis, cell migration, angiogenesis and drug resistance. KSHV-mediated Nrf2 induction may prime the cell and provide an environment conducive to KSHV infection, especially considering that several Nrf2-target genes are induced during de novo KSHV infection. To determine what particular host genes KSHV infection induces in an Nrf2-dependent manner, we created HMVEC-d cells deficient in Nrf2 by transducing them with lentiviral vectors expressing either short hairpin RNA against the Renilla luciferase mRNA (shRL–control) or short hairpin RNA against Nrf2 mRNA (shNrf2). These cells were then infected with KSHV for various time points and the level of Nrf2 activity was assessed for each condition. To assess the efficiency of the knockdown, we performed real-time RT-PCR analysis using Nrf2-specific primers. We observed a consistent ∼50–60% reduction in Nrf2 mRNA in shNrf2 cells compared to shRL cells during the whole course of infection (Fig. 6B). To further verify the knockdown efficiency, Western blot analysis was carried out to determine the protein levels of Nrf2. We observed a drastic reduction in Nrf2 protein levels in shNrf2 cells compared to shRL cells (Fig. S4A). To assess the specificity of the knockdown, we determined the levels of ERK1/2 and β-actin, which did not reveal any significant variation between the conditions (Fig. S4A).
Next, we determined the effect of the knockdown on the induction of several well-characterized Nrf2 target genes involved in ROS homeostasis. NAD(P)H quinone oxidase 1 (NQO1) is an important anti-oxidative agent involved in the clearance of quinones and hydroquinones, and is considered a reliable reporter of Nrf2 transcriptional activity [34], [45], [87]. Western blot analysis revealed that KSHV infection of shRL cells induced NQO1 levels by 2.5 and 1.9-fold at 8 and 24 hr p.i., respectively (Fig. 6C, third panel). Infection of shNrf2 cells, on the other hand, revealed significantly lower basal and induced NQO1 levels (Fig. 6C, third panel). These induction levels closely mirrored the induction and knockdown levels of Nrf2 as observed during the same time points (Fig. 6C, top two blots). Real-time RT-PCR analysis also showed that KSHV infection of shRL cells induced NQO1 mRNA to a significantly higher extent than during the infection of shNrf2 cells (Fig. 6D, left panel, compare black and red bars). Gamma-glutamylcysteine-synthase (GCS) is an essential component in the machinery that synthesizes glutathione, and a well-known Nrf2 target gene [34], [42], [87]. KSHV infection of shRL and shNrf2 cells induced GCS mRNA at 8 hr p.i. by 1.6 and 1.4-fold, respectively, showing that Nrf2 induction was not important for KSHV-mediated GCS upregulation at this time of infection (Fig. 6D, middle panel). At 24 hr p.i., however, KSHV induced GCS expression only in shRL cells (1.5-fold), while failing to do so in shNrf2 cells, indicating that at this time point, KSHV-mediated Nrf2 activation is required for the induction of GCS (Fig. 6D, middle panel). Heme oxygenase 1 (HO1), another well-characterized Nrf2 target gene involved in heme metabolism [34], [43], [44], [87], was significantly induced by KSHV infection of shRL cells (2.1-fold) at 8 hr p.i., but not shNrf2 cells (1.1-fold) (Fig. 6D, right panel). KSHV infection failed to induce HO1 in either condition at 24 hr p.i. (Fig. 6D, right panel).
In addition to stress-related genes, we observed other genes that are induced by de novo KSHV infection and that have also been shown recently to be Nrf2 target genes. Specifically, we observed that KSHV infection induced the anti-apoptotic protein Bcl-2 by ∼2-fold during the course of infection of shRL cells, and such induction was abrogated in the infection of shNrf2 cells (Fig. S4B). Glucose 6-phosphate dehydrogenase (G6PD), transaldolase (TALDO) and transketolase (TKT), three important molecules essential in the pentose phosphate pathway and nucleotide synthesis (Fig. S4D), were induced by KSHV, and this induction was Nrf2-dependent (Fig. S4C), as recently shown by Yamamoto et al (2012) [49]. Nrf2 knockdown did not significantly alter KSHV-mediated NF-κB activation, ruling out a role for Nrf2 in this induction during de novo infection (Fig. S5).
Collectively, these results clearly demonstrated that KSHV-mediated Nrf2 activation induces previously-described Nrf2 target genes, and that such activation can be severely abrogated with lentiviral vectors expressing shNrf2.
Vascular endothelial growth factor (VEGF), an angiogenic factor, plays an essential role in the formation of KSHV pathogenesis and KS histopathology along with angiogenin by mediating neovascularization and proliferation [97]–[104]. We have shown before that KSHV infection of HMVEC-d cells induces VEGF expression and secretion [104]. While certain KSHV genes have been implicated in VEGF induction, the mechanisms of VEGF induction during de novo infection have not been elucidated. According to recent studies, Nrf2 knockdown results in decreased angiogenesis in colonic adenocarcinoma cells and endothelial tube formation assays, both due to decreased expression of HIF-1α [52]–[54]. We examined whether Nrf2 was responsible for the induction of three important members of the VEGF family (VEGF-A, C and D). As expected, KSHV infection of shRL cells increased VEGF-A gene expression by 2.8 and 2.0-fold at 8 and 24 hr p.i., respectively (Fig. 7A, red bars). Upregulation in VEGF-A gene expression was dependent on Nrf2 at 8 hr p.i., since infection of shNrf2 cells did not show any increase at this time, but was independent of Nrf2 at 24 hr p.i. as it showed a 2.4-fold upregulation at this time (Fig. 7A, red bars). KSHV infection of shRL cells induced VEGF-C only by 1.2-fold at 8 hr p.i., and no effect (1-fold) at 24 hr p.i. (Fig. 7A, white bars). More interestingly, we identified a new VEGF, VEGF-D, which was induced by 3.6 and 1.7-fold at 8 and 24 hr p.i., respectively, in shRL cells (Fig. 7A, black bars). Only a mild induction was observed during the infection of shNrf2 cells at 8 hr p.i., but a 2-fold induction was observed 24 hr p.i., suggesting Nrf2 dependence only in the earlier stages of infection (Fig. 7A, black bars).
To determine if VEGF-A secretion levels were dependent on Nrf2 expression similarly to its gene expression, we performed a VEGF ELISA on the supernatant of infected cells. KSHV-infected HMVEC-d cells had elevated VEGF-A levels in their supernatants compared to uninfected cells at 8 hr p.i. (Fig. 7B). While infection of shRL cells exhibited a steady increase in VEGF-A supernatant levels at 8 and 24 hr p.i. (Fig. 7C, black line), infection of shNrf2 displayed only a moderate increase in VEGF-A supernatant levels (Fig. 7C, red line), further confirming the importance of Nrf2 in VEGF-A expression. To verify that the knockdown in the shRL and shNrf2 cells was successful, we performed a real-time RT-PCR analysis using Nrf2-specific primers (Fig. 7D).
These results demonstrated that KSHV-mediated Nrf2 induction early during KSHV infection plays an important role in the expression and secretion of two important members of the pro-angiogenic VEGF family. Moreover, these data also demonstrated for the first time that de novo KSHV infection of HMVEC-d cells induces VEGF-D expression.
We next determined the identity of additional genes that are important in KSHV biology for which Nrf2 activity is indispensable. Cyclooxygenase-2 (COX-2) is an important pro-inflammatory enzyme that catalyzes the conversion of arachidonic acid to prostaglandin H2 (PGH2), a precursor for the synthesis of several prostaglandins, including PGE2 [105]. Our earlier studies have demonstrated that KSHV induces COX-2 expression, which results in elevated levels of PGE2 formation and secretion [25], [106], [107]. These studies also showed that NFAT and CREB are two important transcription factors in inducing the COX-2 promoter, but did not exclude additional factors that may be required for COX-2 induction by KSHV [108]. To our surprise, when COX-2 mRNA was included in real-time RT-PCR analysis as a positive marker for KSHV infection, we observed that shNrf2 knockdown of the cells significantly abrogated the KSHV-mediated COX-2 induction. The induction of COX-2 in shRL cells was 7.0, 2.4 and 1.7-fold at 2, 8 and 24 hr p.i., respectively, contrasting with the modest 3.6, 1.3 and 0.5-fold induction (>50% reduction) in shNrf2 cells (Fig. 8A). Western blot analysis was consistent with the PCR data, showing a steady increase in COX-2 protein levels during the infection of shRL cells, but no observable increase in COX-2 during KSHV infection of shNrf2 cells (Fig. 8B). As expected, no obvious changes in the pattern of expression of the constitutively expressed COX-1 protein were observed (Fig. 8B).
Because Nrf2 is a transcription factor, we explored the possibility of Nrf2 playing an important role in COX-2 induction by direct binding to the COX-2 promoter. When we analyzed the sequence of the COX-2 promoter, we identified two Nrf2 binding sites (ARE-consensus sequence - TGAnnnnGC) on the template strand (Fig. 8C, red arrows). Moreover, the coding strand also provided a sequence pattern containing multiple ARE-like domains (grey arrows), which contained either an extra or missing base-pair between the required TGA and GC (TGAnnnnnGC or TGAnnnGC). These results are in concordance with results in Figs. 8A and B, and suggested that Nrf2 likely induces COX-2 expression through its transcriptional properties.
Sharma-Walia et al have demonstrated that KSHV-mediated COX-2 induction, either during entry or during latency, results in increased release of PGE2 [106], while Arun et al have demonstrated that such secretion mediates signaling in paracrine and autocrine manners through prostaglandin receptors 1–4 (EP 1–4) [97], [106]. Among such signaling pathways is PKC-ζ, which we determined to be important for Nrf2 induction during the early stages of KSHV infection (Fig. 5F) [106]. We hypothesized that PGE2 secretion during the post-entry stages of infection might induce PKC-ζ, which may then lead to sustained Nrf2 induction even post-KSHV entry. We therefore performed a Western blot analysis on the levels of Nrf2 in HMVEC-d cells that were treated with PGE2 for 4 hr, where we observed a robust, dose-dependent tNrf2 and pNrf2 induction (Fig. 8D). Real-time RT-PCR analysis showed that this induction was not mediated at the transcriptional level, as Nrf2 mRNA expression was not significantly affected by PGE2 treatment (Fig. 8E), leading us to believe that signaling pathways mediating Nrf2 stabilization and phosphorylation were involved. We, therefore, pretreated HMVEC-d cells with Myr-PKC-ζ for 1 hr prior to addition of PGE2 for an additional 4 hr. Interestingly, although PGE2 alone was able to induce tNrf2 and pNrf2 by 3.0 and 3.4-fold, respectively, Myr-PKC-ζ was able to fully abrogate such induction (Fig. 8F), establishing PKC-ζ as a key agent in PGE2-mediated Nrf2 activation.
Celecoxib is a specific COX-2 inhibitor that we have previously shown to abrogate PGE2 production during KSHV infection [109]. We utilized the inhibitory effects of Celecoxib to determine if endogenous COX-2 activation and PGE2 release induced by KSHV infection are important for Nrf2 activation. We, therefore, infected HMVEC-d cells in the absence or presence of Celecoxib and assessed Nrf2 levels (Fig. 8G). Infection of mock-treated cells induced Nrf2 accumulation with a pattern similar to that observed in figure 2A, with a robust steady raise until 2 hr p.i., a small dip at 4 hr p.i., and another increase by 8 hr p.i. (Fig. 8G, lanes 1–5). Interestingly, infection of Celecoxib-treated cells induced tNrf2 accumulation and its phosphorylation during the earlier stages (30 min p.i.) of infection, but this induction did not persist throughout the monitored period of 8 hr p.i. (Fig. 8G, lanes 6–10). When we plotted the fold induction of tNrf2 and pNrf2 between these two conditions, we observed no difference at 30 min p.i., but observed a >70% fold reduction in Nrf2 activation in Celecoxib-treated cells compared to mock-treated cells from 2–8 hr p.i. (Fig. 8H).
Collectively, these results demonstrated that KSHV-induced COX-2 and PGE2 autocrine/paracrine signaling activate Nrf2 during the post-entry stages of de novo infection of HMVEC-d cells through induction of a signaling cascade that requires PKC-ζ.
Because COX-2 is also induced by KSHV latency, we wanted to determine if latent gene expression is important for Nrf2 induction during the later stages of de novo infection. Ultraviolet (UV) light treatment of KSHV creates thymidine dimers between adjacent thymine residues of its DNA, and this abolishes its ability to replicate or properly express its genome. This process, however, does not affect the envelope and capsid of the virion, creating a virus that is incapable of establishing proper latent infection, but still capable of inducing early, gene expression-independent events propagated by virion-host interaction [24]. At 2 hr p.i., UV-KSHV induced tNrf2 and pNrf2 by 3.7 and 3.2-fold, respectively, and this was comparable to the induction observed by the untreated virus (Fig. 9A, lanes 1–3). However, we did not observe any induction in Nrf2 activity 24 hr p.i. with UV-KSHV, in contrast to such activation with untreated virus (Fig. 9A, lanes 4–6). IFA analysis also showed that while UV-KSHV could induce nuclear accumulation of pNrf2 at 2 hr p.i., no such effects were seen at 24 hr p.i. (Fig. S6A). These studies demonstrated that KSHV binding to HMVEC-d cell-surface receptors is essential in inducing Nrf2 activity during the early stages of de novo KSHV infection, while expression of latent genes was involved in Nrf2 induction during the late stages of infection.
To assess which KSHV latent gene(s) was responsible for the observed Nrf2 induction, HMVEC-d cells were transduced with lentiviruses expressing four known latent KSHV genes (ORFs 71, 72, 73 and K12) along with two vector controls (pSIN A and pSIN B). Real-time RT- PCR analysis revealed that all latent genes were successfully expressed in their respective transductions (Fig. S6B). Of the four KSHV genes, only transduction with ORF71 (vFLIP) significantly upregulated NQO1 and HO1 mRNA expression, suggesting that vFLIP might be an important inducer of Nrf2 activity during later stages of infection (Fig. 9B). We also observed significant IL-6 upregulation by vFLIP, further verifying the proper activity and specificity of vFLIP expression given its well-known inducing effects on NF-κB (Fig. 9C). Real-time RT-PCR assessment of Nrf2 transcript did not show any significant changes during vFLIP expression (Fig. 9D), indicating that vFLIP induces the Nrf2 axis through non-transcriptional pathways. Moreover, HEK 293T cells transfected with ORF71 showed increased tNrf2 and pNrf2 levels as measured by Western blot analysis (Fig. 9E).
These studies demonstrated that expression of latent genes was involved in Nrf2 induction, and that ORF71 (vFLIP) could be one of the agents responsible for such induction.
Results described above in Fig. 9 demonstrated that latent KSHV induces Nrf2 activity in the whole population of infected cells. At the single-cell level, vFLIP expression may affect Nrf2 activity either i) by inducing intracellular signaling pathways that affect Nrf2 activity within the same cell or ii) by inducing secretion of cytokines (PGE2) and other signaling ligands, which, through autocrine and paracrine signaling, can induce Nrf2 activity in adjacent cells. In the first scenario, only infected cells would exhibit enhanced Nrf2 activity, whereas in the later, both infected and adjacent uninfected cells would exhibit enhanced Nrf2 levels. To differentiate between these two possibilities, we performed an IFA on HMVEC-d cells infected with 5-bromo-2-deoxyuridine (BrdU)-labeled KSHV. Such a technique allowed us to determine precisely which cells were infected with KSHV by staining with an anti-BrdU antibody, while simultaneously observing pNrf2 activity by staining with anti-pNrf2 antibody.
As expected, the cells in the uninfected condition (negative control) showed no BrdU staining and low levels of pNrf2 in their nuclei (Fig. 10A, top row). In contrast, cells in the BrdU-KSHV condition showed significant BrdU staining (Fig. 10A, bottom 3 rows, second column). Colocalization of pNrf2 with BrdU-labeled KSHV genome was readily observed in the infected cell nuclei (Fig. 10A, bottom 3 rows, enlarged box, yellow arrows). More interestingly, in the infected condition, all the cells (14/14 or 100%) that stained for BrdU-KSHV also demonstrated high levels of pNrf2 in their nuclei (Fig. 10A, bottom 3 rows, first column, red arrows). Most uninfected cells (9/13 or ∼70%) showed very low to undetectable levels of pNrf2 staining (white arrows). However, a minority of uninfected cells (4/13 or ∼30%) showed some increase in pNrf2 activity (blue arrows), and these cells were often adjacent to the infected cells, possibly influenced by their paracrine effect. We confirmed that BrdU-KSHV infection induces pNrf2 in the nuclei of the cells through confocal imaging, which provides better resolution. Specifically, two cells, one positive and one negative for BrdU-KSHV staining, were visualized within the same field (Fig. 10B). Corroborating the IFA data, the infected cell exhibited substantially elevated levels of pNrf2 in the nucleus compared to the adjacent uninfected cell (Fig. 10B).
These results demonstrated that viral gene expression plays an important role in inducing pNrf2 levels within the infected cell, while also suggesting that an autocrine/paracrine pathway aids in this induction.
Using IFA we observed substantial colocalization of BrdU-KSHV and pNrf2 in the nuclei of infected cells (Fig. 10A, right-most column, yellow arrows). Because of the low resolution provided by IFA, we confirmed the findings using confocal microscopy, which also revealed multiple colocalization spots between the BrdU-KSHV genome and pNrf2 (Fig. 10C, merged, yellow squares).
We further performed a proximity ligation assay (PLA) between pNrf2 and LANA-1, a technique that can identify interactions between proteins that are <16 nm apart. Compared to uninfected cells, which showed essentially no PLA staining, HMVEC-d cells infected with KSHV for 24 hr showed significant staining (Fig. 11A, white arrows). When quantified, we observed ∼20 dots/nucleus in infected cells compared to the non-specific single dot/nucleus observed in uninfected cells (Fig. 11B), suggesting significant colocalization between LANA-1 and pNrf2.
We next determined whether pNrf2 and LANA-1 colocalized together on the KSHV genome or elsewhere in the nucleus, where they may affect gene expression. To determine possible triple-colocalization we performed PLA for pNrf2 and LANA-1 (green dots) and stained the EdU-labeled virus (red) (Fig. 11C). Using the BrdU-labeled KSHV was not feasible for this experiment because the IFA procedure for BrdU labeling requires DNA denaturation, a process that interferes with the PLA procedure. Interestingly, we observed that about half of the pNrf2-LANA-1 PLA puncta colocalized with the EdU-KSHV genome (Fig. 11C, bottom panel, white arrows), suggesting a possible interaction of all these three agents. In addition, we also observed the pNrf2 and LANA-1 PLA spots in areas of the nucleus where no EdU-KSHV genome was detected (Fig. 11C, bottom panel, blue arrows) which could be in part representing a complex potentially involved in the modulation of host gene expression.
These results suggested that Nrf2 interacts with the KSHV genome by utilizing one of its major regulatory proteins, LANA-1, as an intermediary.
Since Nrf2 associated with the KSHV genome, we wanted to determine its effects on KSHV biology. We first performed a KSHV entry assay on infected shNrf2 and shRL cells and observed that shNrf2-mediated knockdown did not affect the entry levels of KSHV (Fig. 12A). Real-time RT-PCR analysis of Nrf2 mRNA demonstrated that the lentiviral knockdown of shNrf2 cells was successful (Fig. S7A). We also performed the entry assay using chemicals that have been shown in multiple systems to affect Nrf2 activity. Trigonelline, a coffee extract alkaloid that has been shown to inhibit Nrf2 transcriptional activity by inducing its nuclear export [110], abolished NQO1 and GCS mRNA upregulation by KSHV infection as measured by real-time RT-PCR (Fig. S7C). Sulforaphane and tBHQ, two anti-oxidants and well-characterized inducers of Nrf2 protein levels and activity [87], were both able to induce Nrf2 accumulation (Fig. S7D). Bay-11-7082, an NF-κB inhibitor, was included as a negative control as previous studies from our laboratory have shown that NF-κB is not involved in KSHV entry or nuclear delivery. We pretreated HMVEC-d cells with each of these chemicals for 4 hr prior to infection with KSHV for 30 min and performed a KSHV entry assay. Similar to the knockdown experiments, none of these compounds significantly affected KSHV entry (Fig. S7E), further demonstrating that Nrf2 activity does not play an important role in KSHV entry.
The results of the entry experiments were surprising, as we had anticipated Nrf2 knockdown to increase KSHV entry due to increased basal ROS levels. To test whether knockdown of Nrf2 in HMVEC-d cells results in elevated ROS, we performed a ROS measurement experiment in cells transduced with shNrf2 and shGFP (shRNA against green fluorescent protein–negative control). The use of shRL-expressing lentivirus used in most of the experiments was avoided in this experiment because the vector that expresses shRL also expresses GFP (used to monitor transduction efficiency), which has the same excitation/emission spectrum as the CM-H2DCFDA dye, and would interfere with ROS measurement. As we have previously shown, KSHV-infected HMVEC-d cells produced 2–3-fold more ROS than uninfected cells (Fig. S7F). To our surprise, Nrf2 knockdown did not affect the basal ROS levels in uninfected cells (Fig. 12B, compare shGFP U.I. with shNrf2 U.I.). Moreover, KSHV-mediated ROS induction was unchanged between the two cell conditions (Fig. 12B, compare shGFP+KSHV with shNrf2+KSHV). Similarly, Nrf2 knockdown did not affect basal ROS production or KSHV-induced ROS even at 24 hr p.i. (Fig. S7G). Real-time RT-PCR analysis of Nrf2 mRNA demonstrated that the lentiviral knockdown of shNrf2 cells was successful (Fig. S7B).
These results demonstrated that basal Nrf2-levels are dispensable for ROS homeostasis in HMVEC-d cells and that KSHV-mediated Nrf2 induction does not affect the induction of ROS during de novo infection, suggesting no role for Nrf2 in KSHV entry.
We next determined if Nrf2 activity plays a role in KSHV nuclear delivery. We infected shNrf2 and shRL cells with KSHV, isolated the nuclei of the cells 2 hr p.i., and performed real-time DNA PCR using ORF73-specific primers. Interestingly, we observed only a slight decrease (25%) in the nuclear delivery of KSHV in shNrf2 cells compared to control cells (Fig. 12C). Real-time RT-PCR analysis of Nrf2 mRNA demonstrated that the lentiviral knockdown of shNrf2 cells was successful (Fig. S8A). We further determined the degree of nuclear delivery in cells treated with the chemicals used in Fig. S7E, which did not affect the nuclear delivery of KSHV in HMVEC-d cells (Fig. S8B). These data suggested that Nrf2 does not play an important role in KSHV nuclear delivery during de novo KSHV infection.
Our earlier studies have determined that de novo KSHV infection of HMVEC-d cells inevitably results in the establishment of latency [26]. Indeed, we observed a steady increase in ORF73 gene expression (Fig. S8C), whose product, LANA-1, is essential for maintenance of viral latency [26]. However, before latency is established, a transient expression of lytic genes with anti-apoptotic and immune-evasive functions peaks at the earlier stages of infection and subsides as latency takes over [26]. We observed a robust expression of ORF50, the master regulator of lytic gene expression, which decreased by 24 hr p.i. (Fig. S8D).
Since Nrf2 knockdown did not affect KSHV entry or nuclear delivery, and because Nrf2 colocalized with LANA-1 on the KSHV genome (Fig. 11), we further wanted to determine if Nrf2 plays a role in the proper establishment of KSHV latency by affecting the normal course of viral gene expression during the early stages of de novo infection. As expected, infection of shRL cells produced a viral expression pattern similar to that of untransduced cells, with ORF50 peaking at 8 hr p.i., declining by 24 hr p.i., and becoming essentially undetectable by 48 hr p.i. (Fig. 12D, black bars). Interestingly, ORF50 expression during the infection of shNrf2 cells was reduced by 47 and 64% at 8 and 24 hr p.i., respectively (Fig. 12D, red bars). A Western blot analysis of ORF50 expression confirmed the decreased expression of ORF50 at 8 and 24 hr p.i, by 29 and 94% respectively (Fig. 12E). We further assessed the expression pattern of additional lytic genes previously shown to be expressed during early de novo infection, such as ORFK8, K5 and vIRF2, and observed decreased expression during infection of shNrf2 cells compared to shRL cells, suggesting an important role for Nrf2 in the expression of the early lytic KSHV genes during de novo infection of HMVEC-d cells (Fig. S9).
In contrast to lytic genes, ORF73 (LANA-1) gene expression was significantly upregulated by 3-fold at 24 hr p.i. and by 3.6-fold at 48 hr p.i. in shNrf2 cells compared to shRL cells (Fig. 12F). The knockdown for each condition was efficient (Fig. S8E), and KSHV infection was able to induce NQO1 gene expression in shRL cells, but not shNrf2 cells, as we previously determined (Fig. S8F).
To assess how Nrf2 knockdown increased LANA-1 expression, we performed an IFA analysis of LANA-1 during the infection of shRL and shNrf2 cells (Fig. 13A). As expected, the shRL condition consisted of cells expressing GFP, which is part of the shRL lentiviral vector used to measure transduction efficiency. In contrast, the shNrf2 lentiviral vector does not code for GFP. For the purposes of this experiment, fields from the shRL condition containing ∼50% transduction efficiency were obtained to more accurately assess if control vector (shRL) lentiviral transduction had any effects on subsequent KSHV infectivity. As observed by the quantification in Fig. 13B, the same percentage of GFP-positive cells and GFP-negative cells, ∼60%, were infected with KSHV, arguing against a possible role for the control lentiviral transduction in affecting subsequent KSHV infectivity. Furthermore, KSHV infected a similar portion of shRL cells compared to shNrf2 cells, indicating that Nrf2 knockdown does not affect KSHV infection rate (Fig. 13C). Such a finding is consistent with the entry and nuclear delivery assays in Fig. 12A and C. We then quantified the number of LANA-1 puncta per nucleus of each infected cell, and observed that shNrf2 nuclei exhibited, on average, ∼60% more LANA-1 puncta compared to infected shRL cell nuclei (Fig. 13D).
Collectively, these results suggest that Nrf2 knockdown does not affect infectivity with KSHV, but increases LANA-1 transcript expression and puncta formation within the infected cells.
KSHV is the etiological agent of KS, and the importance of oxidative stress in the development of KS pathogenesis has been shown before [83]. In this comprehensive study, we have demonstrated for the first time that de novo KSHV infection of endothelial cells induces the powerful transcription factor Nrf2. Moreover, our studies show that such Nrf2 upregulation is i) present in in vivo KS skin tissues, ii) dependent on ROS induction and signaling through Src, PI3-K and PKC-ζ, and iii) important for the expression of multiple host and viral genes involved in KSHV biology. We also show that Nrf2 induction is required for optimal COX-2 expression seen during KSHV infection, which through its enzymatic product PGE2 further induced Nrf2 activity, creating a feed-forward loop between these two molecules at later time points of infection (Fig. 14). For a better understanding, we have summarized the potential implications of the multiple roles that Nrf2 plays in KSHV biology in the following sections.
Although the importance of Keap1 in Nrf2 inhibition and degradation has long been known, the precise mechanism how this interaction is affected by ROS and other Nrf2 inducers was just recently elucidated. Oxidative stress, Sulforaphane and tBHQ stabilize Nrf2 not by directly dissociating it from Keap1, but by increasing the binding of existent, ubiquitinated Nrf2 with the Keap1-Cul3 ubiquitination axis, thus allowing newly synthesized Nrf2 to accumulate untargeted in the cell (Fig. S1B) [61], [62]. Kobayashi et al (2006) had shown that de novo Nrf2 protein synthesis was required for Nrf2 nuclear translocation and transcriptional activation by ROS, further validating this model [111]. Indeed, our results are in full concordance with this novel mechanism. KSHV infection of HMVEC-d cells did not significantly affect the levels of Nrf2 precipitated by Keap1 pull-down despite a robust increase in cellular Nrf2 protein levels, most of which was in an un-ubiquitinated form (Fig. 4F). This process was ROS-dependent, as inhibition of ROS with the NAC, PDTC and DPI abrogated KSHV-mediated Nrf2 induction. Collectively, these results strongly suggested that KSHV induces ROS, an event that thiolizes the cysteine residues located in the linker region of Keap1, disrupting the fluidity of the “closed-to-open” conformational cycling of the inhibitory machinery, saturating the Keap1-Cul3 system, and allowing Nrf2 to accumulate in the cytoplasm of HMVEC-d cells early during de novo KSHV infection.
Novel pathways that may affect Nrf2 stability are under intense scrutiny. Specifically, it has been shown that GSK-3β mediates phosphorylation of certain serine residues on the Neh6 domain of Nrf2, increasing its affinity for another E3 ubiquitin ligase, β-TrCP, which is believed to mediate Nrf2 ubiquitination and proteasomal degradation [40]. However, these studies were carried out in in vitro systems and the biological significance of such findings remains to be determined. Moreover, although it is well-known that during latency LANA-1 inhibits GSK-3β [112], we saw no effect of de novo infection on GSK-3β activity (Fig. S10), making this pathway unlikely to be involved in Nrf2 activation at the observed stage of infection.
KSHV infection of endothelial cells and fibroblasts induces the host's pre-existing Src, PI3-K/Akt, PKC-ζ and ERK1/2 signal pathways. This induction is important for several functions that aid in KSHV infection, including cytoskeletal rearrangements that help in virus entry and nuclear delivery, as well as viral gene expression [14]–[21]. We have also shown that ROS induction is essential for activation of FAK, Src and EphA2 receptor because pretreatment of the cells with NAC prior to KSHV infection abolished their induction [16]. Interestingly, similarly to kinase inhibitors (i.e. PP2, Wortmannin, LY294002 and Myr-PKC-ζ), ROS inhibitors such as NAC, PDTC and DPI also abolished Nrf2 phosphorylation and stability. The effect on total Nrf2 accumulation is most likely attributed to inhibition of the Keap1-Cul3 axis. The surprising effect of KSHV-induced ROS in Nrf2 phosphorylation is most likely attributed to its important role in amplifying the highly interdependent phosphorylation cascade.
Although Src, PI3-K and PKC-ζ activity was important in mediating Nrf2 activation, their inhibition affected Nrf2 behavior in distinct ways. Specifically, the Src-specific inhibitor PP2 affected Nrf2 accumulation and phosphorylation in both uninfected and infected cells, suggesting that not only is Src important for Nrf2 activity in infected cells, but also that it affects its constitutive phosphorylation levels in uninfected cells. PKC-ζ inhibition, on the other hand, did not affect total Nrf2 levels in either the infected or uninfected conditions, but it fully abolished Nrf2 phosphorylation in both cases. PI3-K inhibition did not affect Nrf2 levels or phosphorylation in uninfected cells. However, while KSHV infection of HMVEC-d cells strongly induced Nrf2 activity, infection of PI3-K-inhibited cells resulted in Nrf2 inhibition. This is a very interesting finding that can be used to affect the Nrf2 response during de novo infection, as modulation of a single agent, PI3-K, may be able to switch the fate of Nrf2 during de novo KSHV infection from induction to inhibition.
Our earlier studies have established a linear signaling pathway that starts with FAK and continues with Src, PI3-K, PKC-ζ and ERK1/2 in sequential order [6]. If the same linearity applied to the signaling cascade that mediates Nrf2 activity, one would expect that interruption of this pathway at any point during its course would have the same effect on Nrf2 stability and phosphorylation. However, given the different effects that the inhibition of Src, PI3-K and PKC-ζ kinases had on Nrf2 stability and phosphorylation (Fig. 5A–F), it is much more likely that this linear network branches off at certain sites to include additional factors important in Nrf2 activity. Interestingly, CKII, MAPK p38, JNK, etc., are kinases that may affect Nrf2 behavior, and further studies that are beyond the scope of the present one are required to decipher their role in KSHV-mediated Nrf2 activity.
In a previous study, we performed global host gene expression mapping at 2 and 4 hr p.i. and determined that KSHV infection of HMVEC-d cells upregulates expression of a myriad of genes involved in anti-apoptotic functions (Bcl2A1, Bcl-3), signal transduction (PKC and MAPK), cytokine signaling (IL-1β), angiogenesis (VEGF-A), and metabolism (COX-2, PFK-2) [113]. Expanding on such findings, in the current study we determined that Nrf2 was essential in upregulating the transcription of several KSHV-induced genes not previously identified, likely due to the time points selected in the previous investigations (2 and 4 hr p.i. in Naranatt et al (2004) vs. 8 and 24 hr p.i. in this study) [113]. Interestingly, several of the genes that KSHV induced in an Nrf2-dependent manner, such as anti-apoptosis (Bcl-2), metabolism (COX-2, G6PD, TALDO/TKT) and angiogenesis (VEGF-A and -D), fall in similar categories to those identified by Naranatt et al (2004) [113]. It is interesting to point out that Bcl-2 is an important anti-apoptotic factor and, like Bcl2A1, helps KSHV infection bypass apoptosis, an important barrier inherent in virus infection, as well as autophagy, a pathway well-known to antagonize KSHV infection [114].
We also identified three new KSHV-induced genes that play an important role in shunting glucose through the pentose phosphate pathway (PPP), TALDO, TKT and G6PD [49]. Such an induction could serve multiple functions, and a recent study by Mitsuishi et al (2012) showed that one of these functions is an increase in ribulose-5-phosphate, a crucial precursor in nucleotide synthesis [49]. This increase may provide a reservoir of nucleotides available for KSHV to utilize in i) the synthesis of host genes that are crucial in its infection or ii) the synthesis of viral genes during its early lytic burst (Fig. S4D for schematic of PPP).
Additionally, although Nrf2 was important for the transcription of several genes, they often followed different expression kinetics. For example, COX-2 and NQO1 peaked as early as 2 hr p.i., while HO1 showed no induction until 8 hr p.i. and subsided thereafter. VEGF and GCS showed Nrf2-dependence at 8 hr p.i. but seemed to be independent of Nrf2 activity at 24 hr p.i., while Bcl-2 steadily increased during the course of infection in an entirely Nrf2-dependent manner. These different patterns of transcript upregulation indicate that while Nrf2 plays an important role in the KSHV induction of these genes, assembly at the promoter of each gene is also likely affected by a cohort of gene-specific transcriptional cofactors that modulate the kinetics of each gene expression induced by Nrf2.
The importance of COX-2 in KSHV biology and pathogenesis has been extensively studied [25], [84], [85], [97], [105]–[107], [109], [115]. Its induction by KSHV and the subsequent elevation in PGE2 secretion and signaling has been shown to be important for KSHV latent gene expression and KS pathogenesis. Given the importance of COX-2 in its biology, it is not surprising that KSHV utilizes Nrf2 to induce its transcriptional levels. It was interesting to observe that PGE2, an enzymatic product of COX-2, induced Nrf2 stabilization and phosphorylation through PKC-ζ activation. This is an exciting finding that establishes a self-amplifying feed-forward loop between two important agents in KSHV biology. Indeed, an initial signaling event that is initiated by virion-host receptor interaction induces Nrf2 activity, which is necessary for COX-2 upregulation and PGE2 secretion, which further induces Nrf2 activity long after the original stimulus mediated by virus binding has subsided (Fig. 14, Phase II). Such sustained Nrf2 activity could additionally help with COX-2 transcription until, and likely even after, vFLIP expression initiates. Interestingly, vFLIP was able to induce transcription of Nrf2 target genes NQO1 and HO1 when lentivirally transduced in HMVEC-d cells. The exact mechanism of vFLIP-mediated Nrf2 induction remains elusive, but it is possible that such induction is dependent on the COX-2/PGE2 axis [109].
In long-term latent models, Ye et al (2011) showed that inhibiting ROS through NAC abolished TPA-induced lytic reactivation of KSHV [27]–[29]. In addition, they showed that addition of H2O2 and upregulation of endogenous ROS induced lytic reactivation of KSHV in latently infected endothelial and PEL cells, and that KSHV-induced signaling kinases were important for such activation [27]. However, the mechanisms of H2O2-induced lytic reactivation were not examined and remained unclear. Our preliminary viral gene expression profile could provide the missing link between H2O2 treatment and lytic reactivation. Our studies demonstrate that knockdown of Nrf2 results in decreased ORF50 and other lytic gene expression, especially at 8 hr p.i., and an increase in ORF73 expression, implicating Nrf2 as a positive factor for ORF50 expression (Fig. 12D–F). It could be possible that H2O2, which we showed to induce Nrf2 activity, and the kinases it induces, stabilize Nrf2 and induce its transcriptional activity on factors that upregulate lytic gene expression. It is very important to note that while these interpretations would involve Nrf2 activity as a lytic agent, such conclusions are preliminary, and current studies are underway to fully elucidate Nrf2 activity on latent and lytic gene expression in latent KSHV infection models. Moreover, the effects of Nrf2 on viral gene expression in endothelial and PEL cells may vary drastically, as developmental epigenetic modifications in these cell types are very different, leading to a completely different set of cellular cofactors present in their respective nuclei.
During de novo infection, virus gene expression is a dynamic process, and the role of Nrf2 on ORF73 and ORF50 expression may vary depending on time and context. Early, ORF50 (Rta) is expressed at high levels as soon as 30 min p.i., reaching a peak around 2–8 hr p.i., and subsequently declining to minimally detectable levels by 24 hr p.i. [26]. In contrast, only low levels of ORF73 (LANA-1) are detected around 2 hr p.i., but steadily increase over time, peaking around 24 hr p.i. and maintained during the observed period of 5 days p.i. [26]. Rta is shown to initiate ORF73 expression from the LTi (inducible) promoter, and subsequent LANA-1 accumulation promotes its own transcription via the LTc (constitutive) promoter while simultaneously repressing the ORF50 promoter and Rta production [5]. The low ORF50 expression in shNrf2-transduced cells and concomitant high expression of ORF73 suggest that during the early stages of KSHV infection of Nrf2-intact cells, Nrf2 could be playing a role in ORF50 induction. These events, in turn, regulate proper ORF73 expression. As LANA-1 accumulates, it could bind to Nrf2, as suggested by our PLA data (Fig. 11), possibly leading to conformational and functional changes in both proteins, modifying the positive effect that Nrf2 has on ORF50 expression. While ORF50 and other lytic gene expression in the absence of Nrf2 was significantly reduced (Fig. 12D–E and Fig. S7), their expression was still detected, likely due to the activity of other well-known ORF50 inducers such as NF-κB and ERK1/2, which in turn can induce ORF73 in an unregulated manner. Induction of very high levels of Nrf2 by H2O2 may lead to low LANA-1 and/or high ORF50 expression during de novo infection. Further studies are essential to dissect out the complexity of this scenario, which are beyond the scope of the current manuscript.
Overall, here we have demonstrated that KSHV induces Nrf2 during de novo infection of HMVEC-d cells through multiple mechanisms, and that this activation served an essential role in the induction of host and viral genes that are important in creating a microenvironment conducive to infection and establishment of latency (Fig. 14). Such knowledge, coupled to the fact that Nrf2 modulators such as Sulforaphane, tBHQ and Trigonelline are easy to administer orally, under clinical trials, and currently with no known side-effects, makes Nrf2 a very appealing target in the fight against KSHV infection.
HMVEC-d cells (CC-2543; Lonza Walkersville, Walkersville, MD) were cultured using endothelial basal medium (EBM2; Clonetics) supplied with growth factors (EGM2) necessary for petridish growth. Cell starvation was achieved by washing thoroughly with PBS and adding growth factor-free media for 8 hr prior to analysis. BCBL-1 cells, the source of KSHV virus used for HMVEC-d infection, were cultured in RPMI media supplied with 10% FBS and 1% penicillin-streptomycin solution as previously described [116]. Human Embryonic Kidney (HEK293T) cells were maintained in DMEM containing 1 mM pyruvate, 2 mM Glutamax, 50 u/ml penicillin, 50 mg/ml streptomycin, and 10% fetal serum (Fetalplex). Formalin-fixed, paraffin-embedded tissue samples from healthy subjects and patients with KS were obtained from the ACSR (AIDS and Cancer Specimen Resource, San Francisco, CA).
KSHV used for infection of primary HMVEC-d cells was extracted, isolated and purified from latently infected BCBL-1 cells after TPA induction as previously described [26], [116]. The quantity of KSHV DNA obtained after purification was quantified by real-time DNA PCR using KSHV ORF73 gene-specific primers as previously described [26].
The pre-sterilized thymidine analogue, 5-bromo-2-deoxyuridine (BrdU labeling reagent; Ref. 00103, Invitrogen) was added to BCBL-1 cells at a 1∶100 dilution along with TPA induction and then again 24 hr later, which allowed for the metabolic labeling of the KSHV genome during lytic production. The BrdU-labeled virus was purified from the supernatant of the treated BCBL-1 cells 5 days after TPA addition as previously described [26]. For IFA detection, the usual protocol was followed and an additional step of DNA denaturing by treatment with 4N HCl for 5 minutes was performed prior to primary antibody incubation. BrdU residues were detected using mouse anti-BrdU antibody.
BCBL-1 cells were induced with neomycin and treated with 10 µM EdU dissolved in DMSO. 72 hr later, an additional dose of 10 µM EdU was supplied to the media to facilitate additional labeling. An additional 72 hr later, EdU-labeled KSHV was isolated as previously described [26]. For IFA detection, after permeabilization with 0.2% Triton X-100, the slides were incubated with Click Reaction Buffer+Cu2SO4+Alexa-Fluor-labeled Picoyl Azide for 30 minutes in the dark at RT, as per manufacturer's protocol (Life Technologies, Grand Island, NY).
To prepare heparin-treated virus, KSHV was incubated with heparin at a final concentration of 20 µg/ml for 1 hr at 37°C, and then used for infection of HMVEC-d cells. UV-light treatment of KSHV was performed using UV-C light for 20 minutes at a distance of 10 cm.
Lentiviral vectors containing short hairpin RNA against Renilla luciferase (shRL), short hairpin RNA against green fluorescent protein (shGFP), short hairpin RNA against Nrf2 (shNrf2), ORF71, ORF72, ORF73, and ORFK12 were prepared using HEK293T cells as previously described [16], [117]. The supernatants containing each vector were used to transduce subconfluent HMVEC-d cells in the presence of polybrene (5 µg/ml). 24–48 hr later the cells were observed for transduction efficiency by using the transduction reporter (GFP) present in the shRL vector, and only experiments where shRL expression was present in >80% of the cells were investigated further. Knockdown was also verified using real-time RT-PCR analysis or WB for Nrf2.
Antibodies against tNrf2, NQO1, pPKC-ζ and pPKC-ζ, and Tyr-216 GSK-3β were obtained from Santa Cruz Biotechnology Inc., Santa Cruz, CA, whereas antibodies for tGSK-3β, Ser-9 GSK-3β, pNF-κB (Ser-536), ERK1/2 and pERK1/2 were obtained from Cell Signaling Technologies, Danvers, MA. pNrf2, Keap1 and p62 antibodies were obtained from Abcam, Boston, MA. Antibodies against β-actin, β-tubulin and TBP were obtained from Sigma-Aldrich, St. Louis, MO. The Ub-K48 antibody was from Millipore, Billerica, MA, and the mouse anti-BrdU antibody was from Life Technologies. ORF50 antibody was obtained from ABBIOTEC, San Diego, CA. HRP-linked anti-mouse and anti-rabbit antibodies used for chemiluminescent detection of Western blot bands were from KPL Inc, Gaithersburg, MD. 49,6-diamidino-2- phenylindole (DAPI), anti-rabbit and anti-mouse Alexa-Fluor 594 or 488 secondary antibodies were from Molecular Probes, Carlsbad, CA. Chemical inhibitors PP2, LY294002, Wortmannin, U73122, Myr-PKC-ζ, U0126, PGE2, and Bay-11-7082 were from Cayman Chemicals, Ann Arbor, MI. Nuclear Extract and TransAM Nrf2 Kits were from Active Motif, Carlsbad, CA. Nuclei EZ Prep Nuclei Isolation Kit was from Sigma-Aldrich. The Human VEGF Quantikine ELISA Kit was from R&D Systems Minneapolis, MN. Heparin, H2O2, N-acetylcysteine (NAC), pyrrolidine dithiocarbamate (PDTC) and Trigonelline were obtained from Sigma-Aldrich. The CM-H2CDFDA ROS-measuring dye was obtained from Life Technologies.
At the end of treatment, cells were suspended in RIPA Lysis Buffer (25 mM Tris-HCl, pH = 7.6, 150 mM NaCl, 1% NP-40, 0.1% SDS, and protease/phosphatase inhibitor cocktails). The lysates were sonicated to shear the DNA remnants and centrifuged at max-speed for 20 minutes at 4°C to rid the insoluble fractions. Protein concentration was assessed using BCA protein assay reagent (Pierce, Rockford, IL). Equal amounts of protein were separated by SDS-PAGE, transferred to a nitrocellulose membrane and probed using protein-specific primary and species-specific, HRP-conjugated, secondary antibodies. Detection was assessed by chemiluminescence assays (Pierce) according to manufacturer's protocol. The bands were digitalized using an Alpha-Imager System (Alpha Innotech Corporation, San Leonardo, CA) and quantified by ImageJ Software.
HMVEC-d cells were infected with KSHV (20 DNA copies/cell) at 37°C for different time points, and the nuclear and cytoplasmic proteins were fractioned per the manufacturer's protocol (Nuclear Isolation Kit, Active Motif). Briefly, cells were trypsinized, washed with PBS supplied with phosphatase inhibitor, lysed with 1× Hypotonic Buffer to disrupt the plasma membrane and collect the cytoplasmic fraction after pelleting the nuclei by centrifugation at 500×g for 2 minutes. The nuclear pellet was washed 2 additional times with 1× Hypotonic Buffer to remove any loosely bound cytoplasmic contaminants prior to resuspension in complete lysis buffer.
Cells were lysed either with denaturing RIPA buffer or non-denaturing buffer (NETM – 100 mM NaCl, 20 mM Tris-HCl, pH = 8.0, 0.5 mM EDTA and 0.5% NP-40) and 200 µg of protein was incubated with 1–2 µg of primary antibody, along with Protein G-sepharose beads. After overnight incubation at 4°C, the beads were pelleted by centrifugation, washed 3 times with the original lysis buffer, and the protein-bead complex was disrupted by resuspending them in SDS-Loading buffer solution and heating to 95°C for 10 minutes prior to SDS-PAGE gel loading and Western blot analysis.
Starved HMVEC-d cells were infected with KSHV (20 DNA copies/cell) for 30 minutes, washed with trypsin-EDTA to remove non-internalized virus, and total genomic and viral DNA was isolated using a DNeasy Blood and Tissue Kit (Qiagen) according to manufacturer's protocol. ORF73 copy number was quantified by real-time DNA PCR amplification from equal amounts of DNA from each condition as previously described [26]. The ORF73 gene cloned in the pGEM-T vector (Promega) was used to create a standard curve to determine the absolute DNA copy number of unknown conditions.
Starved HMVEC-d cells were infected with KSHV (20 DNA copies/cell) for 2 hr, washed with trypsin-EDTA, and subjected to nuclear isolation using a Nuclei EZ Prep Nuclei Isolation Kit (Sigma-Aldrich) as previously described [26]. Briefly, cells were lysed with mild lysis buffer and the nuclei were pelleted by centrifugation at 500×g for 5 minutes. The cytoplasmic contaminants that were loosely attached to the nuclei were removed by repeated washing with the mild lysis buffer. The DNA was isolated using the DNeasy Blood & Tissue Kit (Qiagen). Equal amounts of DNA were quantified with ORF73-specific primers by real-time DNA PCR. The absolute copy number was determined using an ORF73 standard curve as described in the entry experiment.
To assess the gene expression of the two major viral genes, ORF50 and ORF73, we extracted total RNA from KSHV-infected cells (50 DNA copies/cell) using RNeasy Mini Kit (Qiagen) spin columns. Equal amounts of RNA (2–4 µg), as determined by NanoDrop quantification, were subjected to one-step real-time RT-PCR analysis using ORF50- or ORF73-specific primers and Taqman probes (EZ RT-PCR core reagents, Applied Biosystems) as previously described [26]. The absolute copy number of each mRNA was assessed using ORF50 and ORF73-specific RNA standard curves obtained from in vitro-derived transcripts as previously described [26].
To assess changes in host gene expression, total RNA from infected cells was isolated using the RNeasy Mini Kit (Qiagen), and a cDNA library was created using a High-Capacity cDNA Reverse Transcription Kit (Life Technologies). The cycle threshold (Ct) values for each gene were determined using gene-specific primers and Sybr Green Probe-based real-time RT-PCR. ΔCt values relative to β-tubulin were assessed for each condition. Gene expression in the uninfected/untreated conditions was arbitrarily set to 1, and fold-induction was based on the ΔCt differential relative to this condition.
HMVEC-d cells grown in 8 chamber glass slides (Nalge Nunc International) were fixed with 4% paraformaldehyde for 15 minutes at room temperature and then permeabilized with 0.2% Triton X-100 for 5 minutes. The cells were then blocked with Image-iT FX signaling enhancer (Life Technologies) for 15 minutes prior to incubation with specific primary antibodies and fluorescent Alexa-Fluor-conjugated secondary antibodies (Alexa-Fluor) and mounting with DAPI for nuclear staining. Regular imaging was performed with Nikon imaging systems and figure analysis and deconvolution of the images was performed using the Metamorph software. Confocal imaging was performed using the Olympus FV10i microscope, and image analysis was performed using Fluoview1000 (Olympus) software.
To determine the level of VEGF secretion in infected cells, serum-starved HMVEC-d cells' supernatant was collected at various time points, spun at 1,000×rpm to remove any cellular debris, and frozen at −80°C until further use. Levels of the protein were assessed using a Human VEGF Quantikine ELISA Kit per manufacturer's protocol (R & D Systems).
To assess the DNA-binding activity of nuclear Nrf2 we first isolated the nuclear protein fraction using the Nuclear Extraction Kit (Active Motif), and performed Western blot analysis to assess the quality of the fraction by using the TATA Binding Protein (TBP) as a nuclear positive control, β-tubulin as a negative control, and actin as a loading control. The DNA-binding activity of Nrf2 was assessed using TransAM Nrf2 Kit as per manufacturer's instructions (Active Motif). Briefly, equal amounts of nuclear protein (15 µg/condition), as determined by BCA assay, were loaded in each well containing the Nrf2-binding oligo probes (TGAnnnnGC) for 1 hr at room temperature, washed, and sandwiched with an HRP-conjugated anti-Nrf2 antibody for an additional hour prior to chemiluminescent assessment. Thirty µg of MCF-7 cell line nuclear extract was used as a positive control for the assay.
Starved HMVEC-d cells cultured in a 48-well plate were infected for various time points prior to assessing ROS levels as previously described [16]. Briefly, EBM2 medium containing 10 µM of the ROS-detecting dye 5-(and-6)-cholormethyl-2′,7′-dichlorohydrofluorescein diacetate, acetyl ester (CM-H2DCFDA [C6827]); Invitrogen) was added to untransduced, shRL and shNrf2 cells for 30 minutes prior to infection with KSHV (40 DNA copies/cell). Fluorescence measurement and calculations were performed per manufacturer's protocol as previously described [16].
PLA (DuoLink, Sigma-Aldrich) was performed using the DuoLink PLA Kit to detect protein–protein interactions using fluorescence microscopy as per manufacturer's protocol. Briefly, HMVEC-d cells were cultured and infected with KSHV (20 DNA copies/cell) for 24 hr in 8 chamber microscopic slides, fixed with 4% paraformaldehyde for 15 minutes at room temperature, permeabilized with 0.2% Triton X-100 and blocked with DuoLink blocking buffer for 30 minutes at 37°C. Cells were then incubated with primary antibodies against LANA-1 (mouse monoclonal) and pNrf2 (rabbit) diluted in DuoLink antibody diluents for 1 hr, washed and then further incubated for another hour at 37°C with species-specific PLA probes under hybridization conditions and in the presence of 2 additional oligonucleotides to facilitate the hybridization only in close proximity (<16 nm). A ligase was then added to join the two hybridized oligonucleotides to form a closed circle and initiate a rolling-circle amplification using the ligated circle as a template after adding an amplification solution to generate a concatemeric product extending from the oligonucleotide arm of the PLA probe. Lastly, a detection solution consisting of fluorescently labeled oligonucleotides was added, and the labeled oligonucleotides were hybridized to the concatemeric products. The signal was detected as distinct fluorescent dots in the Texas red channel and analyzed by fluorescence microscopy. Negative controls consisted of samples treated as described but with only secondary antibodies.
|
10.1371/journal.pcbi.1000338 | Discovering cis-Regulatory RNAs in Shewanella Genomes by Support Vector Machines | An increasing number of cis-regulatory RNA elements have been found to regulate gene expression post-transcriptionally in various biological processes in bacterial systems. Effective computational tools for large-scale identification of novel regulatory RNAs are strongly desired to facilitate our exploration of gene regulation mechanisms and regulatory networks. We present a new computational program named RSSVM (RNA Sampler+Support Vector Machine), which employs Support Vector Machines (SVMs) for efficient identification of functional RNA motifs from random RNA secondary structures. RSSVM uses a set of distinctive features to represent the common RNA secondary structure and structural alignment predicted by RNA Sampler, a tool for accurate common RNA secondary structure prediction, and is trained with functional RNAs from a variety of bacterial RNA motif/gene families covering a wide range of sequence identities. When tested on a large number of known and random RNA motifs, RSSVM shows a significantly higher sensitivity than other leading RNA identification programs while maintaining the same false positive rate. RSSVM performs particularly well on sets with low sequence identities. The combination of RNA Sampler and RSSVM provides a new, fast, and efficient pipeline for large-scale discovery of regulatory RNA motifs. We applied RSSVM to multiple Shewanella genomes and identified putative regulatory RNA motifs in the 5′ untranslated regions (UTRs) in S. oneidensis, an important bacterial organism with extraordinary respiratory and metal reducing abilities and great potential for bioremediation and alternative energy generation. From 1002 sets of 5′-UTRs of orthologous operons, we identified 166 putative regulatory RNA motifs, including 17 of the 19 known RNA motifs from Rfam, an additional 21 RNA motifs that are supported by literature evidence, 72 RNA motifs overlapping predicted transcription terminators or attenuators, and other candidate regulatory RNA motifs. Our study provides a list of promising novel regulatory RNA motifs potentially involved in post-transcriptional gene regulation. Combined with the previous cis-regulatory DNA motif study in S. oneidensis, this genome-wide discovery of cis-regulatory RNA motifs may offer more comprehensive views of gene regulation at a different level in this organism. The RSSVM software, predictions, and analysis results on Shewanella genomes are available at http://ural.wustl.edu/resources.html#RSSVM.
| RNA is remarkably versatile, acting not only as messengers to transfer genetic information from DNA to protein but also as critical structural components and catalytic enzymes in the cell. More intriguingly, RNA elements in messenger RNAs have been widely found in bacteria to control the expression of their downstream genes. The functions of these RNA elements are intrinsically linked to their secondary structures, which are usually conserved across multiple closely related species during evolution and often shared by genes in the same metabolic pathways. We developed a new computational approach to find putative functional RNA elements by looking for conserved RNA secondary structures that are distinguished from random RNA secondary structures in the orthologous RNA sequences from related species. We applied this approach to multiple Shewanella genomes and predicted putative regulatory RNA elements in Shewanella oneidensis, a bacterium that has extraordinary respiratory and metal reducing abilities and great potential for bioremediation and alternative energy generation. Our findings not only recovered many RNA elements that are known or supported by literature evidence but also included exciting novel RNA elements for further exploration.
| RNA is remarkably versatile [1],[2], acting not only as messengers to transfer genetic information from DNA to protein, but also as critical structural components [3] and catalytic enzymes [4],[5] in the cell. More intriguingly, non-coding RNAs (ncRNA) have been found to play important regulatory roles. They can mediate gene expression post-transcriptionally in two ways: one is to serve as trans-acting antisense RNAs, such as microRNAs, which hybridize with target mRNAs to silence their expression [6],[7]; the other is to form structural cis-elements in the mRNAs, such as riboswitches, which regulate gene expression by mediating transcription termination or translation initiation [8],[9]. The regulatory roles of ncRNAs make them promising drug targets [10] and efficient tools for drug development and gene therapy [11],[12].
In the past a few years, many cis-regulatory RNA structural motifs have been identified in prokaryotes [13]–[15]. They are often located in the 5′ untranslated regions (UTR) of the mRNAs and can sense or interact with cognate factors, including proteins, RNAs, small metabolites, or even temperature changes, to mediate transcription attenuation [8], translation initiation [9], or mRNA stability [16]. The functions of the regulatory RNAs are intrinsically tied to their secondary structures, mostly recognizable as stem-loops or pseudoknots. Moreover, regulatory RNAs are often conserved during evolution: similar regulatory RNA elements can be shared by multiple co-regulated genes in the same metabolic pathway, or conserved in orthologous genes across closely related species [17].
Experimental screenings [18] for cis-regulatory RNAs are highly labor and time consuming. As demonstrated by previous studies [19],[20], a parallel way is to find good candidates computationally followed by targeted experimental validation. Because functional regulatory RNAs are often evolutionarily conserved in their secondary structures, we can identify them by finding significantly conserved RNA secondary structures in orthologous genes across closely related species. To accomplish this, we need two tools: one is to accurately predict common RNA secondary structures in multiple related sequences, and the other is to distinguish functional RNA secondary structures from random foldings of RNA sequences.
A number of algorithms have been developed for common RNA secondary structure prediction, such as RNAalifold [21], Dynalign [22], comRNA [23], CMFinder [24] and FoldAlign [25],[26]. We recently published a new algorithm, called RNA Sampler [27], for predicting common RNA secondary structures and structural alignments in multiple sequences. Both our study [27] and independent studies from other researchers [28],[29] have demonstrated that RNA Sampler provides more accurate structure predictions and generates better structural alignments on sequences of a wide range of identities than other leading software for similar purposes. Moreover, RNA Sampler runs fast and is feasible for common RNA secondary structure prediction on the genome scale.
Studies have shown that for a single sequence RNA secondary structure alone is not sufficient to distinguish functional RNA from random sequence [30],[31]. However, with the availability of multiple RNA sequences from related species, comparative genomics approaches provide additional power to identify functional RNA structures. One strategy is to design a scoring function for the predicted RNA secondary structures and examine the difference between the score distributions of real structures and randomly permutated structures, as employed by the RNA identification pipeline based on CMfinder [32] or comRNA [23]. But one limitation of such an approach is that the user needs to generate a large number of random sequence sets for each set of real sequences and doing structure predictions on these permutated sequence sets is usually time consuming. Besides, it can be difficult to find a score cutoff to make the call between functional and random RNAs. Another strategy is to train a classification model based on features that can distinguish common structures of known functional RNAs from those of random RNAs and then apply the classification model on the newly predicted common RNA structures to determine whether they are of functional or random RNAs. RNA classification algorithms employing this strategy include QRNA, RNAz and Dynalign+LIBSVM. QRNA [33] classifies a pairwise sequence alignment by the posterior probabilities of three probabilistic models, “RNA”, “Coding” and “Null” (position independent). RNAz [34] and Dynalign+LIBSVM [35] both employ support vector machines (SVM) to build the classification models. To train a classification model, the developer still needs to generate a large number of random sequence sets as the negative training sets and make structure predictions on them, but once the classification model is trained, the user can directly utilize the model to identify functional RNAs without the need to generate, and perform folding of, random sequences. The type of sequences used to train the classification models is essential to their classification performance on new sequences. QRNA and Dynalign+LIBSVM only use tRNAs and rRNAs in their training on RNA structures, and RNAz is trained on multiple RNA gene/motif families from the Rfam database but only uses sequence sets with high identities. To avoid overfitting the classification model to specific classes of RNAs, using training sets that cover a wide range of sequence identities and a variety of RNA families is more desirable. In addition, training the classification model using more accurately predicted RNA common structures and alignments is advantageous for more sensitive classification of functional RNAs from random ones. RNAz uses RNAalifold [36] for common RNA structure prediction. When using sequence alignments as its input, RNAalifold performs poorly in predicting RNA structures on sequence sets of low identities [27]. The structure prediction accuracy of RNAalifold may be improved by using structural alignments, but RNAz might need to be re-trained to use structural alignments.
In this paper, we present a new SVM based functional RNA identifier named RSSVM (RNA Sampler+Support Vector Machine). RSSVM applies a set of features to represent common RNA secondary structures and structural alignments generated by RNA Sampler, which predicts RNA structures more accurately than other approaches [27]–[29]. RSSVM is trained with RNA sets with a wide range of sequence identities from all bacterial RNA motif/gene families in the Rfam database [37]. RSSVM is more sensitive in identifying real functional RNAs than other leading RNA classification programs, including RNAz, Dynalign+LIBSVM and QRNA, at the same false positive rate. We applied RSSVM on multiple Shewanella genomes to identify putative cis-regulatory RNA motifs in the 5′-UTRs of orthologous genes.
Shewanella oneidensis is a facultative, gram-negative γ-proteobacterium. It has extraordinary abilities to use a wide variety of metals and organic molecules as electron acceptors in respiration [38]–[40], which gives it great potential to be applied in bioremediation of both metal and organic pollutants. The complete genomic sequences of Shewanella oneidensis and multiple other Shewanella species provide good resources for discovering cis-regulatory RNAs using comparative genomics approaches. Combining with the recent predictions of putative DNA cis-regulatory motifs in S. oneidensis [41], we will have a more complete view of gene regulation in S. oneidensis at different regulation levels.
We examined the performance of RSSVM in identifying RNA regulatory motifs on 1686 positive and 1686 negative test sequence sets (see Methods) and compared its performance with that of RNAz, Dynalign+LIBSVM and QRNA. Both Dynalign+LIBSVM and QRNA only work on two sequences, thus we examined their performance on all unique pairs of RNA sequences for each test set. The sensitivity and false positive rate (FPR) of the predictions were measured by the fractions of true positive classifications on the positive sets and false positive classifications on the negative sets, respectively. For each prediction, RSSVM, RNAz and Dynalign+SVM are able to report an SVM classification probability (P) which measures the confidence of the prediction. The higher the P-value, the more confident the prediction. A P-value cutoff can be selected to call positive predictions. When a lower P-value cutoff is used, although more regulatory RNAs can be identified from the positive sets, more negative test sets may be simultaneously misclassified as regulatory RNAs, leading to a higher false positive rate.
The prediction results from different SVM models at the same P-value cutoff are not readily comparable, because their corresponding sensitivities and false positive rates can be significantly different (Figure 1). Thus, to make fair comparisons, we always compare the performance of two programs at the same false positive rate which may be achieved by using different P-value cutoffs for different programs (Table S1). The Receiver Operating Characteristic (ROC) curves in Figure 1 demonstrate the prediction sensitivities of RSSVM, RNAz and Dynalign+LIBSVM at different FPRs. RSSVM and RNAz have similar sensitivities on all test sets when the FPR is lower than 0.01. However, when a higher FPR is allowed, RSSVM becomes more sensitive. At the FPR of 0.05, the sensitivities of RSSVM and RNAz are 0.86 and 0.75, respectively. We also compared the performance of RSSVM and RNAz on test sets whose average pairwise sequence identities are lower than 70%. On these test sets, RNAz only has slight improvement in sensitivity in the low FPR range comparing to its performance on all test sets. The prediction sensitivities of RSSVM, however, are about 10% higher than those on all test sets at the same FPRs. RSSVM is much more sensitive than RNAz at any FPR. At the FPR of 0.01, the sensitivity of RSSVM (0.77) is higher than that of RNAz (0.64) by 20% (Dataset S1). The higher prediction sensitivity than RNAz at the same FPR makes RSSVM an alternative choice for the whole genome RNA motif search, as it can find more targets while maintaining a low FPR.
At any FPR, Dynalign+LIBSVM has significantly lower sensitivities than RSSVM and RNAz on all test sets and on test sets with low identities, especially in the range of low FPRs (FPR<0.05) (Figure 1). At the FPR of 0.02, the sensitivities of Dynalign+LIBSVM are only 0.28 and 0.42 on all test sets and on test sets with low identities (<70%), respectively. The mediocre performance of Dynalign+LIBSVM in our tests may be attributed to the following reasons: 1) Dynalign+LIBSVM only uses information from two sequences, but RSSVM and RNAz take advantage of covariance information from multiple sequences; 2) Dynalign+LIBSVM was trained only on tRNAs and 5S rRNAs, which may cause overfitting of its classification model to these RNA families. In fact, we did observe a much higher classification sensitivity of Dynalign+LIBSVM on test sets comprising tRNAs and 5S rRNAs than on all test sets at the same FPRs (data not shown). For whole genome RNA motif scan, an ideal tool is required to have a high sensitivity and a low false positive rate. Dynalign+LIBSVM might not be a good choice for large scale scan of RNA motifs.
QRNA does not provide a similar measurement of P-value for its predictions, thus we are not able to generate its ROC curve. But on all test sets, the overall FPR of QRNA is 0.05. At this FPR, RSSVM has a significantly higher sensitivity (0.86) than QRNA (0.51) (Table S1).
We further evaluated the performance of RSSVM on test sets with different ranges of average sequence identities. We use correlation coefficient (CCclassification), the geometric mean of the classification sensitivity and (1−FPR), to measure the overall performance of RSSVM in each identity range. Because the overall FPR of QRNA on all test sets is 0.05, to make fair comparisons, we use different P-value cutoffs for RSSVM, RNAz and Dynalign+LIBSVM to achieve the same FPR of 0.05 on all test sets. As shown in Figure 2A, all algorithms have similar performance on test sets with high identities (≥70%), but RSSVM significantly outperforms all the other algorithms on test sets with low identities (<70%). In general, all tested algorithms tend to have lower FPRs on sequence sets with low identities (<70%) than with high identities (≥70%) (Table S1). The increases in FPRs on high-identity sets may be mainly due to the loss of covariant mutations in the structures. Although Dynalign+LIBSVM and QRNA have low FPRs on low-identity sets, they also make few positive predictions in those sets, leading to low sensitivities.
At the more stringent overall FPR of 0.01 on all test sets, RSSVM (0.68) and RNAz (0.65) have almost the same overall prediction sensitivity (Table S1), and both perform significantly better than Dynalign+LIBSVM, whose lowest possible overall FPR is 0.02 (Figure 2B). However, RSSVM and RNAz outperform each other in different identity ranges. RSSVM is much more sensitive on sequence sets with identities lower than 60%, but RNAz performs better on sequence sets with high identities (≥60%), while both algorithms maintain low FPRs in all identity ranges.
Overall, for the best performance, RNAz, Dynalign+LIBSVM and QRNA are in favor of sequence sets with high identities. RSSVM, however, has consistent and more sensitive performance on the low-identity sets while keeping the same FPRs. These programs can complement each other for the best performance in identifying regulatory RNAs on sequences with a wide range of identities.
Three major improvements may contribute to the better performance of RSSVM compared to RNAz in identifying regulatory RNAs, especially on test sets whose identities are lower than 70%. The first improvement is using the more accurately predicted common RNA secondary structures by RNA Sampler. The accuracy of predicted structures can be measured by the correlation coefficient of structure prediction (CCstructure), which approximates the geometric mean of the sensitivity and specificity of predicted base pairings [27]. RNA Sampler and RNAalifold are the corresponding core algorithms used by RSSVM and RNAz for predicting common RNA secondary structures, respectively. As shown in Figure 3, RNA Sampler gives similar performance to RNAalifold on the high-identity sequence sets (≥80%) but makes much more accurate structure predictions on the low-identity sets (<80%). The more accurately predicted structures and better alignments by RNA Sampler provide a better start point for RSSVM to identify RNA motifs. Second, the additional features used by RSSVM (see Methods), such as the SCI scores calculated based on common structures predicted by RNA Sampler, the information content (IC) which grasps the information of sequence conservation, and the mutual information (MI) which represents covariant mutations in the structural alignments, allow it to generate better SVM models to separate regulatory RNA motifs from shuffled ones, especially on sequence sets with low identities (Figure 1). Third, RSSVM is trained on sequence sets of a wider variety of RNA families and a broader range of sequence identities.
In addition, because the common structures predicted by RNA Sampler are more accurate in general, they may provide insightful hints for inferring the functions of the predicted RNA motifs and guiding the design of experimental validation.
As many known bacterial regulatory RNA sites are located in the 5′-UTR sequences and often conserved during evolution, we applied RSSVM, RNAz and QRNA on multiple Shewanella genomes to identify potential regulatory RNA motifs in the 5′-UTR regions. We retrieved 1002 sets of UTR sequences of orthologous genes from five related Shewanella genomes. The average pairwise sequence identities of the UTR sets range from 25% to 88%, with a mean of 45% and median 42%. The majority of the sequence sets are in the identity range of 40–70%, which is ideal for RSSVM to identify functional RNA motifs. We examined each set of UTR sequences in three overlapping windows that cover the regions of −250 to −100, −200 to −50, and −150 to 20 (1 corresponds to the translation start site). For each UTR set, we report the classification result from the window with the best SVM probability for RSSVM or RNAz. We chose P≥0.95 and P≥0.50 as the confidence probability cutoffs for RSSVM and RNAz, respectively, which give the same overall false positive rate of 0.01 on all test sets. For QRNA, we classified a set as regulatory RNA if more than two pairwise alignments of the sequences were identified as “RNA”.
The total numbers of predicted regulatory RNA motifs by different approaches are listed in Table 1. Of the 1002 orthologous UTR sets, RSSVM, RNAz and QRNA predicted 166, 109 and 112 putative regulatory RNA motifs, respectively. The sensitivities of the predictions can be estimated by the fraction of correctly predicted known RNA motifs/genes. By scanning the orthologous UTR sets with all known bacterial RNA motif models from the Rfam database using the RNA motif searching software Infernal [42], we obtained 19 known RNA motifs that gave infernal scores higher than 10 bits and occurred in at least two orthologous sequences of a UTR set. 6 of the 19 RNA motifs have orthologous sequences from S. oneidensis and E. coli in the Rfam seed alignments. RSSVM, RNAz and QRNA successfully detected 17, 16 and 11 of these 19 known RNA motifs, respectively, and the three approaches combined discovered 18 known RNA motifs (Table 2). It suggests that RSSVM and RNAz have similar sensitivities and both methods are able to discover more known motifs than QRNA. The one missed by all three approaches is the S15 mRNA leader sequence which contains alternative pseudoknot and stem-loop structures. If we slightly lower the P-value cutoff for RSSVM to 0.9, it is able to identify the RNA motif in the S15 UTR set. The success of identifying almost all known RNA motifs in the studied sequence sets demonstrates the high sensitivity of RSSVM.
The predictions by the three approaches overlap significantly with each other, as shown in the Venn diagram in Figure 4. 36 RNA motifs are identified by all three approaches, including 9 matching the known RNA motifs. This suggests that consensus predictions by all approaches may have high specificity. RSSVM and RNAz have additional 44 predicted motifs in common, and 6 of them are known motifs. QRNA has additional 11 and 7 motifs overlapping with the predictions by RSSVM and RNAz, respectively, including 2 matching known motifs. These results suggest that predictions cross-validated by different approaches are more likely to be real. Although a large fraction of the predictions by RSSVM and RNAz overlap, 2 and 1 known RNA motifs are identified only by RSSVM and RNAz, respectively, suggesting that combining predictions from different approaches may find more real RNA regulatory motifs. RSSVM made more predictions than RNAz. Besides the 80 predictions in common, 86 and 29 motifs were identified specifically by RSSVM or RNAz, respectively. The overall sequence identity of the commonly predicted sets by RSSVM and RNAz (mean 50%) is significantly higher than that of the predicted sets only by RSSVM or RNAz (mean 41%), with t-test p-values of 7×10−7 between the common and RSSVM specific predictions and 1×10−3 between the common and RNAz specific predictions, respectively. As seen in Figure S1, 87% of the sets predicted only by RSSVM or RNAz have sequence identities lower than 50%, while only 60% of the commonly predicted sets have identities lower than 50%. 5% more of the RNAz specific predictions are in the high-identity region (≥60%) than the RSSVM specific predictions. As demonstrated with the test sets, RNAz performs better on sequences of high identities, which is consistent with the observation that majority of the RNAz predictions, especially those in common with the RSSVM predictions, have higher identities than the RSSVM specific predictions. The fact that RSSVM gives more independent predictions than RNAz further demonstrates that RSSVM is more sensitive than RNAz on the low-identity sequence sets.
The specificity, the fraction of correct predictions, is difficult to accurately measure because of the poor knowledge on RNA motifs in S. oneidensis. We use the false positive predictions on shuffled sequences to evaluate whether the RNA motifs could be predicted by chance. The RNA Sampler structural alignments or ClustalW alignments of orthologous UTR sets were shuffled using the same approach that generated the negative training and test sets described in Methods and were used as negative controls for RSSVM and RNAz/QRNA, respectively. Both RSSVM and RNAz did not report any RNA motifs in these shuffled sequences, but QRNA had 13 false positive predictions. These results are consistent with the performance of these three approaches on the test sets, with QRNA tending to have more false positives than RSSVM and RNAz.
Besides predictions that match Rfam motifs, we can also assess the accuracy of our predictions by comparing them to other independent types of predictions and to published reports of regulatory motifs or genes undergoing post-transcriptional regulation.
We use the predicted regulatory RNA motif in front of the LeuA operon as an example to illustrate detailed analysis of the predicted RNA structures. The predicted alternative structures in the 5′-UTR of the LeuA operon are shown in Figure 5A. Our predicted alternative structures match the majority of the previously proposed structures [48], including the terminator stem and the anti-terminator stem. Our predicted attenuator structure also includes an additional anti-antiterminator stem in front of the terminator stem. This anti-antiterminator is formed by part of the sequence encoding the leader peptide and half of the anti-terminator stem. The formation of the anti-antiterminator may halt the RNA polymerase, which pauses the transcription and allows translation of the leader peptide to start [8]. During the translation of the leader peptide, the anti-antiterminator stem is opened by the translation machinery and the paused RNA polymerase is able to resume transcription. When tRNALeu is adequate, the leader peptide can be successfully translated, releasing the ribosome at the stop codon of the leader peptide, and the reformation of the anti-antiterminator stem keeps the terminator structure intact which constitutively shuts down the transcription of the downstream genes. When the concentration of tRNALeu is low in the cell, the ribosome is stalled at the region enriched with leucine codons and the anti-antiterminator stem stays opened, which enables the formation of the anti-terminator and prevents the formation of the terminator stem, allowing transcription of the downstream genes. In the structural alignment of the predicted LeuA terminator motifs in the five Shewanella species (Figure 5B), we observed complementary mutations, which provide extra confidence to support the proposed anti-antiterminator structure.
In this paper, we present a new program, RSSVM, based on support vector machines for identifying putative cis-regulatory RNA motifs using the common secondary structures and structural alignments generated by RNA Sampler. By sequentially predicting common RNA secondary structures and alignments from orthologous UTRs and identifying putative RNA regulatory motifs based on the predicted structures and alignments, the combination of RNA Sampler and RSSVM provides a new, fast and efficient pipeline for large-scale searching of RNA regulatory motifs conserved in multiple related species. We applied this strategy to five Shewanella genomes and identified putative conserved cis-regulatory RNA motifs on the genome scale. From 1002 orthologous 5′-UTR sets, we successfully identified 166 5′-UTRs that contain putative regulatory RNA motifs, including 17 of 19 known RNA motifs from Rfam, additional 21 motifs with supporting literature evidence, 72 motifs that overlap with predicted transcription attenuators/terminators, and other novel predicted regulatory RNA motifs. The fact that a large fraction of our predictions are supported by published reports or overlap with predictions by RNAz, QRNA and transcription attenuator/terminator predictors suggests that many of our new predictions are likely to be real, although experimental validation will be needed.
Comparing to other RNA motif identification tools, such as RNAz, Dynalign+LIBSVM and QRNA, RSSVM is more sensitive in detecting functional RNAs at the same FPR, especially on sequences of low identities. The more sensitive performance of RSSVM, compared to that of RNAz and Dynalign+LIBSVM, may be attributed to the following three improvements in the SVM model: first, the common structures and alignments are generated by RNA Sampler, which provides more accurate structure predictions, does not require sequence alignments as input and works well on sequences of low identities; second, more distinctive features are used to represent the common RNA structures and alignments; third, the SVM model is trained with more universal functional RNA structures that cover a large number of RNA motif/gene families and a wide range of sequence identities. We tested a few alternative SVM models which have only one or two of these improvements, such as a modified RNAz that is re-trained using the same training sets for RSSVM, and a modified RNAz that is re-trained using the same training sets for RSSVM and that uses RNA Sampler's structural alignments instead of ClustalW alignments as input. We observed that the sensitivities of these SVM models on all test sets and on test sets with low identities were higher than those of RNAz and similar to those of RSSVM when loose FPRs were allowed, and their sensitivities were gradually improved in the stringent FPR range (FPR<0.02) by adding one improvement at a time (Figure S2). For a tool designed for genome-wide motif prediction, it is essential to keep the FPR as low as possible while achieving a high sensitivity, as with a large number of data sets, low FPRs are always preferred to avoid bringing too many false positives in the predictions. RSSVM, which combines all the three improvements, gives much better sensitivities than other SVM models in the low FPR range (FPR<0.02), suggesting that RSSVM is well qualified for using in large scale predictions.
RNA Sampler and RSSVM run reasonably fast for genome-wide scan of regulatory RNAs. On average, it takes RNA Sampler 125 seconds on a single CPU workstation to predict the common structure of a set of 5 RNA sequences of an average of 150 nt in length. For a project with the similar size of the Shewanella study (1000 orthologous UTR sets, 5 species, 3 scanning windows of 150 nt, one shuffled set for each UTR set), it only takes RSSVM about 200 hours on a single CPU machine to finish the genome-wide scan of RNA motifs. The entire process can be easily run in parallel on multiple-CPU Linux clusters, allowing the whole genome prediction to be done in hours or less. We recommend that users run RNAz as well, since RNAz has better performance on sequences of high identity, which is complementary to the optimum performance of RSSVM on low-identity sequences. Moreover, consensus predictions by both approaches may provide extra confidence in the prediction quality. RSSVM is more advantageous than Dynalign+LIBSVM and QRNA in that RSSVM can take input of multiple sequences, which would provide more information on sequence conservation and complementary mutations than two sequences. Recently, Yao et al. [32] built a pipeline based on CMfinder and successfully identified several new RNA motifs. The major advantages of the CMfinder pipeline lie in its relaxation on the requirement for sequence conservation and integration of motif inference in the genome-wide search. It builds in a scanning procedure which can conveniently look for new instances of motifs in other genomes. However, one major issue of the CMfinder pipeline is that it uses a heuristic composite scoring function to sort all its predictions without giving a clear significance cutoff for confident calls for positive predictions. Also, the pipeline of CMfinder runs considerably slower than RSSVM. Using CMfinder for refinement and new instance finding on interesting predictions from a search by RSSVM may be more efficient and rewarding for users.
The RNA classification of RSSVM is based on the common structures generated by RNA Sampler. These predicted common structures provide preliminary hints for the putative structures associated to the regulatory functions. As demonstrated in the RNA motifs for LeuA, we can infer function and mechanism of the RNA motif from its structure. Although we cannot guarantee that the predicted structures are correct and perfectly match the real structures, they often indicate strong structural conservation information in the potential regulatory regions, which leads to sensitive detection of RNA motifs. Users can use Mfold [49] or other programs to refold the identified regions to obtain sub-optimal structures, which may provide good candidates for alternative structures related to the function of the RNA motifs.
There is always a trade-off between sensitivity and specificity (1 – false positive rate) in computational predictions. Using looser cutoffs (lower P-values in our case) will help increase prediction sensitivities, but at the same time more false positives may appear in the predictions. As seen in the ROC curves generated on the test sets (Figure 1), RSSVM keep high sensitivities even at very stringent FPRs, which makes it a good tool to be used in the genome-wide scan of RNA motifs. In the Shewanella RNA motif study, although we used a very stringent P-value cutoff, which corresponded to FPR<0.01 on the test sets, we still discovered most of the known RNA motifs. However, we noticed that some of the known RNA motifs were scored slightly below this cutoff. Users may consider lowering the cutoff in their own studies depending upon the tolerance to false positives.
Application of RSSVM to find RNA regulatory motifs/genes is not limited to the Shewanella genomes. This approach is fully transferable to other bacterial genomes, or in fact to any set of orthologous RNA segments that are suspected of containing conserved secondary structure motifs. We conducted some pilot tests on the classification performance of RSSVM on RNA sequences from eukaryotic genomes. Without retraining it, we ran RSSVM on 4087 sets of real RNA sequences (positive eukaryotic test sets) from 372 eukaryotic RNA motif families from Rfam and the same number of shuffled sequence sets (negative eukaryotic test sets). Each sequence set contains 3–6 sequences whose average sequence identities ranges from 20% to 100%. The ROC curves on these test sequences are shown in Figure S3. Encouragingly, RSSVM performs well in this test: at the FPR of 0.02, RSSVM gives a good prediction sensitivity of 0.5, lower than it does on the prokaryotic test sets (0.72 at FPR of 0.02) as expected (Dataset S2). Consistent with the prokaryotic tests, RSSVM becomes more sensitive than RNAz when FPR is greater than 0.02. Again, RSSVM performs much better on sequence sets whose identities are lower than 70%, with the overall prediction sensitivity jumping to 0.6 at the FPR of 0.02. On these low-identity sequences, RSSVM starts to outperform RNAz from a very low FPR of 0.005. These results suggest that the RSSVM model trained on prokaryotic RNAs can also be used to search for RNA motifs in eukaryotic sequences. It also verifies that RSSVM can find novel RNA motifs that are distinctive from those in the training sets. By re-training RSSVM on sequence sets from eukaryotic RNA families, its performance may be further improved.
To better serve the Shewanella research community and research groups who are interested in RNA regulatory motifs or post-transcriptionally regulation, we made the RSSVM software, predictions and comprehensive analysis results available online at http://ural.wustl.edu/resources.html#RSSVM.
We use the program, RNA Sampler [27], to predict common RNA secondary structures and generate structural alignments in homologous sequences. RNA Sampler is a probabilistic sampling algorithm that was recently developed by our group. In previous tests [27]–[29], RNA Sampler outperformed other leading algorithms for similar purposes on sequences of a wide range of identities. In this study, the default parameters of RNA Sampler, S = 75 (structure sample size) and i = 15 (iterations), were used in all predictions. Although RNA Sampler is able to predict RNA secondary structures with pseudoknots, we opted to not allow pseudoknots in this study.
Support Vector Machines (SVM) are supervised learning methods widely used for classification and regression. In these methods, labeled data are represented by vectors that are defined by various features, and support vector machines map the feature vectors to a higher dimensional space and construct a maximal separating hyperplane to classify the input data into binary categories. SVM has been used in previous studies [34],[35] to distinguish regulatory RNA secondary structures from random RNA structures. In such methods, the RNA secondary structures or structural alignments of homologous RNAs are represented by a set of predefined features, and the SVM maps the vectors defined by these features to a high-dimensional space. By training on the RNA structures or structural alignments of known functional and random RNAs, SVM is able to maximally separate these two groups of RNAs. Then for any unknown RNA secondary structure, SVM can classify it as either functional or random.
We developed a new SVM classifier for detecting regulatory RNAs. Our SVM classifier differs from the previous ones in three major aspects: first, the recently developed new program, RNA Sampler, is used to predict common RNA secondary structures and structural alignments on any set of homologous RNA sequence, and feature vectors based on such predictions are used to build the SVM classifier; second, a different set of feature parameters are used to represent the common RNA structures and structural alignments; third, the SVM classifier is trained on a larger number of various bacterial RNA gene and motif families that cover a wider range of sequence lengths and identities than previous studies [34],[35].
With the RNA secondary structure prediction algorithm, RNA Sampler, and the RNA motif identification algorithm, RSSVM, we can search putative regulatory RNA structural motifs from any orthologous RNA sequence set. As shown in the flow chart in Figure S5, we first retrieve orthologous RNA sequences from multiple related species, and then use RNA Sampler to predict common RNA secondary structures and structural alignments of these orthologous RNA sequences. Next, RSSVM takes in the output from RNA Sampler and identifies those containing putative RNA motifs/genes. Finally, we evaluate the prediction results by comparing to known RNA motifs or searching for supporting evidence. We applied the combination of RNA Sampler and RSSVM to multiple Shewanella genomes for genome-wide regulatory RNA discovery.
|
10.1371/journal.pntd.0006102 | Interpreting ambiguous ‘trace’ results in Schistosoma mansoni CCA Tests: Estimating sensitivity and specificity of ambiguous results with no gold standard | The development of new diagnostics is an important tool in the fight against disease. Latent Class Analysis (LCA) is used to estimate the sensitivity and specificity of tests in the absence of a gold standard. The main field diagnostic for Schistosoma mansoni infection, Kato-Katz (KK), is not very sensitive at low infection intensities. A point-of-care circulating cathodic antigen (CCA) test has been shown to be more sensitive than KK. However, CCA can return an ambiguous ‘trace’ result between ‘positive’ and ‘negative’, and much debate has focused on interpretation of traces results.
We show how LCA can be extended to include ambiguous trace results and analyse S. mansoni studies from both Côte d’Ivoire (CdI) and Uganda. We compare the diagnostic performance of KK and CCA and the observed results by each test to the estimated infection prevalence in the population.
Prevalence by KK was higher in CdI (13.4%) than in Uganda (6.1%), but prevalence by CCA was similar between countries, both when trace was assumed to be negative (CCAtn: 11.7% in CdI and 9.7% in Uganda) and positive (CCAtp: 20.1% in CdI and 22.5% in Uganda). The estimated sensitivity of CCA was more consistent between countries than the estimated sensitivity of KK, and estimated infection prevalence did not significantly differ between CdI (20.5%) and Uganda (19.1%). The prevalence by CCA with trace as positive did not differ significantly from estimates of infection prevalence in either country, whereas both KK and CCA with trace as negative significantly underestimated infection prevalence in both countries.
Incorporation of ambiguous results into an LCA enables the effect of different treatment thresholds to be directly assessed and is applicable in many fields. Our results showed that CCA with trace as positive most accurately estimated infection prevalence.
| Schistosomiasis is a debilitating disease affecting over 200 million people worldwide, mainly in developing countries. Treatment for schistosomiasis is straightforward and involves treating all school-age children, and sometimes also adults, in areas where schistosomiasis is known to occur. However, detecting intestinal schistosomiasis (Schistosoma mansoni) using Kato-Katz, the most common technique, can fail to detect infection when it is present as Kato-Katz relies on finding eggs within a small stool sample. A new field diagnostic, CCA, is promising as it is simple to use and seems to detect more infections that Kato-Katz. However, assessing the performance of CCA is difficult as we cannot be certain whether or not those positive by CCA but negative by Kato-Katz are truly infected. Additionally, CCA can often return a trace result between negative and positive which is difficult to interpret. We assess the performance of Kato-Katz and CCA in both Côte d’Ivoire and Uganda. We showed that CCA did indeed detect more infections than Kato-Katz, and that CCA accurately estimated the proportion of people truly infected in the population, when trace readings were considered positive. CCA is consequently an important and valuable tool in the fight against schistosomiasis.
| It is estimated that 237 million individuals require treatment for schistosomiasis [1]. Endemic in 56 countries spanning over Africa, the Middle East, South America, and the West Indies, Schistosoma mansoni is the most geographically widespread schistosome species. Nevertheless, despite a growing body of studies looking at distribution (e.g. [2–4]), we do not have a true representation of the number of people infected with S. mansoni as there is no definitive ‘gold standard’ field diagnostic test. Accurate prevalence estimates of those infected are important, as even low infection intensities have associated morbidity [5]. The current recommended control method for schistosomiasis is preventive chemotherapy (PC) of at-risk populations, where all school-aged children (SAC) and, where appropriate, at-risk adults in the community are treated. Frequency of treatment and who receives the drugs are dependent on the prevalence of schistosomiasis in the local area [6], as determined by the parasitological diagnostic test Kato-Katz, where eggs are detected in faecal samples examined microscopically [7, 8]. Kato-Katz has low sensitivity in those with low infection intensities and in areas of low prevalence, as egg output varies both within and between days [9] and infection can easily be missed [10]. However, Kato-Katz is highly specific, as an S. mansoni egg is easily identifiable to a trained technician [11].
A rapid diagnostic test, Circulating Cathodic Antigen (CCA), has recently been endorsed by the World Health Organisation (WHO) for use in mapping and programme impact evaluation [12]. CCA uses a urine sample to test for S. mansoni infection and consequently is much simpler to use than Kato-Katz. However, the presence of an ambiguous result between negative and positive, known as a ‘trace’ result, complicates the interpretation of a CCA test. There is no consensus in the literature on whether trace should be considered as positive or negative (e.g. trace as negative as found to be closest to Kato-Katz:[13], trace assumed to be negative: [14], trace assumed to be positive: [11], [15]). Additionally, it is not clear how to interpret prevalence estimates from CCA tests and whether or not they are reflective of infection prevalence in the population.
Latent Class Analysis (LCA) estimates the sensitivity and specificity of diagnostic tests and the prevalence in the population in the absence of a gold standard, and has been applied in a wide range of fields including human soil transmitted helminthiases [16], malaria [17], and veterinary biology [18]. Analysis of LCA can be within a frequentist or Bayesian framework, with the Bayesian approach having several advantages. Firstly, the distribution of additional parameters such as Positive and Negative Predictive Values (PPV and NPV respectively) are easily calculated from the posterior distributions obtained from the LCA, which enables the results to be more easily interpreted. Secondly, Bayesian analysis enables straightforward comparison of estimated infection prevalence and the prevalence by each test to assess how well the test performs in estimating prevalence rather than infection status of each individual. Finally, the use of posterior distributions enables straightforward testing of differences between countries and between tests.
The aim of this study is to robustly analyse CCA data from two countries, Côte d’Ivoire and Uganda, to determine the effects of considering CCA trace as negative or positive. We particularly focus on assessing the performance of CCA and Kato-Katz at measuring ‘infection prevalence’, which is estimated from the LCA, and is the main use of S. mansoni diagnostics in control programs.
Ethical approval for both surveys, including the consent process, was obtained from Imperial College Research Ethics Committee (ICREC_8_2_2) as well as from the appropriate country: Comité National d’Ethique de la Recherche (CNER; ref: 086/MSHP/CNER-kp) in Côte d’Ivoire and Uganda National Council for Science and Technology (UNCST; ref: HS1993) in Uganda.
The surveys were undertaken as part of the national schistosomiasis control programmes in each country, overseen and approved by the relevant Ministries of Health. As participants were under 18 years of age, written consent was required by a parent or informed guardian. Head-teachers in each school acted as the informed guardian as literacy levels in many areas are low. The head-teacher was informed fully about the study and requested to provide informed consent for field teams to collect urine and stool samples from pupils. Only children who consented orally both before and after selection in the presence of a witness (head-teacher) took part in the survey. Additionally, all children gave urine and stool samples freely following selection, and there were no consequences if a child did not return their samples. All data were entered and analysed anonymously.
A total of 3,035 and 693 children were included in the analysis for Côte d’Ivoire and Uganda, respectively (Table 1). Prevalence by Kato-Katz was higher in Côte d’Ivoire (13.4%) than in Uganda (6.1%), with mean intensity of infection, among all children, being over seven-fold higher in Côte d’Ivoire (26.8 eggs per gram (epg)) than in Uganda (3.4 epg; Table 1). However, prevalence by CCA was similar in both countries, both when trace was assumed to be negative (CCAtn prevalence 11.7% in Côte d’Ivoire and 9.7% in Uganda) and when trace was assumed to be positive (CCAtp prevalence 20.1% in Côte d’Ivoire and 22.5% in Uganda; Table 1). A completed STARD checklist and participant flow diagram are available in S3 and S4 Supporting Informations respectively and the raw data used for analyses are available in S5 (Côte d’Ivoire) and S6 (Uganda) Supporting Informations.
In both countries, just over 75% of pupils were negative on both tests, and 6.5% of children in Côte d’Ivoire and 3.8% of children in Uganda were positive on both tests (Table 1, Fig 1). CCA sometimes failed to detect infection where eggs were found by Kato-Katz: 4.7% of pupils in Côte d’Ivoire and 1.4% of pupils in Uganda had a negative CCA result but were positive by Kato-Katz. Similarly, Kato-Katz sometimes failed to detect infections that were positive (not trace) by CCA: 5.2% and 5.9% of pupils in Côte d’Ivoire and Uganda, respectively, tested positive for CCA but negative by Kato-Katz. Tables of CCA results split by Kato-Katz infection category are available in S7 Supporting Information.
Estimates of parameters obtained from the Bayesian LCA (sensitivity and specificity of each test and prevalence) with the associated BCI are available in Table 2, and Fig 2 shows the posterior distributions of sensitivity and specificity of each test. Table 2 also shows parameters calculated from the posterior distributions of the estimated parameters.
Sensitivity of Kato-Katz in Côte d’Ivoire (59.9%) was significantly higher than in Uganda (32.3%; Table 1, Fig 2), although, there was no evidence of sensitivity estimates of CCAtn and CCAtp differing between the countries (CCAtn = 49.3% and 50.1%; CCAtp = 63.0% and 69.7% in Côte d’Ivoire and Uganda respectively). In both countries, the sensitivity of CCAtp was significantly higher than the sensitivity of CCAtn. However, the countries differed with respect to the patterns of differences between CCA and Kato-Katz. In Côte d’Ivoire, the sensitivity of CCAtn was significantly lower than the sensitivity of Kato-Katz, but the sensitivity of CCAtp was not significantly different from the sensitivity of Kato-Katz. In contrast, in Uganda, the sensitivities of CCAtn and CCAtp were both higher than the sensitivity of Kato-Katz.
The estimated specificity of Kato-Katz did not differ significantly across the two countries (99.0% in Côte d’Ivoire vs. 99.3% in Uganda; Table 1, Fig 2). There was no evidence that the specificity of CCAtn differed between the countries (Côte d’Ivoire = 98.0%, Uganda = 98.8%), but the estimated specificity of CCAtp in Côte d’Ivoire (91.0%) was marginally, non-significantly, higher than in Uganda (87.6%; Table 2, Fig 2). In both countries, the estimated specificity of CCAtp was significantly less than both the estimated specificity of CCAtn and the estimated specificity of Kato-Katz. In Côte d’Ivoire, the estimated specificity of CCAtn was marginally, non-significantly, less than the estimated specificity of Kato-Katz and there was no evidence in Uganda that the estimated specificity of CCAtn and Kato-Katz differed.
Infection prevalence was estimated to be 20.5% in Côte d’Ivoire (Table 1; Fig 3) and 19.4% in Uganda. Indeed, there was no evidence that the estimates of infection prevalence differed between the countries (Table 2; Fig 3).
Both Kato-Katz and CCAtn substantially underestimated infection prevalence (Table 2; Fig 3). In contrast, there was no significant difference between test and infection prevalence by CCAtp in either country, with the CCAtp prevalence falling within 0.4 percentage points of the infection prevalence estimate in Côte d’Ivoire, and within 3.1 percentage points of the infection prevalence estimate in Uganda.
The aim of this study was to understand the implications considering CCA trace as negative or positive using data from two countries, Côte d’Ivoire and Uganda. We particularly focused on assessing the performance of CCA and Kato-Katz at measuring infection prevalence, as this is the main purpose of S. mansoni diagnostics in the control programs. We found that the sensitivities and specificities of CCA were much more consistent between countries than Kato-Katz and that estimates of prevalence by CCA with trace as positive was not significantly different from infection prevalence in either country. We discuss below possible reasons and implications for these results.
We found that trace values treated as positive (CCAtp) had significantly and substantially higher sensitivity and lower specificity than when treated as negative (CCAtn) in both countries. However, the sensitivity estimates indicated that CCAtp still did not detect a substantial proportion of infections (at least 30% in both countries), and this was supported by the raw data where 35% and 24% of the children that were Kato-Katz positive in Côte d’Ivoire and Uganda, respectively were negative by CCA. Although some of this may be due to misidentifying of samples, it seems extraordinary that this could explain the entire pattern.
We used LCA to analyse the data, as there is no ‘gold standard’ test for S. mansoni infection. We put a strong prior distribution (95% certain greater than 80%, with mode at 95%) on the specificity of Kato-Katz which may partly explain why estimated specificity of Kato-Katz (99%) was high and did not differ between countries. However, the lower Bayesian 95% confidence limit of Kato-Katz specificity was greater than the mode of the prior distribution suggesting that the data were not indicating lower specificity of the Kato-Katz than the prior. Consequently, the use of Bayesian analysis enabled us to assess the appropriateness of our assumptions, which would not be possible in a frequentist framework.
We elected to analyse trace positive and trace negative results within a single model. This method of analysis can be applied to any diagnostic where the result is not binary, where the results are in some way graded, and where alternative cut-off points can be assumed positive. The analysis is simple to implement and extend, with the key being that increasing the number of people testing positive on a sliding scale increases the sensitivity but decreases the specificity of the test. The preferred sensitivity and specificity balance of a diagnostic will depend on a number of different factors such as the properties of other diagnostics in use, the expected prevalence in the test population, and cost considerations. An alternative way to approach the analysis would be to analyse CCA trace negative and trace positive results separately. However, this risks the model returning logically impossible sensitivity estimates lower for trace positive than trace negative, or specificity estimates higher for trace positive than trace negative. Fitting trace positive and negative within a single model prevents this and also avoids having to interpret multiple estimates for sensitivities and specificities of other tests, and also for infection prevalence.
We analysed the data in a Bayesian framework, where the use of prior distributions can help overcome issues around degrees of freedom while still letting the data indicate if the assumptions made are not valid, as opposed to the absolute assumptions that can be required in a frequentist framework. Bayesian analysis also outputs the full distribution of each parameter through the iterations saved by the model. Consequently, distributions of additional variables calculated from the parameters are simple to obtain; we used this technique both to test for significance of differences between terms and between studies and to compare test prevalence and infection prevalence for each test.
Comparing between studies, we found that the estimated sensitivity and specificity of both CCAtp and CCAtn did not differ between the countries, in line with previous results that found the performance of CCA to be consistent before and after treatment [13]. However, the estimated sensitivity of Kato-Katz was much higher in Côte d’Ivoire, where the mean infection intensity was also much higher, than in Uganda. Lower sensitivity of Kato-Katz at low infection intensities is a well-known issue [10] and it is possible that this is the reason for our findings. A number of previous papers have assumed Kato-Katz to be the gold standard, including in a recent Cochrane review [27]. It is clear from these, and many other results, that it is not appropriate for Kato-Katz to be considered a gold standard, particularly at low intensities. We echo researchers in this [28] and other [29] fields in concluding that LCA is the most appropriate tool for assessing test sensitivity and specificity in the absence of a gold standard.
Our work adds to a growing body of literature using LCA to assess the performance of CCA [20, 30–35]. Our estimates of sensitivity for both CCA and Kato-Katz were on the lower end of those observed in these other studies, and our specificity estimates for CCA were relatively high. These studies together seem to be reflecting the general pattern of Kato-Katz sensitivity being strongly associated with infection intensities, with CCA perhaps being more consistent between environments [36]. However, this study is the first, to our knowledge, to assess the performance of CCA with respect to its main use in control programs of estimating prevalence within the study population.
Comparison of test and estimated infection prevalence suggested that the prevalence measured by CCAtp was not significantly different from infection prevalence in either country and that both Kato-Katz and CCAtn significantly underestimated infection prevalence in both countries. Consequently, our results imply that CCAtp is revealing substantial numbers of children infected with S. mansoni that were not detectable by Kato-Katz [37]. Although it could be argued that the infected children that are being missed by Kato-Katz are those with the lowest infection intensities, even low levels of schistosomiasis have an associated morbidity burden [38]. Interestingly, estimated infection prevalence did not differ between Côte d’Ivoire and Uganda, despite Kato-Katz prevalence being more than twice as high in Côte d’Ivoire than in Uganda. The difference in Kato-Katz prevalence was presumably due to repeated rounds of treatment in Uganda leading to lower average infection intensities than in Côte d’Ivoire, which are better detected by CCA than Kato-Katz. However, in the absence of historical Kato-Katz and CCA data from the schools in Uganda, it is not possible to assess whether repeated rounds of treatment in Uganda has also been associated with a concurrent decrease in estimated prevalence.
The main weakness of this study is the reliance on only two tests. Additional tests would be expected to increase the robustness of the study through increasing degrees of freedom, and there is clearly a need for additional studies incorporating more tests. However, we tried to mitigate for this weakness by using LCA, and also by incorporating covariances between the tests, which is expected to lead to more robust estimates than simply assuming the properties of different tests to be independent [39]. Additionally, the sample sizes in Côte d’Ivoire were over 4-fold higher than in Uganda, which is likely reflected in the larger confidence intervals around estimates from Uganda. As CCA becomes a more commonly used field tool, we expect sample sizes available for analyses to also increase.
We demonstrated here how ambiguous trace results can be incorporated into LCA using S. mansoni data from Côte d’Ivoire and Uganda, enabling direct comparison of test properties when trace was considered both as negative and positive, and avoiding having to make assumptions as to the nature of trace results. Our results suggested CCA with trace as positive was most reflective of infection prevalence and that both Kato-Katz and CCA with trace as negative substantially underestimated infection prevalence. Consequently, we conclude that CCA is an appropriate tool for field testing for S. mansoni in control programmes.
|
10.1371/journal.pntd.0007004 | Does appreciative inquiry decrease false positive diagnosis during leprosy case detection campaigns in Bihar, India? An operational research study | India contributes ~60% to the global leprosy burden. The country implements 14-day community-based leprosy case detection campaigns (LCDC) periodically in all high endemic states. Paramedical staff screen the population and medical officers of primary health centres (PHCs) diagnose and treat leprosy cases. Several new cases were detected during the two LCDCs held in September-2016 and February-2018. Following these LCDCs, a validation exercise was conducted in 8 Primary health centres (PHCs) of 4 districts in Bihar State by an independent expert group, to assess the correctness of case diagnosis. Just before the February 2018 LCDC campaign, we conducted an “appreciative inquiry” (AI) involving the health care staff of these 8 PHCs using the 4-D framework (Discovery-Dream-Design-Destiny).
To assess whether the incorrect case diagnosis (false positive diagnosis) reduced as a result of AI in the 8 PHCs between the two LCDC conducted in September-2016 and February-2018.
A three-phase quantitative-qualitative-quantitative mixed methods research (embedded design) with the two validation exercises conducted following September-2016 and February-2018 LCDCs as quantitative phases and AI as qualitative phase. In September-2016 LCDC, 303 new leprosy cases were detected, of which 196 cases were validated and 58 (29.6%) were false positive diagnosis. In February-2018 LCDC, 118 new leprosy cases were detected of which 96 cases were validated and 22 cases (23.4%) were false positive diagnosis. After adjusting for the age, gender, type of cases and individual PHCs fixed effects, the proportion of false positive diagnosis reduced by -9% [95% confidence intervals (95%CI): -20.2% to 1.7%, p = 0.068]
False positive diagnosis is a major issue during LCDCs. Though the decline in false positive diagnosis is not statistically significant, the findings are encouraging and indicates that appreciative inquiry can be used to address this deficiency in programme implementation.
| India is the highest leprosy burden country in the world. Government of India’s National Leprosy Eradication Programme (NLEP) launched Leprosy Case Detection Campaign (LCDC)—an active community-based case detection campaign—in 2016 in all high burden areas to detect undiagnosed cases. Following these LCDC, a small validation exercise was conducted in 8 Primary health centres (PHCs) in Bihar State by an independent expert group, to assess the correctness of case diagnosis. It found that ~30% of the cases detected were not true cases, but false positive diagnosis. To reduce false positive diagnosis in the subsequent round of LCDC in 2018, an “appreciative inquiry” involving the health care staff of these 8 PHCs using the 4-D framework (Discovery-Dream-Design-Destiny) was done. In 2018 LCDC, the false positive diagnosis decreased to ~23%. After adjusting for the differences in the patient and health facility characteristics, the decline in false positive diagnosis was estimated to be about 9%. This study shows that false positive diagnosis was a major issue during LCDCs and that appreciative inquiry can be used to address this deficiency in programme implementation.
| Leprosy is a chronic infectious disease caused by the bacteria—Mycobacterium Leprae. It usually affects the peripheral sensory nerves and has a wide range of clinical manifestations. The disease is characterized by long incubation period generally 5–7 years. Leprosy is completely curable with 6–12 months of multidrug therapy. Early diagnosis and treatment of cases is the most effective way of halting transmission and eliminating leprosy from the community [1].
India is the highest leprosy burden country in the world. In 2016, ~135,000 new cases of leprosy were detected by the Government of India’s National Leprosy Eradication Programme (NLEP). This constituted about 66% of total leprosy cases detected in the world in that year [2,3]. In India, in 2016, the Annual New Case Detection Rate (ANCDR) was 9.71 cases per 100,000 population and Prevalence Rate (PR) was 0.66 per 10,000 population. ANCDR and PR have been showing stable trends since 2006. The other leprosy indicator related to the child cases (number and proportion of new cases aged <15 years) is also relatively high indicating on-going transmission in the community [4,5]. The major source of transmission of infection in the community are the hidden undiagnosed and untreated cases. Hence, in order to detect these cases, NLEP introduced yearly Leprosy Case Detection Campaign (LCDC)—community based active case finding campaign—in 2016 in high endemic states [6].
Bihar, a state in the eastern part of India (population of 113 million), is one of the highest leprosy burden states in the country. It is reporting 16,000 to 20,000 new cases of leprosy every year since 2005 (15–20% of the cases in the country). In 2016, Bihar reported 16,185 new cases of leprosy. LCDC was carried out in Bihar in 2016 in 20 out of 38 districts and this yielded 4517 new cases of leprosy [2]. The campaign was organised for 14 days from 5–18 September 2016.
Damien Foundation India Trust (DFIT), is a charitable Non-Governmental Organization working for leprosy and tuberculosis control in Bihar. DFIT provides technical support to NLEP in planning, implementing, monitoring, and evaluation. DFIT organised a validation exercise in collaboration with the State Leprosy Programme Officer, Bihar. The validation exercise was carried out by an independent expert group to assess the quality of diagnosis among the cases detected during the campaign. Two blocks in each of the four districts- Nalanda, Sitamarhi, Gopalganj and Araria (which reported highest number of cases during LCDC) were selected for validation. It was found that about 30% of cases detected during LCDC were wrongly diagnosed as leprosy cases (false positive cases). False positive diagnosis leads to unnecessary medication, causes stigma, isolation, loss of employment and discrimination that can lead to considerable mental trauma and agony in the patients and their families [7,8]. In addition, it also discredits the LCDC campaign. Thus, there was an urgent need to understand the reasons for false positive diagnosis and undertake suitable corrective measures to address this issue.
Diagnosis of leprosy requires specific clinical expertise. Anecdotal discussions with the programme staff indicated that with a general decline in leprosy cases over the last few decades, there has been a decline in the clinical expertise within the public health system to diagnose leprosy due to retirement of trained leprosy personnel without new recruitments, inadequate trainings, transfer of existing leprosy trained workforce to other public health programmes etc. In Indian public health programme settings, the traditional approach for problem-solving is generally characterised by fault finding and penalization. In contrast, we wanted to test a flexible and friendly approach for reducing false positive diagnosis.
Appreciative Inquiry (AI)—is a philosophical approach to organizational learning, change management and research. It is a process which shifts the focus of programme or organization from problem identification, defensiveness and denial of facts towards discovery of programme strengths and building on what works well in the given setting and context [9]. This approach has been found effective in improving obstetric referral system in Cambodia [10], improvement of community-based mental health services [11], improvement in nursing care in hospital setting in the United Kingdom [12], and development of better health care work environment in NHS [13]. AI offers a framework which positively influences organizational growth by generating common goals and actions to be achieved by the programme staff [11]. It is emerging as a promising approach for staff motivation and programme sustainability in public health programmes in low and middle-income countries.
Therefore, in 2017–2018, we conducted an operational research study to assess whether AI with health staff reduces the number (and proportion) of false positive diagnosis of leprosy cases during the LCDC in February 2018 when compared to LCDC in September 2016.
This is a three-phase mixed methods study (embedded design). The quantitative part contained a before-after study design and the qualitative intervention comprised of appreciative inquiry (Fig 1).
The study population included all leprosy cases detected during LCDCs in September 2016 and in February 2018 in 8 blocks of 4 districts and validated by the DFIT team. For the qualitative part, the following staff were invited to the AI meeting: at least one Medical Officer from each PHC, Block Community Mobiliser, Block Health Manager, District Nucleus Team, Communicable Disease Officer.
The validation was undertaken within four weeks of LCDC. In 2016, four teams were formed for the exercise, each consisting of a Medical Officer, a supervisor with more than 10 years of experience in leprosy diagnosis from the State level and another supervisor from the district nucleus team. This team attempted to validate all the new leprosy patients diagnosed during LCDC and assessed whether these cases were true positive cases or false positive cases using the same clinical diagnostic criteria given in Table 1. In this process, they also collected socio-demographic and clinical data from the patients and noted their findings using a structured data collection case sheet. Cases were examined either at the primary health centres or at the patients’ residences.
For the quantitative part, the individual patient wise data of all cases diagnosed as leprosy during LCDC conducted in 2016 & 2018 in these 8 blocks were available with the State NLEP Office in Patna. The patient wise data collected during validation exercise in 2016 & 2018 was available at the DFIT office in Patna. The principal investigator (ANW) obtained these data for its usage in this study. The patient wise data contained information on the name of the patient, age, sex, type of case (PB or MB), PHC, district, block and disability grading in accordance with the NLEP guidelines.
We followed the Appreciative Inquiry framework to plan the intervention One appreciative inquiry meeting was held in each of the four districts in the month of November-December 2017. Formal permissions from the district health authorities were obtained for this meeting. It was facilitated as a group activity. A total of 43 personnel belonging to to various health cadre as mentioned above participated in these meetings (>90% participation). The participants were informed about the purpose of the meeting and were oriented to the philosophy of AI at the time of the meeting. Each meeting had four sequential phases—Discovery, Dream, Design and Destiny (4D)—as per the AI framework (Box 1). Discovery: After creating a climate of open exchange, this step was implemented to explore the strengths and positive experiences on what is working well in the programme from each of the participant. Dream: This phase of the meeting was facilitated on the broad themes emerging in the ‘discovery’ phase to challenge the status-quo and dream for the better programme achievements. The participants were asked to share their suggestions to improve the programme activities further. Design: In this phase, participants were asked to design the action plan for improvement or change in the desired direction based on the collective dream. Destiny: In this phase pre-conditions crucial for change or improvement to happen were discussed.
The district leprosy officer (Communicable Disease Office) was involved and briefed about the AI approach and its philosophy to seek his full co-operation in the improvement process.
In ‘Appreciative Inquiry’ approach (AI), the questions pertained to the following: Experience—based on your experience, what is the current status of the leprosy programme?; Opinion—What is your opinion on the current status of the leprosy programme?; Suggestions—What could be the ways to improve the current status of the programme?; Discover—Tell me that high point in the leprosy programme which makes you feel high; Dream—What do you wish to improve in leprosy programme in the future?[14]
Quantitative: All quantitative data analysis was done using EpiData [version 2.2.2.183, EpiData Association, Odense, Denmark] and Stata [Version 15, StataCorp, College Station, Texas, United States]. The demographic and clinical characteristics has been summarized using frequencies and percentage. We compared the demographic and clinical characteristics of patients detected during LCDC and patients reached during validation in 2016 and 2018 using Chi-square test. We used log binomial models with robust standard error estimates to obtain the adjusted differences in the proportion of false positive cases (in those validated) between 2016 and 2018 after adjusting for the differences in age, sex, type of case and the PHCs from which these cases were detected. A P-value < 0.05 was considered for statistical significance. Qualitative: For the analysis of qualitative interview we used the AI framework. The issues that emerged from the four meetings were grouped into three broad themes. The themes were similar to the Discovery, Dream, Design concept of the AI framework. The themes were ‘strengths of the program’, ‘imagined future outcome of the program’, ‘suggestions to improve the program in future’[14]. The similar issues within a theme was grouped into categories. Two investigators did the analysis independently. grouped the issues into these themes. Any discrepancies were sorted out by discussion. The final analysis was finally reviewed by another investigator.
We obtained ethics approval for this study from the Ethics Advisory Group of the International Union Against Tuberculosis and Lung Disease, Paris, France and from the ethics review board of the Sri Manakula Vinayagar Medical College and Hospital Pondicherry, India. We obtained administrative approvals for conducting this study from the State and the four District Leprosy Officers. For the quantitative component of the study, which involved the retrospective review of patient records, we got a waiver from obtaining informed consent from patients. However, we obtained written informed consent from all the participants who were part of the Appreciative inquiry meetings.
In 2016, 303 leprosy cases were detected during LCDC in the 8 PHCs of which 196 cases could be validated. Of those validated, 58 (29.6%) were false positive cases (Fig 3). The proportion of cases validated when compared to detected cases did not differ by age, gender and type of leprosy cases. However, proportion validated differed across the 8 PHCs (Table 2).
As planned, four AI meetings were held, one in each district. The themes that emerged during these meetings pertaining to discovery, dream, design is summarised in (Table 3). The major strengths of the programme were availability of manpower and infrastructure, availability of commodities for management of leprosy, administrative support from government and other external sources. The imagined future outcome of the program was leprosy free society without stigma, discrimination and a well-informed society. The proposed action plan to achieve the future outcomes included reorientation training of all the programme staff, financial and administrative support, improved intersectoral co-ordination, better referral system, strengthening supervision and monitoring, health education of the community and implementation of the social welfare schemes.
In 2018, 118 leprosy cases were detected during LCDC in the same 8 PHCs—62% decline in the number of new cases diagnosed when compared to LCDC conducted in 2016. Of the 118 cases detected, 94 cases were validated. Of those validated, 22 cases (23.4%) were false positive cases (Fig 3). The proportion of cases validated differed from the cases detected in LCDC by gender, type of cases, across districts and PHCs (Table 2).
After adjusting for the age, gender, type of cases and individual PHCs fixed effects, the prevalence ratio of false positive cases between 2016 and 2018 was 0.67 (95% CI: 0.44–1.03, p = 0.068) indicating a 33% decline in the relative prevalence of false positive cases in 2018 across 8 PHCs when compared to 2016 (Table 4). From the coefficients of the model used to derive the adjusted prevalence ratios, the adjusted estimated decline in the proportion of false positive cases between 2016 and 2018 was -9% (95% CI: -20.2% to 1.7%). The proportion of false positive cases across PHCs varied widely and it ranged from 3.6% to 46% with the false positive cases in some PHCs were almost 3–4 times higher than the others.
This is one of the first studies from India in recent years, describing the proportion of false positive diagnosis during LCDC campaigns and to assess the effect of appreciative inquiry as an intervention to reduce false positive diagnosis. The study had three important findings. First, in 8 PHCs of 4 districts in Bihar, 303 new leprosy cases were diagnosed during LCDC in September 2016 and a repeat LCDC conducted in February 2018 reduced the number of new cases diagnosed to 118 cases (~62% decline). Second, when a sample of these new cases detected during the two LCDCs was independently validated by a group of experts, the proportion of cases found to be false positive declined from 29.6% in September 2016 LCDC to 23.4% in February 2018 LCDC (6.2% decline). In-between the two rounds of LCDC an appreciative inquiry was conducted by the study investigators involving the district leprosy programme officer and the health care providers of these PHCs. Our inferences based on these study aspects/findings are as follows:
First, there was 62% decline in the total number of cases diagnosed in the 8 PHCs between the two rounds of LCDCs in 2016 and 2018. Though we do not have a control group of PHCs to compare this decline, we had aggregate data on the overall decline in the number of cases detected in the same and neighbouring districts (where AI was not implemented) from the programmatic reports. On an average, the decline in the number of cases was ~42% (range from -92% to +5%). Therefore, we feel that the decline in the number of cases seen in the 8 intervention PHCs is due to the overall decline that can be anticipated between the two rounds of LCDC and is not unique to these 8 PHCs (i.e., it is unrelated to AI).
Second, the adjusted average decline in the proportion of false positive cases between the two rounds of LCDCs was -9% (95% CI: -20% to +1.3%). We feel that this decline is programmatically relevant. However, 95% confidence intervals (CI) are wide and crosses the null value (0%) and therefore we do not have the statistical evidence at the 95% CI level to say that there is conclusive statistical proof about the reduction in the proportion of false positive diagnosis. The wide confidence intervals were due to relatively small sample size (during the February-2018 LCDC) and due to the huge variations in the proportion of false positive diagnosis at the PHC levels. Therefore this should not be termed as “absence of evidence” and result in inaction or rejection of the findings [15]. We therefore estimated the 90% confidence intervals for the adjusted decline and it was -18% to -0.1%. Based on this, we feel that though we do not have statistical evidence for the decline in false positive diagnosis at 95% CI level, we have statistical evidence for this decline at the 90% CI level. We feel that our study provides “proof of concept” that the intervention, has the potential to decrease the false positive cases [16,17].
Third, did AI as an intervention lead to these changes in these 8 PHCs? The ideal study design to provide a confirmatory answer to this question would have been a cluster randomised before and after study. Since we were in a programmatic setup and not a research setup, this ideal study design was operationally not feasible. Even if we were to select a control group of PHCs now, measuring and ensuring that the intervention and control PHCs were almost similar in all aspects except for the intervention in question, is practically impossible. Therefore, we are unable to give a confirmatory answer to this key question. However, our current study design resembles a single arm before and after study design. In 7 out of 8 PHCs the medical officers who had diagnosed the cases in 2016 and 2018 remained the same. They were given an identical refresher training on how to diagnose and treat leprosy before both the LCDC campaigns in 2016 and 2018. However, the only major difference was that, in 2018, they had information on false positive diagnosis. This information was given in a friendly manner using the principles of AI. The health staff who participated in AI meetings quoted that they liked this strategy of change management than the usual hierarchical approach. We therefore believe that AI could have played a role in reducing the false positive diagnosis and the change could have happened through the re-trainings and supportive supervision and monitoring.
Fourth, the most important message for the NLEP from this study is that false positive diagnosis is a major issue during LCDC. This has been highlighted in one of the validation studies done in India during 2004 where 9.4% (95% CI: 7.4%-11.4%) of the cases were found to be wrongly diagnosed as leprosy [18]. And therefore, sufficient measures must be undertaken to address this issue. To our knowledge there are no published studies in the literature since 2004 describing the magnitude of false positive diagnosis during LCDC. Hence, we are unable to compare and contrast our study findings with the false positive diagnosis in other settings or describe the circumstances under which false positive diagnosis is likely to be high or low. Furthermore, our study does not provide information on false negative diagnosis (i.e., the number and proportion of true cases of leprosy missed during the LCDC) which is essential to reduce transmission. These issues have to be explored in future through more operational research studies or validation exercises.
Fifth, the occurrence of false positive diagnosis and false negative diagnosis is due to the “subjectivity” in the diagnosis of leprosy cases due to its dependence on clinical criteria. There are several commentaries/studies on how using clinical criteria can lead to misdiagnosis [19–21]. There are serological tests to assess infection of leprosy that could be used for difficult cases (antibodies against Phenolic glycolipid (PGL-1) Mycobacterium leprae antigen) [22] or use split skin smears [23]. We need to explore this on a programmatic perspective to reduce misdiagnosis. Therefore, in order to reduce the errors in diagnosis, we strongly feel that NLEP must consider making the diagnostic criteria more ‘objective’, introduce more rigorous/comprehensive methods for training of medical officers and/or constitute a committee of two or more trained medical officers at the PHC level to arrive at diagnosis of leprosy. Given the human resource shortages at PHC level in Bihar, we are not sure whether this suggestion is practically feasible or not. Assessing which of these measures will help in reducing misdiagnosis of cases under routine programmatic setting is an area for future research.
Lastly, we strongly believe that the validation exercise conducted by DFIT in the limited number of PHCs helped identify an important operational problem and therefore this needs to be done in all other districts and other states of India. The protocols for validation have been developed by NLEP but the validations are not carried out as envisaged. The NLEP must focus on routine validation exercises in future.
In conclusion, about one in three cases diagnosed as leprosy during LCDC in 2016 in 8 PHCs of Bihar was found to be false positive. This reduced to one in four cases during the LCDC conducted in February 2018 due to the implementation of AI. Though the decline in proportion of false positive diagnosis is not statistically significant at 95% CI level, we believe the findings are programmatically important.
|
10.1371/journal.ppat.1002791 | Pathogenicity and Transmissibility of North American Triple Reassortant Swine Influenza A Viruses in Ferrets | North American triple reassortant swine (TRS) influenza A viruses have caused sporadic human infections since 2005, but human-to-human transmission has not been documented. These viruses have six gene segments (PB2, PB1, PA, HA, NP, and NS) closely related to those of the 2009 H1N1 pandemic viruses. Therefore, understanding of these viruses' pathogenicity and transmissibility may help to identify determinants of virulence of the 2009 H1N1 pandemic viruses and to elucidate potential human health threats posed by the TRS viruses. Here we evaluated in a ferret model the pathogenicity and transmissibility of three groups of North American TRS viruses containing swine-like and/or human-like HA and NA gene segments. The study was designed only to detect informative and significant patterns in the transmissibility and pathogenicity of these three groups of viruses. We observed that irrespective of their HA and NA lineages, the TRS viruses were moderately pathogenic in ferrets and grew efficiently in both the upper and lower respiratory tracts. All North American TRS viruses studied were transmitted between ferrets via direct contact. However, their transmissibility by respiratory droplets was related to their HA and NA lineages: TRS viruses with human-like HA and NA were transmitted most efficiently, those with swine-like HA and NA were transmitted minimally or not transmitted, and those with swine-like HA and human-like NA (N2) showed intermediate transmissibility. We conclude that the lineages of HA and NA may play a crucial role in the respiratory droplet transmissibility of these viruses. These findings have important implications for pandemic planning and warrant confirmation.
| North American triple reassortant swine (TRS) influenza A viruses have caused sporadic human infections, but human-to-human transmission has not been established. We wished to elucidate potential human health threats posed by the TRS viruses and to identify determinants of virulence in the TRS and closely related 2009 H1N1 pandemic viruses. We used a ferret model to evaluate the pathogenicity and transmissibility of North American TRS viruses with the HA and NA antigenic proteins of swine viruses and of human viruses. We observed that the North American TRS viruses grew efficiently in both the upper and lower respiratory tracts and caused moderate pathogenicity in ferrets. The viruses were readily transmissible via direct contact, irrespective of their HA and NA lineages. However, transmissibility via respiratory droplets was substantially greater when the viruses carried the HA and NA of human influenza A viruses rather than of swine influenza A viruses. Because ferrets are a useful model of human influenza infection, this finding helps to predict features that increase the risk to human health.
| For nearly 70 years, swine influenza virus in North America was relatively stable, dominated by the classical-swine H1N1 (cH1N1) subtype [1]. However, H3 seasonal human influenza A viruses were circulating at low frequency in U.S. swine [2]. In 1998, influenza epidemiology in North American swine changed dramatically with the emergence of double-reassortants (combining gene segments of cH1N1 and seasonal human H3N2 influenza A viruses) and triple-reassortants (adding gene segments from avian influenza lineages). The triple-reassortants gained predominance in North American swine and continued to evolve, further reassorting with cH1N1 and contemporary seasonal human influenza viruses [3], [4]. All of the currently circulating North American triple-reassortant swine (TRS) influenza A viruses contain a similar constellation of internal genes (avian PA and PB2, human PB1, and classical swine-lineage M, NP, and NS), but their surface glycoproteins are derived from different lineages (classical swine-lineage H1 and N1 and seasonal human-lineage H1, H3, N1 and N2).
Sporadic infections with TRS H1N1 (swine-like HA and NA) and H1N2 (swine-like HA, human-like NA) viruses have been reported in humans exposed to swine in North America [5]. Some have included severe lower respiratory tract disease and diarrhea. H3N2 (human-like HA and NA) TRS viruses have also been isolated from humans [6], [7], [8]. In 2009, TRS viruses with human-like H1 and N1 (closely related to A/Brisbane/59/2007 [H1N1]) caused cough, fever, nasal congestion, rhinorrhea, sneezing, malaise, and dizziness in humans [9]. These symptoms were very similar to those caused by the 2009 H1N1 pandemic viruses, which possessed six gene segments (PB2, PB1, PA, HA, NP, and NS) closely related to those of North American TRS viruses [10]. However, unlike the 2009 H1N1 pandemic viruses, the TRS viruses were not reported to be transmissible among humans.
Despite extensive recent studies of the pathogenicity and transmissibility of pH1N1 viruses in different animal models [11]–[14], there is very little information of this kind about North American TRS viruses. A/swine/Kansas/77778/2007 (H1N1), a triple reassortant similar to H1N1 viruses that infected humans and pigs at an Ohio county fair in 2007, was isolated from swine herds in the Midwestern United States. This virus is highly virulent in swine and is readily transmitted to sentinel pigs [15]. TRS virus A/Swine/Texas/4199-2/98 (H3N2) was also shown to be transmissible from infected swine to direct-contact swine and from them to a second group of direct-contact swine [16]. Belser and co-workers found two North American H1N1 TRS viruses (with swine-like HA and NA) isolated from humans to be pathogenic in mice [17]. In ferrets, these viruses showed pathogenicity similar to that of 2009 pandemic H1N1 influenza virus but less efficient transmissibility [18]. We have shown that the TRS virus A/swine/Arkansas/2976/02 (H1N2) and the Eurasian avian-like swine virus A/swine/Hong Kong/NS29/09 (H1N1) are not transmissible via respiratory droplets in ferrets [19]. The TRS virus A/swine/Guangdong/1222/2006 (H1N2) and the Eurasian avian-like swine virus A/swine/Fujian/204/2007 (H1N1) were recently shown not to be transmissible by direct contact in guinea pigs [20]. Very recently, Pearce et al. demonstrated that H3N2 TRS viruses isolated from humans were efficiently transmitted via respiratory droplets (RD) in ferrets [8].
Most of the North American TRS viruses belong to three subtypes: H1N1, H1N2, and H3N2; H1 and N1 are of the classical swine or seasonal human lineages, while H3 and N2 are of seasonal human lineage only. TRS viruses with human-like HA and NA have recently become the predominant influenza viruses isolated from swine in the US (St. Jude swine influenza surveillance program, unpublished data), but their transmissibility has not been tested in the ferret model. Ferrets are an established small-animal model that appears to recapitulate the pathogenicity and transmissibility of human seasonal influenza A viruses [11]–[13] and the poor human transmissibility of H5 and H7 avian influenza A viruses [21]–[23].
In the present study, we used the ferret model to evaluate the pathogenicity and transmissibility of three distinct groups of North American TRS viruses (H1N1 viruses with classical swine-like HA and NA; H1N2 viruses with classical swine-like HA but human-like NA; and H1N1, H1N2, and H3N2 viruses with human-like HA and NA) and of the Eurasian avian-like swine virus A/sw/Italy/1369-7/1994 (H1N1) (Italy/94). Because a limited number of ferrets could be used, the study was designed to detect patterns in the transmissibility and pathogenicity of TRS viruses that, once confirmed, will have important implications for pandemic preparedness. Italy/94 virus was less efficiently transmissible than the North American TRS viruses. The North American TRS viruses, regardless of their HA and NA lineages, were readily transmissible to co-housed (direct contact, DC) ferrets, while viruses with human-like HA and NA or with human-like NA alone showed enhanced transmission via respiratory droplets.
Ferrets inoculated with 106 pfu of each virus (a dose previously found to result in consistent infection [11]–[13], [18], [22]) showed only mild clinical signs of illness, mild to moderate weight loss (∼3% to 9%), and infrequent sneezing (Table 1). Among the ferrets that lost weight, weight loss was maximal during days 4 to 6 pi, and then weight began to increase. Ferrets that did not lose weight showed no significant change until day 4 pi and then began gaining weight. A few inoculated ferrets (e.g., two inoculated with Italy/94) gained weight continuously, starting on day 1. A few ferrets (mainly those inoculated with TRS viruses with human-like HA and NA) had elevation of body temperature (maximum increase, 1.5°C). We observed no significant lethargy and no ruffled fur. Infectious virus was observed in nasal washes until approximately day 6 post-inoculation (p.i.), with peak titers on day 2 p.i. TRS virus A/sw/NC/47834/2000 (H1N1, with swine-like HA and NA) caused the least nasal virus shedding (mean peak titer, 5×104 pfu/ml vs. 106 pfu/ml for the other two viruses in this group) (Fig. 1).
One of the TRS isolates (A/sw/IN/9K035/1999 [H1N2]) with swine-like HA but human-like NA caused peak nasal virus shedding (mean peak titer, 106 pfu/ml) substantially higher than that caused by the other two viruses in this group (mean peak titers, 2×105 pfu/ml and 1×105 pfu/ml). TRS isolates containing human-like HA and NA caused the highest peak nasal wash titers (mean, 106 pfu/ml). Infectious virus particles were identified at various titers in the lungs of all ferrets inoculated with the studied TRS viruses (Table 1). Ferrets inoculated with the Italy/94 virus had low peak nasal wash titers (mean, 1.7×105 pfu/ml) and exhibited almost no clinical signs. However, infectious virus particles were obtained from the lungs of both inoculated ferrets.
These findings show that overall, North American TRS viruses grow efficiently in both the upper and lower respiratory tracts of ferrets irrespective of their HA and NA lineages and cause moderate pathogenicity, similar to the reported pathogenicity of the 2009 pandemic H1N1 viruses [11]–[14].
All of the North American TRS viruses caused bronchitis, bronchiolitis, alveolitis and alveolar wall interstitial changes, with varying degrees of involvement and severity. The degree of involvement and the severity also varied to some extent in different ferrets inoculated with the same virus and in different lobes of the same lung. Fig. 2 shows representative changes. The bronchitis featured intraluminal granulocytes and/or mucus, bronchial epithelial hyperplasia with submucosal mucus gland loss, and mixed inflammatory-cell infiltrates. The bronchiolitis featured intraluminal cellular debris, sloughed epithelial cells, and inflammatory cells (macrophages and/or granulocytes) with or without bronchiolar epithelial cell necrosis and/or regenerative epithelial cell hyperplasia and hypertrophy. In the alveolitis, the alveoli surrounding the bronchioles contained a mixture of inflammatory cell infiltrates (granulocytes, lymphocytes, plasma cells and macrophages) and foci of pneumocyte hyperplasia. The interstitia (alveolar walls) were either normal or thickened by increased cellularity. Italy/94 virus caused similar morphologic changes but they were far less severe than those caused by the North American TRS viruses.
Unlike the 2009 H1N1 pandemic viruses, North American TRS viruses are not known to be transmissible among humans [5], [9]. To investigate factors that affect transmissibility, we assessed the transmission of different TRS viruses by direct contact (DC; in co-housed ferrets) and by respiratory droplets (RD). Italy/94 (Eurasian avian-like swine) virus was transmitted to only one of two DC ferrets and to neither RD ferret (Table 2, Fig. 1B), indicating poor transmission efficiency.
Among the TRS viruses with swine-like HA and NA, A/sw/NC/47438/2000 (H1N1), which had the lowest mean peak nasal wash titer (≤105 pfu/ml), was transmitted only by direct contact (2/2 DC, 0/2 RD) (Fig. 1C). Although the A/sw/NC/18161/2002 virus replicated efficiently (mean peak nasal wash titer, 106 pfu/ml) in donor ferrets, transmission was observed in only one of two DC ferrets and in neither RD ferret. Infectious A/sw/MN/6998/2003 was detected on day 7 post-exposure (p.e.) (day 8 p.i.) in the nasal wash of one of the two RD ferrets. Overall, only one of the three viruses in this group was transmitted via RD, to one of two ferrets; the remaining two viruses were not transmitted via RD. Therefore, TRS viruses with swine-like HA and NA were poorly transmitted via RD. Poor RD transmission of TRS viruses containing swine-like HA and NA has been reported previously [16].
The TRS viruses with human-like HA and NA replicated efficiently (mean peak nasal wash titer, ∼106 pfu/ml) and were efficiently transmitted (2/2 DC, 2/2 RD) in ferrets (Fig. 1E and Table 2), consistent with a recent report on the RD transmissibility of TRS viruses containing human-like HA and NA (H3N2) [8]. The transmission efficiency of TRS viruses with swine-like HA and human-like NA (N2) varied from strain to strain. Among these viruses, A/sw/MN/1182/2001 was as efficiently transmitted (2/2 DC, 2/2 RD) as TRS viruses with human-like HA and NA. In contrast, infectious A/sw/MN/5763/2003 and A/sw/IN/9K035/1999 viruses were detected in the nasal wash of only one of two RD ferrets (Fig. 1D, Table 2). A/sw/MN/5763/2003 virus was first detected in a RD ferret on day 7 p.e., while the other two viruses in this group were first detected on day 3 p.e.
Taken together, these results suggest that in the ferret model, 1) transmissibility of Italy/94 (Eurasian avian-like swine) virus is poor, 2) North American TRS viruses are readily transmissible by direct contact irrespective of their HA and NA lineages, and 3) transmissibility via RD of TRS viruses with swine-like HA and NA, swine-like HA but human-like NA, and human-like HA and NA is poor, moderate, and efficient, respectively.
Our results show that unlike seasonal H1N1 viruses, whose replication is reported to occur primarily in the upper respiratory tract [11]–[13], TRS viruses grow efficiently in ferret lungs and cause substantial lung pathology, similar to that reported for pandemic H1N1 viruses [11]–[13]. As the temperature is higher in the lower than the upper respiratory tract, ability to grow at a higher temperature might favor virus growth in the ferret lung. To better understand the pathogenicity of the swine isolates studied, we examined the multi-cycle growth kinetics of these swine viruses, of seasonal human H1N1 virus A/Brisbane/59/2007, and of the 2009 pandemic H1N1 virus A/Mexico/4482/2009 in MDCK cells at different temperatures. MDCK cell monolayers were inoculated with the viruses at a multiplicity of infection (MOI) of 0.001 and incubated at 33°C, 37°C, and 39.5°C. At different h p.i., supernatants were harvested and virus was titrated by pfu assay (Fig. 3). At 33°C, Italy/94 (Eurasian avian-like swine) virus and TRS viruses with swine-like HA and NA had similar growth kinetics, with peak titers of ∼108 pfu/ml. TRS viruses with human-like HA and NA grew to substantially higher titers (∼109 pfu/ml), while TRS viruses with swine-like HA and human-like NA grew to titers similar to those of seasonal H1N1 or 2009 pH1N1 viruses (108–109 pfu/ml, Fig. 3). At 37°C, although final yield increased only slightly, replication of all viruses was accelerated, as noted by significantly higher titers at 12 and 18 h p.i. At 39.5°C, replication of all swine viruses and of the 2009 pH1N1 virus A/Mexico/4482/2009 was less (by a factor of 10 to 100) than their replication at 37°C. However, the reduction of virus titer at 39.5°C was greatest for seasonal human H1N1 virus A/Brisbane/59/2007 (6×103 pfu/ml vs. 106–108 pfu/ml for swine and pH1N1 viruses).
At all three temperatures, TRS viruses with human-like HA and NA grew to the highest titers, whereas those with swine-like HA and NA grew to the lowest titers. These replication characteristics somewhat paralleled the viruses' overall respiratory droplet transmission efficiency.
Despite sporadic human infections with North American TRS influenza A viruses, their human-to-human transmission has not been established, and pathogenicity and transmission studies in animal models have been very limited. This study of the pathogenicity and transmissibility of North American TRS viruses containing both swine- and human-like HA and NA found that the viruses grow efficiently in both the upper and lower respiratory tracts and cause moderate pathogenicity similar to that reported for the 2009 pandemic H1N1 viruses [11]–[14]. The TRS viruses were readily transmissible by direct contact in ferrets, irrespective of their HA and NA lineages. However, RD transmissibility varied significantly with the lineages of HA and NA: TRS viruses with swine-like HA and NA, swine-like HA but human-like NA, and human-like HA and NA were transmitted poorly, moderately, and efficiently, respectively, via respiratory droplets.
Multiple viral factors, including HA receptor specificity, human-specific amino acid residues (e.g., 627K/701N) in PB2, and balance between HA and NA, are known to influence transmission and pathogenicity in humans [24]–[28]. Like the 2009 pH1N1 viruses, all TRS viruses studied here have avian-origin PB2 containing 627E and 701D. However, SR polymorphism (590S and 591R) within pH1N1 PB2 was shown to partly compensate for the absence of 627K in polymerase activity and virus replication in human A549 cells, suggesting that this polymorphism plays a role in efficient growth of pH1N1 viruses in the human upper respiratory tract [29]. E627K substitution in PB2 was later shown not to alter the growth of pH1N1 virus in MDCK cells at 33°C, 37°C, or 39°C or to significantly alter its virulence and replication in mouse and ferret lung tissues [30]–[32]. Importantly, the PB2 of Italy/94 (Eurasian avian-like swine) virus, containing 627E/701D, was associated with less lung pathology and transmissibility; it also lacks the SR polymorphism, instead containing 590G/591Q, which were shown to reduce the polymerase activity of 2009 pH1N1 virus by 50% [29]. Like pH1N1, all of our TRS viruses contain the avian-specific amino acids 627E and 701D and the SR polymorphism (590S and 591R) in PB2, (with the exception of A/sw/NC/47834/2000, which contains 590S and 591Q and showed lower nasal wash titers and poor transmissibility in ferrets), indicating involvement of other factors in their differential RD transmission. In our experiments, although the TRS viruses with human-like HA and NA replicated efficiently (mean peak nasal wash titer, ∼106 pfu/ml) and were efficiently transmitted in ferrets (Fig. 1E and Table 2), transmission did not always parallel virus shedding. For example, the TRS virus A/sw/NC/18161/2002 (H1N1), with swine-like HA and NA, had an average peak nasal wash virus titer (106 pfu/ml) similar to those of viruses with human HA and/or NA but was least transmissible (1/2 DC, 0/2 RD) in ferrets. In contrast, the TRS virus A/sw/MN/1192/2001(H1N2), with swine-like HA and human-like NA, had a significantly lower average peak nasal wash virus titer (2×105 pfu/ml) than A/sw/NC/18161/2002 (106 pfu/ml) but was transmitted efficiently in ferrets.
H1 HA of North American swine isolates comprises four distinct phylogenetic groups, H1α (cH1N1), H1β (TRS H1N1-like), H1γ (TRS H1N2-like), and H1δ (human-like) [33], [34]. The HAs of some recent TRS H1N1 isolates with swine-like HA and NA are closely related to H1γ (Fig. S1), as are pH1N1 HAs. Recently, two TRS viruses isolated from humans, A/Texas/14/08 (H1N1, H1β) and A/Ohio/2/07 (H1N1, H1γ), showed poor aerosol transmissibility in ferrets [18]. In our study, RD transmission of H1N1 TRS viruses with swine-like H1β HA and swine-like NA was poor in ferrets. However, the H1N2 TRS viruses A/sw/MN/1182/2001 (H1N2) and A/sw/IN/9K035/1999 (H1N2) which possess H1γ HA, were readily transmitted via respiratory droplets, suggesting that human-like NA (N2) is responsible for RD transmissibility of North American TRS viruses containing swine-like H1γ HA. However, studies using reassortant RG viruses are needed to confirm that human-like NA (N2) can enhance the RD transmissibility of TRS viruses containing swine-like HA. Our group and, more recently, others have demonstrated that pH1N1 NA and M can enhance RD transmission of TRS viruses in ferrets [19], [35]. However, others have shown that in a guinea pig model that Eurasian avian-like swine virus (A/sw/Fujian/204/2007[H1N1]) NA and M are not sufficient to alter the non-transmissibility of North American TRS virus (A/sw/Guandong/1222/2006[H1N2]) and that the HA and NS of 2009 pandemic H1N1 virus (A/Beijing/7/2009) contributes to its transmissibility [20].
We found that the North American TRS viruses grew well at 39.5°C. The TRS viruses yielded 2×106 to 2×108 pfu/ml at 39.5°C, while seasonal human H1N1 A/Brisbane/59/07 yielded only 6×103 pfu/ml under identical growth conditions (Fig. 3). This finding may explain the growth of TRS viruses in ferret lungs. At 33°C and 37°C, the yield of seasonal human virus (2×108 pfu/ml) and TRS viruses was similar. Interestingly, at all three temperatures the growth kinetics of the TRS viruses with swine-like HA was very similar to that of the 2009 pandemic A/Mexico/4482/2009 (H1N1) virus (Fig. 3). It was reported that unlike seasonal H1N1 viruses, whose replication is primarily restricted to the upper respiratory tract, the 2009 pandemic H1N1 viruses replicated efficiently in ferret lungs [11]–[13]. A recent study found that replacing the HA of seasonal H1N1 virus A/New York/312/2001 with the HA of 2009 pH1N1 A/Mexico/4108/2009 (swine-like) virus reduced surfactant protein D binding and increased lung pathology in mice, although it did not increase lung virus titers [36]. In our experiments, TRS viruses with either swine-like or human-like HA caused significant lung pathology, yielded high lung virus titers, and replicated efficiently in MDCK cells at 39.5°C. As the lower respiratory tract is warmer, ability to grow at a higher temperature may be responsible at least in part for the efficient lung growth and significant lung pathology in ferrets. Further studies of growth characteristics in primary human respiratory epithelial cells, which may more closely recapitulate the human respiratory tract, are warranted. The molecular determinants of the efficient in vitro growth of the TRS and pH1N1 viruses at 39.5°C and the relation of this growth to their efficient replication in ferret lungs are of interest for future studies.
TRS viruses containing human-like HA and NA showed the highest RD transmissibility in the ferret model, likely reflecting a higher rate of replication (high virus titers in vitro and in vivo), efficient release of progeny virions in the presence of human-like NA, and efficient re-infection in the presence of human-like HA. In contrast, this group of TRS viruses causes only sporadic human infection, indicating a possible limitation of the ferret model. Importantly, however, ferrets used in this study (and in all transmissibility studies) were farm-raised and sero-negative for influenza A viruses. Acquired immunity within the human population, in addition to viral and environmental factors, plays a critical role in human-to-human transmission of influenza A viruses [37]. It is possible that vaccination and pre-exposure to human-like HA and NA in the human population inhibits the spread of this group of viruses in humans. Unlike HA-specific antibodies, NA-specific antibodies do not prevent influenza virus infection, and NA immunity is referred to as infection-permissive [38]. However, humoral immunity induced by NA can markedly reduce virus replication and release, moderating the severity and duration of illness [39]–[42]. Human infections with H1N2 TRS viruses containing swine-like HA have been reported [5], but in humans, unlike the ferret experimental model, transmission is likely to be partially inhibited by NA-mediated immunity to seasonal influenza viruses, including H3N2. Therefore, the rapid worldwide human spread of pH1N1 may be partially explained by its acquisition of Eurasian avian-like swine virus NA and M in a North American TRS genetic background (with swine-like HA) in two ways. First, its RD transmissibility could have been enhanced by the presence of the Eurasian NA and M and second, its pandemic potential could have been enhanced by the absence of immunity to the swine-like HA and NA in the human population.
The pandemic 2009 H1N1 virus is now the predominant human H1N1 influenza virus worldwide. Vaccine against seasonal human H1N1 does not offer significant protection against 2009 pH1N1, and therefore seasonal H1N1 has been replaced by 2009 pH1N1 (with North American swine-like HA and Eurasian swine-like NA) in the World Health Organization's recommended trivalent vaccine. TRS viruses are reported to cause severe lower respiratory tract disease and diarrhea in humans [5]. Here we have shown that unlike seasonal H1N1 (e.g., A/Brisbane/59/2007), whose replication is reported to be restricted primarily to the upper respiratory tract [11], TRS viruses grow efficiently in the lung and cause substantial lung pathology in ferrets. Most importantly, we have shown that H1N1 TRS viruses with human-like HA and NA (which reportedly do not cross-react with antibody to 2009 pandemic H1N1 [43]) are efficiently transmitted in the ferret model, indicating that in the absence of pre-exposure or vaccination to seasonal H1N1, these viruses may be transmissible among humans, especially young children, and therefore are a public health concern.
Four- to six-month-old male ferrets (Triple F farms, Sayre, PA; Marshall Farms, Hazle Township, PA) that were serologically negative for currently circulating influenza viruses by hemagglutination inhibition (HI) assay were used. All animal experiments were conducted in an Animal Biosafety Level 2+ (level 2 with enhanced biocontainment for pandemic H1N1 influenza A virus) facility at St. Jude Children's Research Hospital, in compliance with the policies of the National Institutes of Health and the Animal Welfare Act and with the approval of the St. Jude Children's Research Hospital Animal Care and Use Committee.
MDCK cells were maintained in Dulbecco modified Eagle's medium (DMEM; Invitrogen Corporation, Grand Island, NY) supplemented with 10% fetal bovine serum and antibiotics-antimycotic (Sigma, St. Louis, MO; 100 U/ml penicillin, 100 µg streptomycin, and 0.25 µg amphomycin per ml). Stock viruses were propagated in embryonated chicken eggs at 37°C for 48 h. All isolates underwent a limited number of passages in eggs to maintain their original properties. The genome sequences of A/sw/OK/011521-5/2008 (H1N2), A/sw/OK/011506/2007 (H3N2), A/sw/IN/9K035/99 (H1N2), and A/Mexico/4482/2009 (H1N1) have been described previously [43]–[45]. Our group previously performed genome sequencing and GenBank submission of A/Brisbane/59/2007 (H1N1). For all other viruses, complete genomes were sequenced as described previously [46] and sequences were submitted to GenBank. Sequence alignment and phylogenetic analysis of HA and NA (Fig. S1) and the other six gene segments (data not shown) confirmed Italy/94 (H1N1) as a Eurasian avian-like swine virus and the other nine swine viruses as North American triple reassortants of the subtypes and HA and NA lineages shown in Table 1.
MDCK cell monolayers were inoculated with viruses at an MOI of 0.001 and maintained in virus growth medium (modified Eagle's medium [Invitrogen] containing 1% BME vitamins [Sigma], 0.2% BSA [Calbiochem], 1.6 mg/ml NaHCO3 [Invitrogen], antibiotics-antimycotic [Sigma], and 1.0 µg/ml of tosylsulfonyl-phenylalanyl-chloromethyl-ketone [TPCK]-treated trypsin [Sigma]) (1.6 ml per well in 6-well plates) at different temperatures. Plaque assays were done in MDCK cells in the presence of 1.0 µg/ml TPCK-treated trypsin in agarose overlay medium (virus growth medium containing 0.0015% DEAE-dextran hydrochloride [prepared from dextran of mean molecular weight 500,000; Sigma] and 0.9% ultra-pure low-melting-point agarose [Invitrogen]), as reported previously [47]. After incubation at 37°C for 60 h the plaques were visualized by staining with 0.1% crystal violet solution containing 10% formaldehyde.
Baseline body weight and temperature were documented before inoculation or contact exposure. Four donor ferrets per virus were housed in the lower cages of isolators (configured as shown in Fig. 1A) in ABSL2+ facilities. Air was uniformly circulated at 52 to 57 air changes per hour. Ambient temperature and relative humidity were maintained. The donor ferrets were lightly anesthetized with isoflurane and inoculated with 106 pfu of virus in 0.5 ml PBS (250 µl per nostril). The next day (day 1 p.i.), two donor ferrets were moved into separate cages, each containing one naïve direct contact (co-housed) ferret (Fig. 1A). For respiratory droplet transmission assay, two naïve ferrets were housed separately in cages adjacent to the donor ferrets but separated by double-layered (3 inches apart) grills to allow unobstructed airflow but prevent direct contact. A borazine gun (Zero Toys, Concord, MA) was used to ensure smooth air flow from the left cages to the right cages within each isolator. The donor and recipient ferrets remained housed together from day 1 p.i to day 20 p.i. Weight, temperature, and clinical signs (sneezing, lethargy, and ruffled fur) were recorded every other day for 14 days. Nasal washes were collected on days 2 (donors only), 4, 6, and 8 p.i. by flushing nostrils with total 1.0 ml PBS, and pfu titers were determined in MDCK cells. The two donor animals remaining in the lower cages (Fig. 1A) were euthanized on day 5 p.i. for histopathology of the lung and virus titration of the nasal turbinates, trachea (upper and lower), and lung.
Lung tissue was collected from control (un-inoculated) and virus-inoculated ferrets on day 5 p.i., fixed in 10% neutral buffered formalin, and embedded in paraffin. 5-µm sections were stained with hematoxylin and eosin and examined by microscopy in a blinded fashion. Histopathology was examined separately for the bronchi, bronchioles, alveoli and alveolar interstitial walls of each lung lobe.
Serum samples were collected from ferrets at day 20 p.i., treated for 18 h at 37°C with receptor-destroying enzyme, heat-inactivated at 56°C for 30 min, and tested by HI assay with 0.5% packed chicken red blood cells as described previously [48].
CY058484-91, CY098465-72, CY098473-80, CY098481-88, CY098489-96, CY098497-504, CY098505, CY098506-12, and CY098513-20 for A/Brisbane/59/2007 (H1N1), A/swine/MN/1192/2001 [H1N2], A/swine/NC/47834/2000 [H1N1], A/swine/MN/6998/2003 [H1N1], A/swine/MN/5763/2003 [H1N2], A/sw/Italy/1369-7/1994 [H1N1], A/Mexico/4482/2009 (H1N1), A/swine/NC/38448-1/2005 [H1N1], and A/swine/NC/18161/2002 [H1N1], respectively.
|
10.1371/journal.ppat.1005791 | Two Different Virulence-Related Regulatory Pathways in Borrelia burgdorferi Are Directly Affected by Osmotic Fluxes in the Blood Meal of Feeding Ixodes Ticks | Lyme disease, caused by Borrelia burgdorferi, is a vector-borne illness that requires the bacteria to adapt to distinctly different environments in its tick vector and various mammalian hosts. Effective colonization (acquisition phase) of a tick requires the bacteria to adapt to tick midgut physiology. Successful transmission (transmission phase) to a mammal requires the bacteria to sense and respond to the midgut environmental cues and up-regulate key virulence factors before transmission to a new host. Data presented here suggest that one environmental signal that appears to affect both phases of the infective cycle is osmolarity. While constant in the blood, interstitial fluid and tissue of a mammalian host (300 mOsm), osmolarity fluctuates in the midgut of feeding Ixodes scapularis. Measured osmolarity of the blood meal isolated from the midgut of a feeding tick fluctuates from an initial osmolarity of 600 mOsm to blood-like osmolarity of 300 mOsm. After feeding, the midgut osmolarity rebounded to 600 mOsm. Remarkably, these changes affect the two independent regulatory networks that promote acquisition (Hk1-Rrp1) and transmission (Rrp2-RpoN-RpoS) of B. burgdorferi. Increased osmolarity affected morphology and motility of wild-type strains, and lysed Hk1 and Rrp1 mutant strains. At low osmolarity, Borrelia cells express increased levels of RpoN-RpoS-dependent virulence factors (OspC, DbpA) required for the mammalian infection. Our results strongly suggest that osmolarity is an important part of the recognized signals that allow the bacteria to adjust gene expression during the acquisition and transmission phases of the infective cycle of B. burgdorferi.
| Borrelia burgdorferi, the Lyme disease agent, exploits a multifaceted enzootic cycle that requires a tick vector for successful transmission between mammalian hosts. Two different regulatory systems control genes that are required to complete this infective cycle. The Hk1/Rrp1 two-component system affects genes required for successful transfer between mammal and tick vector while the Rrp2-RpoN-RpoS regulatory cascade modulates genes essential for the transmission from the tick to a new vertebrate host. Data presented in this study indicate that fluctuations in osmolarity in the tick midgut directly affect these two regulatory pathways. Osmolarity in the lumen of the tick adjusts to the osmolarity of the incoming blood (blood meal) to promote water and ion flux into tick tissues. A positive water flux is essential to generate sufficient saliva for prolonged feeding. We propose that B. burgdorferi uses this physiological parameter as an important signal to adapt and regulate genes required for survival in the tick (through Hk1/Rrp1) and transmission to a new host (through Rrp2-RpoN-RpoS).
| Borrelia burgdorferi, the Lyme disease agent, survives and grows in mammals and various vertebrate hosts. However, the bacteria are not transmitted directly to a new host. Instead they are acquired by a hematophagous arthropod (Ixodes scapularis) and transmitted to a new host. Cycling between a host and the vector requires the bacteria to adapt and survive in both mileus. The environment of the host is well defined: nutrient-rich, constant temperature, stable pH, established ion concentrations and osmolarity [1]. Overall, the mammalian host provides the bacteria with a very steady environment for survival provided that they can successfully evade an aggressive host immune system. In contrast, the tick presents a more variable environment with parameters that are gradually changing before, during and after feeding. Acquisition of B. burgdorferi begins when uninfected ticks begin feeding on infected mammals. Initially, this colonization is characterized by rapid growth of the bacteria and regulation of gene expression by the two-component system (TCS), Response Regulator 1 (Rrp1) and Histidine Kinase 1 (Hk1) [2–4]. As the blood meal is consumed, a feast-famine succession that lasts for several weeks slowly converts B. burgdorferi from rapid growth to stationary phase. During this progression, Borrelia adjusts gene expression for long-term survival via regulatory networks mediated by RelBbu (RelA/SpoT homolog), the Borrelia oxidative stress regulator (BosR) and σS (RpoS) [5–7]. After molting to the next developmental stage, the ticks begin the next feeding and parameters in the midgut revert. B. burgdorferi, localized specifically to the midgut, begin to grow and alter gene expression according to reconstituted feeding conditions (replenished nutrients, temperature, etc.). Some of these conditions act as signals to upregulate key virulence and transmission factors (OspC, DbpA, BBA66, etc.) via the Rrp2/RpoN/RpoS regulatory cascade [8, 9].
The way the midgut environmental conditions affect the expression of this regulatory system and virulence factors required for the successful transmission has been extensively studied [10]. In several cases, in vitro conditions have been used to mimic parameters that are suspected to exist or have been measured in the tick midgut or the blood meal [11–15]. In addition, the transcription of virulence related genes has been assayed directly from B. burgdorferi RNA extracted from feeding ticks [16]. Interestingly, one parameter that has been completely overlooked is osmolarity. Because of the extended feeding time of Ixodes ticks (5 to 6 days), water from the blood meal must be recycled through the hemolymph to the salivary glands to generate adequate saliva for prolonged feeding. This water flux is followed by a corresponding flux of ions such as Na+, K+ and Ca2+. In Dermacentor andersonii, Kaufman and Phillips demonstrated, by directly measuring the ion concentration, that the osmolarity changes throughout feeding [17–19]. Early studies suggest that, in Ixodes ricinus, the salivary glands function in osmoregulation and facilitate the recycling of 70% of the water from the blood meal to the salivary glands [20, 21]. These studies suggest that B. burgdorferi should encounter osmotic conditions in the feeding tick midgut that are generated by water and ion flux necessary to produce the saliva required for successful feeding.
Bacteria respond to the physiological changes associated with changes in osmolarity by a process known as osmoadaptation [22–24]. Osmoadaptation is classically associated with the synthesis or uptake of a limited set of molecules called compatible solutes [22, 25, 26]. There are two categories of compatible solutes: solutes that have no effect on growth, and those that do have an effect on growth (osmoprotective molecules) [27]. Bacteria use osmoprotective molecules to modulate their intracellular osmolarity so they can grow and divide [22, 28]. In E. coli, the increasing K+ concentration is directly related to the increase of the environmental osmolarity [29, 30]. Potassium is also known to activate key regulatory proteins that are involved in the regulation of intracellular pH [31, 32]. Bacteria also accumulate, by transport or de novo synthesis, specific amino acids like proline or glutamate [15, 22, 27, 33]. The major role of the glutamate is to offset the uptake of K+, which inhibits enzyme activity [32, 34].
In this study, we determined the effects of changes in osmolarity on the virulence and physiology of B. burgdorferi. First, we measured the osmolarity (mOsm) in the bloodmeal, saliva and hemolymph isolated from feeding ticks, then tested B. burgdorferi cells for their ability to grow over a range of osmolarities. Surprisingly, Borrelia was only able to grow normally between 250 mOsm and 650 mOsm, which very closely matched the range of osmolarity in the bloodmeal during tick feeding (~300–600 mOsm). The growth, morphology and motility were dramatically affected by osmolarity outside of this narrow range. Interestingly, at low osmolarity, Borrelia cells expressed increased levels of virulence factors (OspC, DbpA) required for successful transmission. Finally, we analyzed the osmoadaptation by following the expression of the genes putatively involved in osmoregulation (proU, gltP, etc.), or virulence (ospC, dbpA, etc.) at various osmolarities. Mutants that did not express the putative L-glutamate transporter (gltP) or proline transport (proU) system were more sensitive to changes in osmolarity than wild-type cells, suggesting that L-glutamate and proline were osmoprotective molecules. We hypothesize that bloodmeal osmolarity may directly affect the expression of key virulence factors and may serve as a physiological signal to trigger B. burgdorferi to migrate from the midgut to the salivary glands during transmission.
Based on previous observations in I. ricinus and D. andersonii [17–20], we hypothesized that the osmolarity of the Ixodes tick midgut changes throughout its feeding cycle. To test this hypothesis, we measured the osmolarity of the midgut contents, hemolymph and saliva of I. scapularis in feeding nymph or adult ticks (Fig 1). Initially, we harvested feeding ticks from host animals at specific times (days) after attachment to analyze midgut contents. However, because of a lack of consistency in feeding efficiency between individuals in a feeding cohort, we decided to use scutal index to measure feeding progress. The scutal index is the ratio of the length of the idiosoma to the maximum width of the scutum (see methods, Fig 1A and 1B) [35] and this proved to be an effective method to evaluate the duration of feeding. In both nymph and adult ticks, it was not possible to measure the osmolarity of midgut contents before 24 h of the initiation of feeding due to the extremely low volume of recoverable midgut material.
Therefore, feeding was interrupted by detaching ticks, scutal index were measured, the midgut contents were extracted, osmolarity was measured and the data were plotted as a function of the scutal index (Fig 1C) [35]. In adult ticks, the osmolarity began at ~ 550 mOsm at a scutal index of 2 and then decreased to ~300 mOsm (with the lowest value measured at 264 mOsm) at a scutal index between 5–7. Finally, the osmolarity increased as feeding finished, and in replete ticks, returned to an osmolarity of ~550 mOsm at a scutal index between 7–8 (Fig 1C, Adults). We also measured the changes in osmolarity in the midguts of feeding nymphs. Again, we were unable to measure the osmolarity at early time points. As the scutal index reached 3.5–4, the osmolarity approached ~600 mOsm and then decreased to 300 mOsm at a scutal index of 5 (Fig 1C, Nymphs). Again, as observed in the adult ticks, the osmolarity rebounded to ~500–600 mOsm at the conclusion of feeding. These data showed a very similar pattern of amplitude and fluctuation in osmolarity in the midgut of feeding adults and nymphs.
We also measured the osmolarity of mouse (335 mOsm ± 2.8) and rabbit (304.7 mOsm ± 9.5) blood confirming previously published values [1]. Additionally, hemolymph and saliva were collected from feeding ticks and the osmolarity of biological triplicate samples were measured. Hemolymph (311.4 mOsm ± 33.3) and saliva (323.5 mOsm ± 6.4) had osmolarities very similar to those measured in host blood. Taken together, these data suggest that B. burgdorferi encounters little change in osmolarity in the hemolymph, saliva or in the mammalian host but faces a variation in osmolarity (~275–600 mOsm) in the tick midgut during feeding.
After characterizing the osmolarity in the tick and mammalian blood, we attempted to understand if this dynamic affected B. burgdorferi growth and physiology. Considering that most bacteria and some spirochetes (Leptospira) tolerate a wide range of osmolarity [22, 24, 36], we were skeptical that the range of osmolarity observed in the tick midgut would have much effect on B. burgdorferi. To define the range of osmotolerance of B. burgdorferi, we monitored the growth rate in BSK-II medium at various osmolarities (150 to 1,250 mOsm) and in different concentrations of oxygen (Fig 2A, S1 Fig). As a control, we also monitored the growth rate of E. coli MG1655 in LOS medium over the same osmolarity range. Unlike E. coli, which can tolerate a range of osmolarity between 50 to 1,050 mOsm, B. burgdorferi was only able to grow between 250 and 650 mOsm in microaerobic or anaerobic conditions, with an optimal growth rate between 250 and 550 mOsm. At osmolalities <200 mOsm and >750 mOsm, most of the cells lysed. Transferring “survivors” to fresh BSK-II media (450 mOsm) indicated that these cells could not recover. Under aerobic conditions, B. burgdorferi was more sensitive to the osmolarity (Fig 2A). These data indicated that B. burgdorferi could tolerate a relatively narrow range of osmolarity (250 and 650 mOsm). However, considering the range of osmolarities in the feeding tick, it seems that B. burgdorferi is well adapted to survive in the tick midgut environment. Because of these results, all subsequent experiments were done between 250 and 650 mOsm under microaerobic conditions.
During evaluation of the growth rates of B. burgdorferi at different osmolarities, we observed an effect on motility in higher osmolarity. In addition, cell morphology also was affected. High and low osmolarity are known to have a global effect on the cell physiology and gene regulation in many bacteria [24]. In B. burgdorferi, cellular morphology is critical for proper motility as the cells utilize endoflagella to perform waveform motility [37]. To understand how differing osmolarities affect these aspects of B. burgdorferi physiology, we observed the morphology and motility of cells at the three physiologically relevant osmolarities: 250, 450 and 650 mOsm using dark-field microscopy at mid-log phase of growth (4–5 X 107 cells/ml) (Fig 2B and 2C). At 450 mOsm, the cells displayed normal morphology, i.e. long waveform-shaped cells (Fig 2B). At lower osmolarity (250 mOsm), the cells were slightly longer, with normal motility (Fig 2C). At an osmolarity of 650 mOsm (Fig 2B), ~80% of the cells were non-motile and ~10% had altered motility (twitching) (Fig 2C), and were shorter (Fig 2B). We confirmed by plating that non-motile cell were viable. These observations suggested that osmolarity affected both cellular morphology and motility in B. burgdorferi. The changes in cell shape could indicate an adaptation to a change in water flux. The observed effects on motility at higher osmolarity may reflect physical constraints on flagellar function or may indicate an effect on membrane potential and/or cellular energy.
Because of the observed changes in osmolarity in the feeding tick midgut, we analyzed the production of key proteins involved in successful transmission at different osmolarities. The levels of virulence factors OspC, DbpA and BBA66 increased at an osmolarity of 250 mOsm while OspA increased slightly at higher osmolarity (650 mOsm) (Fig 3A, S2 Fig). Key regulatory proteins involved in the regulation of these virulence related proteins were also assayed. The levels of Rrp2, Rrp1 and BosR did not change significantly at any osmolarity tested. However, RpoN and RpoS, which have been shown to regulate these and other virulence factors, increased at an osmolarity of 250 mOsm (Fig 3A, S2 Fig). More importantly, we directly tested the production of virulence factors in strains B31-A3ΔrpoN and B31-A3ΔrpoS at different osmolarities (Fig 3B and 3C, S2 Fig). While these mutants grew normally at all osmolarities tested compared to wild-type B31-A3, no changes were observed in the production of OspC, DbpA or BBA66 when the mutants were grown at 250 mOsm. These data strongly suggest that the RpoN-RpoS regulatory cascade was involved in the regulation of these virulence factors at lower osmolarity. Also, immunoblots of cell lysates of B. burgdorferi grown at 250, 450 and 650 mOsm were probed using serum from mice infected with B. burgdorferi B31-A3 by tick bite (Fig 3D). Spirochetes grown at 250 mOsm, corresponding to the osmolarity measured at the midpoint in a feeding tick bloodmeal, in tick saliva or in mammalian blood, showed increased reactivity with infected serum (Fig 3D).
We also tested the expression of the genes encoding these proteins by qRT-PCR. At 250 mOsm, similar to osmolarity that was measured in the blood, at the mid-point of feeding and in tick saliva, expression of the sigma factors rpoN and rpoS increased 2.5 and 4.5-fold respectively (Fig 4). We measured the expression of RpoS-dependent virulence factors (ospC, dbpA, bba66, bb0844) and found that similar to the rpoN-rpoS expression pattern, the expression of these four genes increased significantly (6.9, 5.9, 5.7 and 6.8-fold, respectively) at 250 mOsm (Fig 4). Although both rpoN and rpoS showed changes in gene expression in response to lower osmolarity, the sigma factor, rpoD, did not change in response to osmolarity (Fig 4). Regulation of rpoS and RpoS is transcriptional, translational and post-translational [7, 8, 38–41]. Expression analysis of the Borrelia oxidative stress regulator (BosR), which is thought to directly regulate rpoS, indicated that there was no change in transcription or translation of bosR in response to changes in osmolarity (Figs 3 and 4). Additionally, the transcription and translation of rrp2, which is required to activate the RpoN-RpoS cascade, was not affected by osmolarity. Since both BosR and Rrp2 are transcriptional activators, their regulatory effects on RpoN and RpoS might only require activation of these proteins (e.g., oxidation of BosR, phosphorylation of Rrp2) rather than an increase in the transcription or translation of the genes encoding them. Taken together, these data indicate that lower osmolarity could trigger an increase in the expression of key virulence factors in actively growing cells and this increase was directly linked to the RpoN-RpoS regulatory cascade.
B. burgdorferi cells colonizing ticks are exposed to a distinct range of osmolarities during the tick lifecycle (Fig 1C). Previous studies show that the Hk1-Rrp1 TCS is required for tick midgut colonization. hk1 or rrp1 mutants are unable to be acquired by ticks fed on infected mice or introduced by artificial feeding [2]. Because TCSs are known to be involved in osmoregulation (e.g., OmpR) [42], we investigated the possibility that Rrp1 might be involved in the adaptation of B. burgdorferi to tick midgut osmolarities.
To test our hypothesis, we monitored the growth rate of B. burgdorferi strains 5A4, 5A4 Δhk1 and 5A4 Δrrp1 at 250, 450, 650 mOsm (Fig 5A and 5B). The mutant strains were not affected at low osmolarity (250 mOsm) but were dramatically affected at osmolarities >550 mOsm. In fact, Δrrp1 mutant cells in BSK-II media at increased osmolarity lysed completely mimicking the phenotype that has been reported for these mutant strains in ticks (Fig 5B)[2]. Further, hk1 expression has been shown to increase in the bacteria during acquisition by the tick from the host [2]. In this study, hk1 expression increased 5-fold as osmolarity increased from 250 to 650 mOsm (Fig 5C). In contrast, rrp1 and Rrp1 expression did not change at the osmolarities tested (Figs 3 and 4). However, because Rrp1 has been shown to have diguanylate cyclase activity, we measured the levels of cyclic di-GMP (c-di-GMP) in cells at different osmolarities. At 650 mOsm, the intracellular levels of c-di-GMP increased from 220 nM/mg protein (250 mOsm) to 1120 nM/mg protein (Fig 5D). These data suggest that: i) Hk1 and Rrp1 are required for the transition of B. burgdorferi from the mammal (300 mOsm) to the initial conditions in the tick midgut at the beginning of feeding (600 mOsm), ii) Rrp1 enzymatic activity dramatically increases at 650 mOsm, iii) Hk1 and Rrp1 could be sensing changes in osmolarity, and iv) c-di-GMP could act as an effective secondary messenger for the successful acquisition and long-term survival of B. burgdorferi in ticks.
Changes in osmolarity affect B. burgdorferi morphology, motility and virulence factor expression. We next sought to characterize factors demonstrated to aid in osmoadaptation in other bacteria. Osmoadaptation involves both the efflux and influx of osmolytes, as well as ions. Among all of the characterized osmolyte transporters in E. coli or B. subtilis, only the ProU system is found in the B. burgdorferi genome [43]. This system is an ATP-dependent transporter for glycine betaine, proline, and/or choline [44–46]. The ProU locus consists of proV (ATP-binding subunit), proW (integral membrane protein), and proX (periplasmic glycine betaine binding protein) and the genes are found in that order in the B. burgdorferi genome [43]. In other bacteria, the ProU system protects bacterial cells from high osmolarity by scavenging glycine betaine, proline or choline from the growth media [44–46]. To determine whether the ProU system served as an osmoprotectant system in B. burgdorferi, we first analyzed the expression of proV, proW, proX at 250, 450 and 650 mOsm (Fig 6A). proV and proX showed a significant increase in expression at 250 and 650 mOsm when compared to 450 mOsm while proW did not change at any osmolarity tested. This may suggest that these genes may be transcribed from different promoters or a full length transcript may be post-transcriptionally modified. We also measured the gene expression of the proV gene during nymph feeding (Fig 6B). Surprisingly, proV expression remained unchanged at different points of nymph feeding.
To determine if the ProU system played a role in the osmotolerence of B. burgdorferi, we inactivated the ProU locus by deleting proX and tested the pro mutant for survival at different osmolarities. We attempted to delete the entire locus (proX, proW and proV) but we were unable to do so, probably because choline is used to synthesize phosphatidylcholine (a major phospholipid in B. burgdorferi [47]). We grew B31-A3 and B31-A3ΔproX at various osmolarities and the proX mutant showed a narrower range of osmotolerance than the wild-type (Fig 6C). Because of the effect of the proX mutation on growth, we tested the effect of proX inactivation on virulence and strain B31-A3proX was fully virulent in mice (S1 Table). While the growth rates were slower in this mutant than in the wild-type B31-A3 or the complemented strain B31-A3proX pSABG1, B31-A3proX cells, at an osmolarity of 300 mOsm (Fig 6C, denoted by the arrow), were motile and showed an increase in the expression of OspC that was characteristic of wild-type cells at lower osmolarity (Fig 6D). Overall, these data suggest that the ProU locus facilitates osmoadaptation but over a very narrow range of osmolarities and glycine betaine, proline or choline do not expand the osmotolerance of B. burgdorferi cells. It seems remarkable that B. burgdorferi is so very well adapted to living within a narrow range of osmolarities that directly reflects its immediate environment during the infective cycle.
L-glutamate has been described to be involved in osmoadaptation in bacteria and is readily available in mammalian blood [15, 22, 24, 27, 33, 48]. Classically, bacteria, such as E. coli, increase the synthesis of L-glutamate to promote growth at high osmolarity (~1000 mOsm) [26, 48, 49]. Because B. burgdoferi is unable to synthetize L-glutamate [43, 47], we searched for L-glutamate uptake systems in the B. burgdorferi genome and identified two putative transporters for L-glutamate: bb0729 (gltP) and bb0401. Gene expression analyses revealed that only gltP, not bb0401, was differentially regulated in response to changes in osmolarity, increasing 3-fold at 250 mOsm (Fig 7A). We also measured gltP expression before, during and after feeding in nymphs. The expression of gltP increased 5.2-fold during feeding at a scutal index of ~4 and decreased 2-fold in replete ticks compared to unfed ticks (Fig 7B). These data suggested that glutamate might function as an osmoprotectant at lower osmolarity.
Because these data suggest a role for L-glutamate as an osmoprotective molecule, we tested this more directly. An insertion inactivation mutant (B31-5A18gltP) was obtained from the transposon mutant library [50] and this mutant was tested for growth and survival at different osmolarities. As expected from the expression (Fig 7A), strain B31-5A18gltP had a slower growth rate at 300 mOsm (blood osmolarity) than the wild-type strain B31-5A18 or the complemented strain B31-5A18gltP pSABG2 (Fig 7C, denoted by the arrow). As was observed in B31-A3proX, B31-5A18gltP cells, at 300 mOsm, showed normal motility and increased expression of OspC (Fig 7D). The effect of gltP inactivation on virulence was tested and, as with B31-A3proX, B31-5A18gltP was fully virulent in mice (S1 Table). These data showed that: i) gltP expression responded to low osmolarity both in vitro and in vivo, ii) exogenous L-glutamate played a role in osmoprotection at low osmolarity, and iii) L-glutamate transport does not affect survival in mice. Currently, we are trying to test the role of osmoprotectants, such as glutamate, glycine betaine and proline, in ticks by (i) measuring the levels of these molecules in the tick bloodmeal, (ii) generating a B31-A3ΔgltP-ΔproX double mutant, and (iii) testing all mutants in mice and ticks.
To investigate the role of ion transport in osmotolerance, we analyzed the gene expression profiles of ion transport systems identified in the genome of B. burgdorferi [43](Fig 8A). These included the ktrAB transport system (potassium uptake), the K+/Na+/Ca2+ transport system (bb0164), the three Na+/H+ antiporter systems (bb0447 and nhaC-1, nhaC-2) and the Mg2+ uptake system (mgtE, bb0380). The expression of both nhaC-1 and nhaC-2 increased 10-fold at 250 mOsm osmolarity, suggesting an import of H+ and export of Na+ was involved in osmoadaptation (Fig 8A). Expression of the K+/Na+/Ca2+ antiporter system increased 3-fold, suggesting an adaptive flux of K+, Na+ and/or Ca2+ (Fig 8A). Furthermore, the expression of the ktrAB system increased 4.6-fold suggesting that the flux of K+ could augment osmoadaption (Fig 8A). mgtE expression was not affected by the changes in osmolarity which was expected since magnesium has never been shown to have a role in osmotolerence (Fig 8A). Taken together, the gene expression data suggest that the flux of ions would promote survival at low osmolarity.
To confirm that the in vitro analysis was consistent with observed in vivo expression, we analyzed the gene expression of each of the previously mentioned transporters during nymph feeding (Fig 8B). The three Na+/H+ antiporters (bb0447, nhaC-1, nhaC-2) were induced during the feeding, increasing 8.6-fold, 2.1-fold and 2.7-fold respectively (Fig 8B). The expression of bb0447 and nhaC-1 in replete ticks returned to the initial expression level observed in unfed ticks (Fig 8A). Only nhaC-2 stayed at the levels of expression observed during tick feeding (Fig 8B). Taken together, these data suggest that B. burgdorferi alters the expression of its ion transport systems which may allow the bacterium to adapt to changing osmotic conditions in the tick midgut during feeding and in its mammalian hosts. It is also possible that other factors such as ion availability (e.g. sodium) may be affecting the regulation of these transport systems.
B. burdorferi lives in two distinctly different environments: the mammalian host and the tick vector. As the bacteria shuttles back and forth between host and vector, they encounter conditions that are distinct to each setting. For example, when B. burgdorferi are colonizing a mammalian host, they must switch their surface proteins from OspC to VlsE to evade the host immune system [51, 52]. While the exact signal to trigger this change has not been identified, it is clear that the host immune system provides selective pressure to eliminate bacterial cells that have not made the necessary antigenic changes [52]. Surviving cells colonize immune privileged sites existing in a nutrient rich environment with stable physiological parameters (temperature, pH, oxygen, osmolarity, etc.) Conversely, the tick midgut is the locale where B. burgdorferi faces a different set of conditions. Physiological conditions change between flattened and feeding ticks but most would hardly be considered to be extreme. For example, temperature (23°–34°C), oxygen (mostly anaerobic to ~2–3% O2 during feeding), pH (6.8 in flattened or feeding ticks) do not vary significantly while nutrients (nutrient rich to starvation) and reactive oxygen (ROS) or reactive nitrogen (RNS) species may be considered more variable challenges. What is remarkable is that B. burgdorferi is very well adapted to these conditions and senses minor changes in the tick “environment” to regulate expression of key virulence factors. In this study, we characterized another physiochemical parameter, osmolarity, that changed during tick feeding and may be a signal triggering the expression of essential virulence factors (e.g., OspC, DbpA, etc.).
As previously described in other species and genera of ticks, osmolarity fluctuates during acquisition of a blood meal [17–20]. This seemed to be the case for I. scapularis. Midgut contents, isolated from feeding ticks, showed an interesting, triphasic shift from ~600 mOsm to ~300 mOsm returning to ~600 mOsm during the sequential stages of feeding. The physiological reasons for this shift are certainly related to ion and water flux required to balance the effects of non-diffusible or non-transportable anionic polypeptides concentrated in the bloodmeal (Gibbs-Donnan equilibrium) [53]. Clearly, water and ion fluctuations are required for the recycling of water and solutes necessary to generate the amounts of saliva that are required for long-term, successful feeding of I. scapularis.
Interestingly, experiments on wild-type B. burdorferi at different osmolarities indicated that the cells had a narrow range of osmotolerance (Fig 2A) compared to E. coli. Normal doubling times were observed over a range of ~250 to ~650 mOsm under anaerobic and microaerobic conditions, mimicking the conditions observed in the bloodmeal during feeding. The initial observations of cells by dark-field microscopy at different osmolarities indicated that motility was affected as osmolarity reached 650 mOsm. This was of particular interest because Dunham-Ems et al. reported that B. burgdorferi cells have two phases of motility in the midgut of ticks during feeding [37, 54]. Cells were observed to have normal motility and evenly distributed throughout the bloodmeal or were nonmotile and clumped associating with the interior face of the midgut lining. Our observations of the motility of B. burgdorferi suggest that increased osmolarity may be partially responsible for altered motility observed in feeding ticks [37].
Other interesting trends occurred in B. burgdorferi cells at physiologically relevant osmolarities. Immunoblots of protein isolated from cells grown at low osmolarity, indicated that the cells increased the expression of virulence related proteins such as OspC, DbpA and BBA66 (Fig 3A). It has been shown that these proteins are required for the successful transmission and survival of B. burgdorferi in mammalian hosts [6, 7, 14, 55–60]. Analysis by qRT-PCR of RNA isolated from cells grown at 250 mOsm showed an increase in the transcription of ospC, dbpA and bba66 correlating with the increase in expression of these proteins in immunoblots. An increase in the expression of rpoN and rpoS were observed at low osmolarity. Additionally, immunoblot analysis of B31-A3ΔrpoN and B31-A3ΔrpoS indicated that OspC, DbpA and BBA66 were not induced in these mutants at low osmolarity. Since it has been shown that OspC, DbpA and other virulence factors are controlled by the Rrp2-RpoN-RpoS regulatory cascade, it seems very likely that low osmolarity is directly affecting this regulatory network. It is interesting to note that the increased expression of important virulence factors at low osmolarity corresponds to the osmolarity measured at the midpoint of feeding (Fig 1B). It has been shown that transmission of B. burgdorferi occurs ~2 days after the initiation of the feeding, which correlates with the drop in osmolarity measured in the bloodmeal of B. burgdorferi infected ticks. Additionally, these changes were observed in actively growing (mid-log phase), motile cells. It is interesting to speculate that a drop in osmolarity could also serve as a signal to trigger the migration of B. burgdorferi from the midgut to the hemolymph and ultimately to the salivary glands during feeding. However, at this time, we do not have any direct experimental evidence supporting this hypothesis.
High osmolarity (650 mOsm) occurs in the midgut of an unfed tick, at the initiation of feeding and after feeding is complete. Except for a slight increase in the expression of OspA, the expression of other virulence factors remained unchanged at high osmolarity (650 mOsm) compared to cells grown in BSK-II (450 mOsm) (Fig 3). However, high osmolarity not only affected motility but also had another very interesting effect on B. burgdorferi. Analysis of a B31-5A4Δrrp1 mutant indicated that this strain was exquisitely sensitive to osmolarities >500 mOsm compared to strain B31-5A4 and cells rapidly lysed after less than 4h of exposure to increased osmolarity. Rrp1 is the response regulator in the Hk1-Rrp1 TCS and functions as a di-guanylate cyclase [2–4, 61]. C-di-GMP acts as a secondary messenger for signal transduction in bacteria and the levels of c-di-GMP increased dramatically at high osmolarity (Fig 5D). Rrp1 has also been shown to be required for tick colonization, motility and the regulation of genes involved in glycerol metabolism [2, 3, 62]. In addition to its regulatory functions, Caimano et al. showed that B31-5A4Δrrp1 was virulent in mice but this mutant rapidly lysed after being acquired by ticks fed on mice infected with this strain [2]. Also, B31-5A4Δrrp1 rapidly lysed when introduced into ticks by artificial feeding. Collectively, these data suggest that at high osmolarity, Rrp1: i) was required for survival; ii) had increased diguanylate cyclase activity; iii) is required for tick colonization; and iv) could putatively regulate B. burgdorferi motility.
Lastly, we investigated the osmoadaptation of B. burgdorferi. In bacteria, the response to changes in external osmolarity happens at two levels. To restore a conductive intracellular environment, cells transport ionic solutes like K+, Na+ and compatible solutes glutamate, proline and glycine betaine [22, 23]. At low osmolarity, ionic solutes (primarily K+) accumulate while at high osmolarity, compatible solutes accrue to support a high intracellular osmotic pressure without the deleterious effects that ionic solutes have on the activity of metabolic and biosynthetic enzymes [24]. When compatible solutes are not available in the extracellular milieu, the cells will increase their intracellular concentrations by accelerating the synthesis of these important osmoprotectants. Together, these osmoadaptive systems allow bacteria like E. coli and Salmonella typhimurium to tolerate osmolarities from 50–1400 mOsm.
Unlike E. coli or other spirochetes like Treponema denticola and L. interrogans [22–24, 36], the B. burgdorferi genome does not harbor the genes encoding proteins to synthesize osmolytes (e.g., proline, choline or glutamate). However, the genome does have three putative osmolyte transport systems: the proU system for the transport of glycine betaine, proline or choline, as well as bb0729 (gltP) and bb0401 both of which are annotated as glutamate transporters. Transcription of the proU system increased at low and high osmolarity in vitro suggesting that this transport system might be involved in osmoprotection. Additionally, a B31-A3proX mutant strain showed a narrower range of osmotolerance than wild-type B31–A3. However, choline, proline and glycine betaine did not increase the range of osmotolerance of B31-A3. These data indicate that these compatible solutes are required for the survival of B. burgdorferi within the narrow range of osmolarities encountered in the bloodmeal of feeding ticks.
The results for glutamate are distinctly different from what was expected based on previously published information on the role of glutamate in protecting E. coli and S. typhimurium at high osmolarity [48]. As previously mentioned, compatible solutes (e.g., glutamate, proline) protect cells at high osmolarity while ionic solutes (e.g., K+) protect cells at low osmolarity [22, 27, 32, 34]. The mutant strain B31-5A18gltP was more sensitive to low osmolarity while high osmolarity had no effect on the growth and survival of this mutant compared to B31-5A18. Predictably, the expression of the genes encoding ionic solute transport systems such as ktrAB (K+ transport), bb0164 (K+/Na+/Ca2+), bb0447, nhaC-1 and nhaC-2 increased at low osmolarity (250 mOsm). Currently we do not understand why a compatible solute like glutamate is required as an osmoprotectant for B. burgdorferi at low osmolarity but we suspect that it plays a role in the accumulation of ionic solutes in cells as they respond and adapt to low osmolarity. This may be an important function since it has been shown that >70% of the K+ is cycled into the hemolymph and saliva during the feeding of I. ricinus and D. andersonii [17–21]. Clearly, the inability of B. burgdorferi to synthesize compatible solutes has narrowed the limits of their osmotolerance but, despite this, they are finely adapted to the narrow range of osmolarities that they encounter in the tick bloodmeal/midgut and the mammalian host.
While B. burgdorferi cells are well adapted to a narrow range of osmolarity, what was remarkable was that they were using these relatively small changes in osmolarity as a signal to affect at least two regulatory pathways. First, high osmolarity (650 mOsm) has a dramatic effect on motility in wild-type B31-A3, B31-5A18 and B31-5A4. In addition, strains B31-5A4Δhk1 and B31-5A4Δrrp1 did not survive at osmolarities above 500 mOsm. As important, the levels of the secondary messenger molecule, c-di-GMP, increased dramatically at high osmolarity, most likely due to an increase in the diguanylate cyclase activity of Rrp1 [63]. These data suggest a role for Hk1 and Rrp1 in the adaptation to and survival of B. burgdorferi cells at osmolarities of 600 to 650 mOsm. Second, analyses of protein and gene expression in B31-A3, B31-A3ΔrpoN and B31–A3ΔrpoS suggested that B. burgdorferi cells express key virulence factors, such as OspC, DbpA and BBA66 at low osmolarity and this increase in expression was dependent on RpoN and RpoS.
Our current working model (Fig 9) is that as B. burgdorferi cells are acquired by feeding ticks, they rapidly transition from osmolarities of ~300 mOsm in mammalian blood and tissue to ~600 mOsm in the tick midgut at the beginning of feeding. It seems very likely that Hk1-Rrp1 and c-di-GMP are essential for this transition. At the end of acquisition, in replete ticks, the osmolarity returns to ~600 mOsm and motility is impaired, potentially limiting spread of B. burgdorferi and trapping them in the midgut. Long-term survival of cells through the molt is most likely mediated by RelBbu (RelA/SpoT homolog) [5]. At the midpoint of the second feeding, as the osmolarity cycles from ~600 to ~250, motility increases and the cells respond to lower osmolarity by expressing ionic solute transport systems. Most importantly, the RpoN-RpoS regulatory cascade is also stimulated by low osmolarity and triggers the expression of vertebrate virulence-related proteins. At this point, the cells are actively growing, have normal motility and are expressing proteins necessary to promote successful transmission to the next mammalian host. It seems clear that changing osmolarity can affect two different regulatory pathways, Hk1-Rrp1 and RpoN-RpoS, and is potentially a major signal sensed by B. burgdorferi during acquisition and transmission.
The strains used in this study are described in the S2 Table. B. burgdorferi strains were grown in BSK-II medium, pH 6.8 at 34°C [64] under microaerobic environment (5% O2, 5% CO2) and, when indicated, under anaerobic (5% CO2, 5% H2, balance N2) or aerobic condition. Cell densities were determined by dark-field microscopy (Eclipse E600, Nikon, Melville, NY). The osmolarity of the BSK-II medium is 450 mOsm. To obtain high-osmolarity medium, NaCl was added. To obtain low-osmolarity medium, ddH2O was added. BSK-II medium for plating contained 0.6% agarose. Importantly, low-osmolarity BSK-II media was tested to ensure that essential nutrients were not too dilute to support normal growth. This was accomplished by adding NaCl to the dilute BSK-II to adjust the osmolarity to 450 mOsm. Restoring the osmolarity of diluted BSK-II to 450 mOsm restored normal growth of wild-type B31-A3 [65]. For survival assays, various wild-type and mutant strains were grown in BSK-II medium at different osmolarities starting at 1 x105 cells/ml to early stationary phase of growth. Every 24 h an aliquot of each culture was examined by dark-field microscopy and plated on BSK-II. Plates were incubated at 34°C under microaerobic conditions for 7–14 days to allow enumeration of CFU. The cell length (40 cells per slide, 5 slides from 5 independent cultures) was measured using ImageJ software.
E. coli strains were grown in Lysogeny broth [66] or in low osmolarity medium, called LOS (4 g of casein hydrolysate, 0.5 mg of FeSO4, 18 mg of MgCl2, 200 mg of (NH4)2SO4 and 175 mg of K2HPO4 per liter, pH 7.2) [67]. The LOS medium osmolarity is 70 mOsm. To obtain high-osmolarity medium, NaCl was added.
Growth rates were defined during the exponential phase [68]. Briefly, the growth rate is defined by 1/doubling time and expressed in 1/h.
All osmolarities were measured with a Wescor vapor pressure osmometer at 21°C (model 5500, Wescor, Inc., Logan UT, USA) and expressed as milli-osmolar (mOsm).
The proX::himar1-Gm was amplified by PCR from B31-5A18 NP1 proX::himar1-Gm (proUF ACAGATGAGGTTGTAGCAGCA and proUR GCATATACAAACCTACCTGCTC) and cloned into TopoZeroBlunt (Invitrogen, Carlsbad, CA) to obtain Topo0ProX::Gm vector. The resulting plasmid was transformed into low-passage B. burgdorferi B31-A3 strain as described previously [69] and gentamicin-resistant colonies were analyzed by PCR to confirm the inactivation of proX. Mutants were screened using plasmid specific primer sets [25]. Mutant strain B31-A3proX harbored all plasmids except cp9 was used for further characterization.
For the complementation, the proU operon was amplified by PCR using proUF and proUR primers and cloned into PCR-XL-TOPO following the manufacturer’s recommendations (Invitrogen, Carlsbad, CA). The resulting plasmid was digested with SacI-PstI and the proU fragment was cloned into the pKFSS1 [70] shuttle vector digested with the same restriction enzyme to obtain pSABG1.
The gltP gene was synthesized by Genscript, USA and cloned into the pKFSS1 shuttle vector digested with SacI-PstI to obtain pSABG2. The resulting plasmids were transformed into low-passage B. burgdorferi mutants strains as described previously [69] and spectinomycin-resistant colonies were analyzed by PCR to confirm the construction.
RNA samples were extracted from B. burgdorferi cultures using the RNeasy mini kit (Qiagen, Valencia, CA) according to the manufacturer’s protocol. Three independent cultures were used for each osmolarity. Total RNA from ticks was isolated from 3 pools of 7 nymphs fed on mice infected by needle inoculation with B31-A3. RNA samples was extracted using RNeasy mini kit (Qiagen, Valencia, CA). Ticks were frozen at -80°C directly and crushed. TRIzol (Life technologies, Carlsbad, CA) was added with chloroform. After centrifugation, the upper phase was mixed with ethanol 70% (1:1) and loaded onto the provided Qiagen column according to the manufacturer’s instructions. Digestion of the genomic DNA was performed using TURBO DNA-free DNase I (Life Technologies, Carlsbad, CA). The cDNA was synthesized using the Superscript III reverse transcriptase with random primers (Invitrogen, Carlsbad, CA). To determine gene expression levels, a relative quantification method was employed using the enoS gene as a reference gene (S3 Fig). All samples were performed in at least three biological replicates and three technical replicates on a Roche LightCycler 480 System using Green PCR Master Mix (Life technologies, Carlsbad, CA). All primers used for the study are listed in S3 Table. To determine relative gene expression, the LightCycler 480 software version 1.5 was used. The relative quantification was performed following the E-Method using the enoS as a housekeeping gene [71].
For analysis of cell lysates by Western-blot, bacteria were grown to mid-log phase at 34°C in microaerobic conditions. The cells were harvested by centrifugation, washed twice in HN buffer (50 mM HEPES pH7.5, 50 mM NaCl), resuspended in 0.25M Tris-HCl pH 6.8 and lysed by sonication. The protein concentration was determined with Take3 micro-volume plate in a Synergy 2 Multi-Mode plate reader (BioTek Instruments, Winooski, VT, USA). 40 μg of protein was loaded in a 4–20% pre-cast SDS-PAGE gel (Bio-Rad, Hercules CA, USA) and transferred to a nitrocellulose membrane using a Trans-Blot TurboTM blotting system (Bio-Rad, Hercules CA, USA) with a pre-programmed protocol (2.5A, up to 25V, 3 min). Western blotting was performed using standard protocols, i.e. membrane blocking 1 h in 5% nonfat milk in PBS-T (0.1% Tween 20), then incubating 1hr in PBS-T with primary antibodies, washing in PBS-T and then incubating 30 min in PBS-T with Rec Protein A-HRP (1:4,000; Life technologies, Carlsbad, CA, USA) or with the anti-IgY conjugated to HRP (1:50,000; for α-BBA66, Aves Laboratories, Tigard, OR, USA). For the primary antibodies, the following dilutions were used: α-OspC 1:1,000 [72], α-DbpA purified antibody 1:1,000 (Rockland Immunochemicals, Gilbertsville, PA, USA), α-BBA66 1:4,000 [73], α-RpoS 1:500 [74], α-RpoN 1:1,000, α-BosR 1:500, α-OspA 1:2000 (Rockland Immunochemicals, Gilbertsville, PA, USA), α-Rrp1 1:1,000, α-Rrp2 1:2,000 or infected-mouse serum 1:200 (mice infected with wild-type B31-A3 spirochetes by tick bite). Blots were imaged by chemiluminescent detection using Super Signal Pico chemiluminescent substrate kit (Thermo Scientific, Rockford, IL, USA).
Rabbit polyclonal antisera directed against Rrp2 or BosR protein was prepared according to a previously published protocol [72]. Rabbit polyclonal antisera directed against Rrp1 protein was prepared by Rockland Immunochemicals, Gilbertsville, PA, USA.
I. scapularis egg masses (Oklahoma State University) were allowed to hatch and mature in a controlled temperature, humidity and photoperiod environment. RML mice were needle inoculated by intradermal injection with 100 μl of BSK-II containing 1 x 105 B. burgdorferi B31-A3 and after three weeks, infection confirmed by culturing ear punch biopsies. Larval ticks were fed to repletion on infected mice (naïve mice for non-infected cohort), collected and allowed to molt into nymphs and cure in a controlled environment. Nymphal ticks were then fed on naïve RML mice and mechanically removed periodically during the feeding and further processed for osmolarity measurement or RNA isolation as indicated. For infected nymphs, mice were sacrificed 3–6 weeks post inoculation and tissues (ankle joint, bladder and ear) were cultured to verify infection of the ticks through transmission to the naïve animal. Several of the nymphs (infected and non-infected cohorts) were fed to repletion, collected and allowed to molt into adults. After maturation, these ticks were fed on New Zealand White rabbits. Ticks were removed during feeding and further processed for osmolarity determination or RNA extraction as indicated.
c-di-GMP was quantified from B. burgdorferi cultures using the cGMP Direct Biotrak EIA (GE Healthcare, UK) according to the manufacturer’s protocol. Four independent culture samples were used for each condition. Protein was quantified using a Microplate with Synergy 2 plate reader (BioTek, VT, USA)
Scutal index in feeding ticks was determined as previously described [35]. Briefly, for nymphs, the width of the scutum and length of the body (Fig 1A) were measured under a dissecting microscope configured with an ocular micrometer calibrated to a stage micrometer at a given magnification. For adult ticks, a similar procedure was performed except that a hand held magnifying micrometer was used. Because the width of the scutum remains constant and the length of the body increases proportionately during tick feeding, its ratio provides the most reliable and reproducible indicator of feeding progress.
In triplicate, RML mice were inoculated intradermally with 1 x 105 cells in 100 μl BSK-ll with B. burgdorferi strains B31-A3, B31-5A18, B31-A3proX and B31-5A18gltP. Four weeks post-infection, the mice were sacrificed, tissues dissected (ankle joint, bladder and ear) and cultured in BSK-II to confirm the presence of spirochetes. Rocky Mountain Laboratories (RML), NIAID, NIH in Hamilton, MT are accredited by the International Association for Assessment and Accreditation of Laboratory Animal Care.
The blood meal from the midgut of fed Ixodes scapularis adults and nymphs was collected from interrupted and replete ticks using the following methods. Ticks were held behind the basis capituli with fine pointed forceps. With a second set of forceps, the abdomen was pierced and the contents extruded with slight downward pressure into a microfuge tube. Adults were collected individually and nymphs of similar scutal index were pooled to provide sufficient sample subsequent for analyses.
Mouse infection studies were carried out in accordance with the Animal Welfare Act (AWA 1990), the guidelines of the National Institutes of Health, Public Health Service Policy on Humane Care (PHS 2002) and Use of Laboratory Animals and the United States Institute of Laboratory Animal Resources, National Research Council, Guide for the Care and Use of Laboratory Animals. All animal work was done according to protocols approved by the Rocky Mountain Laboratories, NIAID, NIH Animal Care and Use Committee (Protocol Number 2014–021). The Rocky Mountain Laboratories are accredited by the International Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC). All efforts were made to minimize animal suffering.
Prism 6 software (v6.00, GraphPad, San Diego, CA) was used for all statistical analyses. The data were analyzed using an unpaired t test. P<0.05 was considered significant.
|
10.1371/journal.pcbi.1000894 | Reinforcement Learning on Slow Features of High-Dimensional Input Streams | Humans and animals are able to learn complex behaviors based on a massive stream of sensory information from different modalities. Early animal studies have identified learning mechanisms that are based on reward and punishment such that animals tend to avoid actions that lead to punishment whereas rewarded actions are reinforced. However, most algorithms for reward-based learning are only applicable if the dimensionality of the state-space is sufficiently small or its structure is sufficiently simple. Therefore, the question arises how the problem of learning on high-dimensional data is solved in the brain. In this article, we propose a biologically plausible generic two-stage learning system that can directly be applied to raw high-dimensional input streams. The system is composed of a hierarchical slow feature analysis (SFA) network for preprocessing and a simple neural network on top that is trained based on rewards. We demonstrate by computer simulations that this generic architecture is able to learn quite demanding reinforcement learning tasks on high-dimensional visual input streams in a time that is comparable to the time needed when an explicit highly informative low-dimensional state-space representation is given instead of the high-dimensional visual input. The learning speed of the proposed architecture in a task similar to the Morris water maze task is comparable to that found in experimental studies with rats. This study thus supports the hypothesis that slowness learning is one important unsupervised learning principle utilized in the brain to form efficient state representations for behavioral learning.
| Humans and animals are able to learn complex behaviors based on a massive stream of sensory information from different modalities. Early animal studies have identified learning mechanisms that are based on reward and punishment such that animals tend to avoid actions that lead to punishment whereas rewarded actions are reinforced. It is an open question how sensory information is processed by the brain in order to learn and perform rewarding behaviors. In this article, we propose a learning system that combines the autonomous extraction of important information from the sensory input with reward-based learning. The extraction of salient information is learned by exploiting the temporal continuity of real-world stimuli. A subsequent neural circuit then learns rewarding behaviors based on this representation of the sensory input. We demonstrate in two control tasks that this system is capable of learning complex behaviors on raw visual input.
| The nervous system of vertebrates continuously generates decisions based on a massive stream of complex multimodal sensory input. The strength of this system is based on its ability to adapt and learn suitable decisions in novel situations. Early animal studies have identified learning mechanisms that are based on reward and punishment such that animals tend to avoid actions that lead to punishment whereas rewarded actions are reinforced. The study of such reward-based learning goes back to Thorndikes law of effect [1]. Later, the mathematically well-founded theory of reinforcement learning, which describes learning by reward, has been developed [2], [3].
In a general reinforcement learning problem, an agent senses the environment at time via a state , where is the state space of the problem. The agent then chooses an action , which leads to state according to some (in general probabilistic) state-transition relation. The agent also receives some reward signal , which depends probabilistically on the state . By choosing an action the agent aims at maximizing the expected discounted future rewardwhere denotes the expectation and is some discount rate. This general theory has a huge influence on psychology, systems neuroscience, machine learning, and engineering and numerous algorithms have been developed for the reinforcement learning problem. By utilizing these algorithms, many impressive control applications have been developed. Several experimental studies connect the neural basis for reward-based learning in animals to well-known reinforcement learning algorithms. It has been shown that the activity of dopaminergic neurons in the ventral tegmental area is related to the reward-prediction error [4], a signal that is needed for parameter updates in temporal difference learning [3]. These neurons in turn have dense diffuse projections to several important areas including the striatum. In the striatum it was shown that dopamine influences synaptic plasticity [5]. Hence, the principal basis of reward-based learning in this sub-system, although not well understood yet, could be related to well-known reinforcement learning algorithms. However, the learning capabilities of animals such as rodents are still far from reach with current reinforcement learning algorithms. Since physiological experiments are consistent with quite standard reward-based learning schemes, it is reasonable to speculate that the superior learning capabilities of animals is to a high degree based on the ability to autonomously extract relevant features from the input stream such that subsequent reward-based learning is highly simplified (We note that the distinction between feature extraction and reward-based learning is most likely not so strict in the brain. For example, acetylcholine is a prominent neuromodulator in sensory cortical areas which could be utilized for task-dependent feature extraction). In fact, one of the most crucial design questions in the design of a reinforcement learning system is the definition of the state space . Most reinforcement learning algorithms are only applicable if the state space of the problem is sufficiently small. Thus, if the sensory input to a controller is complex and high-dimensional, the first task of the designer is to extract from this high-dimensional input stream a highly compressed representation that encodes the current state of the environment in a suitable way such that the agent can learn to solve the task. In contrast, the nervous system is able to learn good decisions from high-dimensional visual, auditory, tactile, olfactory, and other sensory inputs autonomously. The autonomous extraction of relevant features in the nervous system is commonly attributed to neocortex. The way how neocortex extracts features from the sensory input is still unknown and a matter of debate. Several principles with biologically plausible neural implementations have been postulated. Possible candidates are for example principal component analysis (PCA) [6], [7], independent component analysis [8]–[10], and information bottleneck optimization [10], [11]. One learning algorithm that exploits slowness information is slow feature analysis (SFA) [12]. SFA extracts the most slowly varying features in the input stream (see below). One important property of SFA is that it can be applied in a hierarchical fashion, first extracting local features on the raw input data which are then integrated to more and more global and abstract features. This hierarchical organization is similar to cortical organization for example in the visual system (we note however that the characteristic recurrent organization of cortex where multiple loops provide feedback from higher-level to lower-level processing is not yet exploited in hierarchical SFA architectures). Furthermore, the features that emerge from SFA have been shown to resemble the stimulus tunings of neurons both at low and high levels of sensory representation such as various types of complex cells in the visual system [13] as well as hippocampal place cells, head-direction cells, and spatial-view cells [14].
These features have been extracted from visual input. This hints at the usefulness of SFA for autonomous learning on high-dimensional input streams. In fact, it was shown in [15] that important stimulus features such as object category, the position of objects, or their orientation can be easily extracted by supervised training with high precision from the slow features of a high-dimensional visual input stream. It should be noted that the SFA algorithm is only one particular implementation of learning based on slowness, and there have been various earlier approaches, e.g., [16]–[19]. Slowness has previously been used in some hierarchical models as well [20]–[22].
Unsupervised learning based on the slowness principle (i.e., learning that exploits temporal continuity of real-world stimuli) has recently attracted the attention of experimentalists [23], [24]. It was shown in monkey experiments, that features in monkey infero temporal cortex are adapted in a way that is consistent with the slowness principle [23].
In this article, we propose a learning system where the state space representation is constituted autonomously by SFA. A subsequent neural circuit is then trained by a reward-based synaptic learning rule that is related to policy gradient methods or Q-learning in classical reinforcement learning. We apply this system to two closed-loop control tasks where the input to the system is high-dimensional raw pixel data and the output are motor commands. We thus show in this article for two control tasks on high-dimensional visual input streams that the representation of the SFA output is well suited to serve as a state-representation for reward-based learning in a subsequent neural circuit.
The learning system considered in this article consists of two components, a hierarchical SFA network and a subsequent control network, see Figure 1. The SFA network reduces the dimensionality of the state-space from 24025 to a small number that was chosen to be 64 or less in this article. The decisions of the subsequent control network are based solely on the features extracted by the SFA network.
We tested this learning system on two different control tasks where an agent (a fish) navigates in a 2D environment with analog state- and action-space: a task similar to the Morris water-maze task [25] and a variable-targets task, see section “Tasks”. The state of the universe at time (see below for details) was used to render a 155 155 dimensional 2D visual scene that showed the agent (a fish; for one of the tasks two fish-types with different visual appearance were used) at a position and potentially other objects, see Figure 2. This visual scene constituted the input to the learning system. These tasks are to be seen as generic control tasks of reasonable complexity. The bird's eye perspective used here is of course not realistic for animal agents. As demonstrated in [14] our model should also be able to deal with a first-person perspective, especially in the Morris water-maze. For the variable-targets task this would introduce some complications like the target not being in the field of view or being hidden behind the distractor. On the other hand it would simplify the task, since the agent would not need to know its own position and angle (it could simply center its field of view on the target).
For the training of the system, we distinguish two different phases. In a first phase the SFA network is trained. In this phase, the fish, the target, and the distractor are floating slowly over the 2D space of the environment. The type of fish is changed from time to time (see section “Training stimuli of the hierarchical network”).
In a second phase the control circuit is trained. This phase consists of several learning episodes, an episode being one trial to reach a defined target from the initial fish-position. An episode ends when the target is reached or when a maximum number of time-steps is exceeded.
The hierarchical network described in the next section is based on the Slow Feature Analysis Algorithm (SFA) [26], [27]. SFA solves the following learning task: Given a multidimensional input signal we want to find instantaneous scalar input-output functions that generate output signals that vary as slowly as possible but still carry significant information. To ensure the latter we require the output signals to be uncorrelated and have unit variance. In mathematical terms, this can be stated as follows:
Optimization problem: Given a function space and an I-dimensional input signal find a set of real-valued input-output functions such that the output signals (1)under the constraints(2)(3)(4)with and indicating temporal averaging and the derivative of , respectively.
Equation (1) introduces the -value, which is a measure of the temporal slowness (or rather fastness) of the signal . It is given by the mean square of the signal's temporal derivative, so that small -values indicate slowly varying signals. The constraints (2) and (3) avoid the trivial constant solution and constraint (4) ensures that different functions code for different aspects of the input. Because of constraint (4) the are also ordered according to their slowness, with having the smallest .
It is important to note that although the objective is slowness, the functions are instantaneous functions of the input, so that slowness cannot be achieved by low-pass filtering. Slow output signals can only be obtained if the input signal contains slowly varying features that can be extracted instantaneously by the functions . Note also that for the same reason, once trained, the system works fast, not slowly.
In the computationally relevant case where is finite-dimensional the solution to the optimization problem can be found by means of Slow Feature Analysis (SFA) [26], [27]. This algorithm, which is based on an eigenvector approach, is guaranteed to find the global optimum. Biologically more plausible learning rules for the optimization problem exist [28], [29].
The visual system is, to a first approximation, structured in a hierarchical fashion, first extracting local features which are then integrated to more and more global and abstract features. We apply SFA in a similar hierarchical manner to the raw visual input data. First, the slow features of small local image patches are extracted. The integration of spatially local information exploits the local correlation structure of visual data. A second layer extracts slow features of these features (again integrating spatially local patches), and so on. Such hierarchical architecture is promising because SFA has been applied successfully to visual data in a hierarchical fashion previously [15], [30]. A hierarchical organization also turns out to be crucial for the applicability of the approach for computational reasons. The application of non-linear SFA on the whole high-dimensional input would be computationally infeasible. Efficient use of resources is also an issue in biological neural circuits. It has been suggested that connectivity is the main constraint there [31], [32]. Since a hierarchical organization requires nearly exclusively local communication, it avoids extensive connectivity.
The hierarchical network consists of a converging hierarchy of layers of SFA nodes, and the network structure is identical to that used in [30] (there this part of our model is also discussed in greater length). All required building blocks for the hierarchical network are available in the “Modular toolkit for Data Processing” (MDP) library [33].
We employed neural implementations of two reinforcement learning algorithms, one is based on Q-learning and one is a policy-gradient method.
Neural versions of Q-learning have been used in various previous works on biological reward-based learning, see e.g. [34], [35]. The popularity of Q-learning stems from the finding that the activity of dopaminergic neurons in the ventral tegmental area is related to the reward-prediction error [4], [36], [37], a signal that is needed in Q-learning [35]. In Q-learning, decisions are based on a so-called Q-function that maps state-action pairs onto values that represent the current estimate of the expected total discounted reward given that action is executed at state . For a given state, the action with highest associated Q-value is preferred by the agent. However, to ensure exploration, a random action may be chosen with some probability. We implemented the neural version of Q-learning from [35] where the Q-function is represented by a small ensemble of neurons and parametrized by the connection weights from the inputs to these neurons. The system learns by adaptation of the Q-function via the network weights. In the implementation used in this article, this is achieved by a local synaptic learning rule at the synapses of the neurons in the neuron ensemble. The global signal that modulates local learning is the temporal difference error (TD-error). We do not address in this article the question how this signal is computed by a neuronal network. Several possible mechanisms have been suggested in the literature [37]–[39].
The Q-function was represented by a set of linear neurons that receive information about the current state from the output of the SFA circuit. The output of neuron is hence given by .
Each neuron has a dedicated preferred direction . The Q-value of a movement in direction for the given state is hence given by . The activities of these neurons imply a proposed action for the agent which is a movement in the direction given by the population vector . Here, is the angle of the vector(5)where the vector is the unit vector in direction .
The Q-function is parametrized by the weight values and it is learned by adapting these weights according to the Q-learning algorithm (see [35]):
See Supporting Text S1 for parameter settings.
The second learning algorithm employed was a policy gradient method. In this case, the action is directly given by the output of a neural network. Hence, the network (which receives as input the state-representation from the SFA network) represents a policy (i.e., a mapping from a state to an action). Most theoretical studies of such biologically plausible policy-gradient learning algorithms are based on point-neuron models where synaptic inputs are weighted by the synaptic efficacies to obtain the membrane voltage. The output of the neuron is then essentially obtained by the application of a nonlinear function to the membrane voltage. A particularly simple example of such a neuron model is a simple pseudo-linear rate-based model where a nonlinear activation function (commonly sigmoidal) is applied to the weighted sum of inputs :(6)Here, denotes the synaptic efficacy (weight) of synapse that projects from neuron to neuron , is a bias, and denotes some noise signal. We assume that a reward signal indicates the amount of reward that the system receives at time . Good actions will be rewarded, which will lead to weight changes that in turn make such actions more probable. Reinforcement learning demands exploration of the agent, i.e., the agent has to explore new actions. Thus, any neural system that is subject to reward-based learning needs some kind of stochasticity for exploration. In neuron model (6) exploration is implemented via the noise term . Reward-based learning rules for this model can easily be obtained by changing the weights in the direction of the gradient of (7)where denotes the low-pass filtered version of with an exponential kernel, and is a small learning rate. In our simulations we used for the filtered reward. The update equations for the bias is analogous with .
A single neuron of type (6) turns out to be too weak for some of the control tasks considered in this article. The standard way to increase the expressive power is to use networks of such neurons. The learning rule for the network is then unchanged, each neuron tries to optimize the reward independently from the others [40], but see [41]. It can be shown that such a greedy strategy still performs gradient ascent on the reward signal. However, the time needed to converge to a good solution is often too long for practical applications as shown in Results. We therefore propose a learning rule that is based on a more complex neuron model with nonlinear dendritic interactions within neurons [42] and the possibility to adapt dendritic conductance properties [43].
In this model, the total somatic input to neuron is modeled as a noisy weighted linear sum of signals from dendritic branches(8)where describes the coupling strength between branch and the soma and is a bias. Again, models exploratory noise. At each time step, an independent sample from the zero mean distribution is drawn as the exploratory signal . In our simulations, is the uniform distribution over the interval . The output of neuron at time is modeled as a nonlinear function of the total somatic input:(9)Each dendritic branches itself sums weighted synaptic inputs followed by a sigmoidal nonlinearity (10)where denotes the synaptic weight from input to the dendritic branch of neuron . Update equations that perform gradient ascent on a reward-signal are derived in Supporting Text S2. The derived update rules for the parameters are(11)(12)where and are small learning rates. The update rules can be extended to use eligibility traces that collect the information about recent pre-and postsynaptic states at the synapse in a single scalar value. In this way, previous states of the synapse can be incorporated in the weight change at time , which is driven by the momentary reward signal . In this article however, we rely on the update rules (11) and (12) without eligibility traces. See [41] for an alternative rule of similar flavor.
In our simulations, we needed two control variables, one to control the speed of the agent and one for its angular velocity . Each control variable was computed by a single neuron of this type where each neuron had branches. The nonlinearity in the branches was the tangens hyperbolicus function . Also a logistic sigmoidal was tested which is a scaled version of the tangens hyperbolicus to the image set . Results were similar with a slight increase in learning time. The nonlinearity at the soma was the tangens hyperbolicus for the angular velocity and a logistic sigmoid for the speed . The noise signal was drawn independently for each neuron and at each time step from a uniform distribution in . Detailed parameter settings used for the simulations can be found in Supporting Text S1.
We tested the system on two different control tasks: a task similar to the Morris water-maze task and a variable-targets task.
We implemented this task with our learning system where the decision circuit consisted of the Q-learning circuit described above. In this task, the slowest components as extracted by the hierarchical SFA network were used by the subsequent decision network. The results of training are shown in Figure 5. The performance of the system was measured by the time needed to reach the target (escape latency). The system learns quite fast with convergence after about 40 training episodes. The results are comparable to previously obtained simulation results [34], [35], [44] that were based on a state representation by neurons with place-cell-like behavior. Figure 5B shows the direction the system chooses with high probability at various positions in the water maze (navigation map) after training. Using only the 16 slowest SFA components for reinforcement learning, the system has rapidly learned a near-optimal strategy in this task. This result shows that the use of SFA as preprocessing makes it possible to apply reinforcement learning to raw image data in the Morris water maze task.
The Morris water maze task is relatively simple and does not provide rich visual input. We therefore tested the learning system on the variable-targets task described above, a control task where two types of fish navigate in a 2D environment. In the environment, two object positions were marked by a cross and a disk, and these positions were different in each learning episode. A target object was defined for each fish type and the task was to navigate the current fish to its target by controlling the forward speed and the change in movement direction (angular velocity). The control of angular velocity, the arbitrary target position, and the dependence of the target object on the fish identity complicates the control task such that the Q-learning algorithm used in the water-maze task as well as a simple linear decision neuron like the one of equation (6) would not succeed in this task. We therefore trained the leaning system with the more powerful policy gradient algorithm described above on the slowest 32 components extracted by the hierarchical SFA network.
In order to compute the SFA output fast, we had to perform the training of the control network in batches of 100 parallel traces in this task (i.e., 100 training episodes with different initial conditions are simulated in parallel with a given weight vector. After the simulation of a single time step in all 100 episodes, weight changes over these 100 traces are averaged and implemented. Then, the next time step in each of the 100 traces is simulated and weights are updated). When the agent in one of the traces arrived at the target, a new learning episode was initiated in this trace while other traces simply continued. As will be shown below, the training in batches has no significant influence on the learning dynamics.
Results are shown in Figure 6A,B. The reward converges to a mean reward above which means that the agent nearly always takes the best step towards the target despite the high amount of noise in the control neurons. Figure 7 shows that the trajectories after training were very good. Interestingly, the network does not learn the optimal strategy with respect to the forward speed output. Although it would be beneficial to reduce the forward speed when the agent is directed away from the target, first rotate the agent, and only then move forward, the output of the speed neuron is nearly always close to the maximum value. A possible reason for this is that the agent is directed towards the target most of the time. Thus, the gain in reward is very small and a relatively small fraction of training examples demands low speed.
We compared the results to a learning system with the same control circuit, but with SFA replaced by a vector which directly encoded the state-space in a straight-forward way. For this task with two fish identities and two objects, we encoded the state-space by a vector(17)where is the position of the agent, is its orientation, is its identity, and is the position of the object. Figure 6C,D shows the results when the control network was trained with identical parameters but with this state-vector as input. The Performance with the SFA network is comparable to the performance of the system with a highly informative and precise state encoding.
For efficiency reasons, we had to perform the training of the control network in batches of 100 traces (see above). Because no SFA is needed in the setup with the direct state-vector as input, we can compare learning performance of the control network to performance without batches. The result is shown in in Figure 6C,D (gray dashed lines). The use of small batches does not influence the learning dynamics significantly.
In the environment considered, movement is mirrored if the agent hits a boundary. Since this helps to avoid getting stuck in corners we performed control experiments where the movement in the direction of the boundary is simply cut off but no reflection happens (i.e., the dynamics of the position of the fish is given by and , compare to equations (14),(15)). Results are shown in Figure S1. As expected, the system starts with lower performance and convergence takes about twice as long compared to the environment with mirrored movements at boundaries. Interestingly, in this slightly more demanding environment, the SFA network is converging faster than the system with a highly informative and precise state encoding.
In another series of experiments we tested how the performance depends on the number of outputs from the SFA network that are used as input for the reinforcement learning. Since the outputs of the SFA network are naturally ordered by their slowness one can pick only the first outputs and train the reinforcement learning network on those. For the variable-targets task we tested the performance for 16, 22, 28, 32, and 64 outputs. For 16 outputs the average reward value always stayed below and rose much slower than in the case of 32 outputs. For 28 outputs the performance was already very close to that of the 32 outputs. Going from 32 outputs to 64 did not change the average reward, but in the case of 64 outputs the trajectories of the agent occasionally showed some errors (e.g., the agent initially chose a wrong direction and took therefore longer to reach the target).
We compared performance of the system to a system where the control network is a two-layer feed-forward network of simpler neurons without dendritic branches, see Equation (6). We used two networks with identical architecture, one for each control variable. Each network consisted of 50 neurons in the first layer connected to one output neuron (increasing the number of neurons in the first layer to 100 did not change the results). Every neuron in the first layer received input from all SFA outputs. The learning rates of all neurons were identical. See Supporting Text S1 for details on parameters and their determination. Results are shown in Figure S2. The network of simple neurons can solve the problem in principle, but it converges much slower.
We also compared performance of the system with SFA to systems where the dimensionality of the visual input was reduced by PCA. In one experiment the SFA nodes in the hierarchical network were simply replaced by PCA nodes. We then used 64 outputs from the network for the standard reinforcement learning training. As shown in Figure 8 the control network was hardly able to learn the control task. This is also evident in the test trajectories, which generally look erratic.
In another experiment we used PCA on the whole images. Because of the high dimensionality we first had to downsample the image data by averaging over two by two pixels (reducing the dimensionality by a factor of four) before using linear PCA. The performance was very similar to the hierarchical PCA experiment (the average reward hovered below ). A direct analysis of the PCA output with linear regression [15] indicates that except for the agent identity, no important features such as position of the agent or the targets can be extracted in a linear way from the reduced state representation. For hierarchical SFA, such an extraction is often possible [15]. This hints at the possibility that the state representation given by PCA cannot be exploited by the control network because the implicit encoding of relevant variables is either too complex or too much important information has been discarded.
Several theoretical studies have investigated biologically plausible reward-based learning rules [46]–[55]. On the synaptic level, such rules are commonly of the reward-modulated Hebbian type, also called three-factor rules. In traditional Hebbian learning rules, changes of synaptic plasticity at time are based on the history of the presynaptic and the postsynaptic activity, such that the weight change of a synapse from a presynaptic neuron to a postsynaptic neuron is the product between some function of the presynaptic activity history and some function of the postsynaptic activity history. A third signal that models the local concentration of some neuromodulator which in turn signals some reward, is in many models modulating these Hebbian updates. Such update rules are either purely phenomenological [53], [55] or derived from a reward-maximization principle [47]–[51]. From the viewpoint of classical reinforcement learning, the latter approach is related to policy-gradient methods. Since the learning algorithms in these previous works are based on simple neuron models, they are too weak for the variable-targets task considered in this article. The policy-gradient method used in this article extends the classical single-neuron based policy-gradient approach in the sense that it is based on a more expressive neuron model with nonlinear branches. In this model, both, synaptic weights and branch strengths are adapted through learning. Our approach is motivated by recent experimental findings where it has been shown that not only synaptic efficacies but also the strengths of individual dendritic branches are plastic [43]. Furthermore, it was shown that this type of plasticity is dependent on neuromodulatory signals. Our results (compare Figure 6 to Figure S2) indicate that the neuron model with nonlinear branches can be trained much faster than networks of point-neuron models. This hints at a possible role of nonlinear branches in the context of reward-based learning.
The Morris water-maze task has been modeled before. In [45], a network of spiking neurons was trained on a relatively small discrete state-space that explicitly coded the current position of the agent on a two-dimensional grid. The authors used a neural implementation of temporal difference learning. In contrast to the algorithms used in this article, their approach demands a discrete state space. This algorithm is therefore not directly applicable to the continuous state-space representation that is achieved through SFA. In [34] and [44] the input to the reinforcement learning network was explicitly coded similar to the response of hippocampal place-cells. In [35], the state-representation was also governed by place-cell-like response that were learned from the input data. This approach was however tailored to the problem at hand, whereas we claim that SFA can be used in a much broader application domain since it is not restricted to visual input. Furthermore, in this article SFA was not only used to extract position of an agent in space but also for position of other objects, for object identity, and for orientation. We thus claim that the learning architecture presented is very general only relying on temporal continuity of important state variables.
Although the variable-targets task considered above is quite demanding, the learning system gets immediate feedback of its performance via the reward signal defined by equation (16). By postulating such a reward signal one has to assume that some system can evaluate that “getting closer to the target” is good. Such prior knowledge could have been acquired by earlier learning or it could be encoded genetically. An example of a learning system that probably involves such a circuitry (the critique) is the song-learning system in the songbird. In this system, it is believed that a critique can evaluate similarity between the own song and a memory copy of a tutor song [56]. However, there is no evidence that such higher-level critique is involved for example in navigational learning of rodents. Instead, it is more natural to assume that an internal reward signal is produced for example when some food-reward is delivered to the animal. One experimental setup with sparse rewards is the Morris water maze task [25] considered above. In principle, this sparse reward situation could also be learned if the learning rules (11), (12) are amended with eligibility traces [48]. However, the learning would probably take much longer.
Given the high-dimensional visual encoding of the state-space accessible to the learning system, it is practically impossible that any direct reinforcement learning approach is able to solve the variable-targets task directly on the visually-induced state-space. Additionally, in order to scale down the visual input to viable sizes, a hierarchical approach is most promising. Here, hierarchical SFA is one of the few approaches that have been proven to work well. Linear unsupervised techniques such as principal component analysis (PCA) or independent component analysis (ICA) are less suited to be applied hierarchically. To understand the results, it is important to note that SFA is quite different from PCA or other more elaborate dimensionality reduction techniques [57], [58]. Dimensionality reduction in general tries to produce a faithful low-dimensional representation of the data. The aim of SFA is not to produce a faithful representation in the sense that the original data can be reconstructed with small error. Instead, it tries to extract slow features by taking the temporal dimension of the data into account (this dimension is not exploited by PCA) and disregards many details of the input. Although it is in general not guaranteed that slowly varying features are also important for the control task, slowly varying features such as object identities and positions are important in many tasks. In fact, the removal of details may underlie the success of the generic architecture, since it allows the subsequent decision circuit to concentrate on a few important features of the input. This may also explain the failure of PCA. The encoding of the visual input produced by PCA can be used to reconstruct a “blurred” version of the input image. However, it is very hard to extract from this information the relevant state variables such as object identity or position. But this information can easily be extracted from the SFA output, see [15].
We compared the preprocessing with SFA to PCA preprocessing but not to more elaborate techniques [57], [58] since the focus of this paper is on simple techniques for which some biological evidence exists. Another candidate for sensory preprocessing instead of SFA is ICA. However, ICA does not provide a natural ordering of extracted components. It is thus not clear which components to disregard in order to reduce the dimensionality of the sensory input stream. One interesting possibility would be to order the ICA components by kurtosis in order to extract those components which are most non-Gaussian. Another interesting possibility not pursued in this paper would be to sparsify the SFA output by ICA. This has led to place-cell like behavior in [14] and might be beneficial for subsequent reward-based learning. Information bottleneck optimization (IB) is another candidate learning mechanism for cortical feature extraction. However, IB is not unsupervised, it needs a relevance signal. It would be interesting to investigate whether a useful relevance signal could be constructed for example from the reward signal. Finally, the problem of state space reduction has also been considered in the reinforcement learning literature. There, the main approach is either to reduce the size of a discrete state space or to discretize a continuous state-space [59],[60]. In contrast, SFA preserves the continuous nature of the state-space by representing it with a few highly informative continuous variables. This circumvents many problems of state-space discretization such as the question of state-space granularity. Thus, there are multiple benefits of SFA in the problem studied: It can be trained in a fully unsupervised manner (as compared to IB). By taking the temporal dimension into account, it is able to compress the state-space significantly without the need to discretize the continuous state-space (as compared to [59], [60]). It provides a highly abstract representation that can be utilized by simple subsequent reward-based learning (compare to the discussion of PCA). The possibility to apply SFA in a hierarchical fashion renders it computationally efficient even on high-dimensional input streams, both in conventional computers and in biological neural circuits where it allows for mainly local communication and thus avoids extensive connectivity [31], [32]. The natural ordering of features based on their slowness implies a simple criterion on the basis of which information can be discarded in each node of the hierarchical network (compare to ICA), resulting in a significant reduction of information that has to be processed by higher-level circuits. Finally, SFA is relatively simple, its complexity is comparable to PCA and it is considerably simpler than other approaches for state-space reduction [57]–[60]. Accordingly, biologically plausible implementations of SFA exist [28], [29]. Together with the fact that experimental evidence for slowness learning exists in the visual system [23], this renders SFA an important candidate mechanism for unsupervised feature extraction in sensory cortex.
In this article, we provided a proof of concept that a learning system with an unsupervised preprocessing and subsequent simple biologically realistic reward-based learning can learn quite complex control tasks on high-dimension visual input streams without the need for hand-design of a reduced state-space. We applied the proposed learning system to two control tasks. In the Morris water maze task, we showed that the system can find an optimal strategy in a number of learning episodes that is comparable to experimental results with rats [25]. The application of the learning system to the variable targets task shows that also much more complex tasks with rich visual inputs can be solved by the system. We propose in this article that slowness-learning in combination with reward-based learning may provide a generic (although not exclusive) principle for behavioral learning in the brain. This hypothesis predicts that slowness learning should be a major unsupervised learning mechanism in sensory cortices of any modality. Currently, such evidence exists for the visual pathway only [23]. We showed that learning performance of the system in this task is comparable to a system where the state-representation extracted by SFA is replaced by a highly compressed and precise hand-crafted state-space. Finally, our simulation results suggest that performance of the system is quite insensitive to the number of SFA components that is chosen for further processing by the reinforcement learning network as long as enough informative features are chosen.
Altogether this study provides, on the one hand, further support that slowness learning could be one important (but not necessarily exclusive) unsupervised learning principle utilized in the brain to form efficient state representations of the environment. On the other hand, this work shows that autonomous learning of state-representations with SFA should be further pursued in the search for autonomous learning systems that do not - or much less - have to rely on expensive tuning by human experts.
|
10.1371/journal.pntd.0001139 | Seasonality and Prevalence of Leishmania major Infection in Phlebotomus duboscqi Neveu-Lemaire from Two Neighboring Villages in Central Mali | Phlebotomus duboscqi is the principle vector of Leishmania major, the causative agent of cutaneous leishmaniasis (CL), in West Africa and is the suspected vector in Mali. Although found throughout the country the seasonality and infection prevalence of P. duboscqi has not been established in Mali. We conducted a three year study in two neighboring villages, Kemena and Sougoula, in Central Mali, an area with a leishmanin skin test positivity of up to 45%. During the first year, we evaluated the overall diversity of sand flies. Of 18,595 flies collected, 12,952 (69%) belonged to 12 species of Sergentomyia and 5,643 (31%) to two species of the genus Phlebotomus, P. duboscqi and P. rodhaini. Of those, P. duboscqi was the most abundant, representing 99% of the collected Phlebotomus species. P. duboscqi was the primary sand fly collected inside dwellings, mostly by resting site collection. The seasonality and infection prevalence of P. duboscqi was monitored over two consecutive years. P. dubsocqi were collected throughout the year. Using a quasi-Poisson model we observed a significant annual (year 1 to year 2), seasonal (monthly) and village effect (Kemena versus Sougoula) on the number of collected P. duboscqi. The significant seasonal effect of the quasi-Poisson model reflects two seasonal collection peaks in May-July and October-November. The infection status of pooled P. duboscqi females was determined by PCR. The infection prevalence of pooled females, estimated using the maximum likelihood estimate of prevalence, was 2.7% in Kemena and Sougoula. Based on the PCR product size, L. major was identified as the only species found in flies from the two villages. This was confirmed by sequence alignment of a subset of PCR products from infected flies to known Leishmania species, incriminating P. duboscqi as the vector of CL in Mali.
| Female sand flies transmit a parasite called Leishmania that causes a disease called cutaneous leishmaniasis (CL). Several species of sand flies are found in West Africa, but only one species, Phlebotomus duboscqi, has been proven to transmit the parasite. Cutaneous Leishmaniasis has also been reported from Mali, Central West Africa, but the sand fly transmitting the parasite and its annual abundance has not been established, until now. Sand flies were collected during three consecutive years from two neighboring villages in Central Mali, Kemena and Sougoula, where CL is present. P. duboscqi was collected year-round and was the dominant sand fly inside of and surrounding human dwellings. Other sand fly species, known not to be vectors of CL, were primarily found outside the village. Additionally, P. duboscqi females were found infected with L. major, the same Leishmania species identified from human CL cases in Mali. The estimated infection prevalence of P. duboscqi females was 2.7%. Interestingly, the sand fly abundance and infection prevalence was similar in the two villages despite a previous report indicating a disparate L. major exposure rate in humans. This study greatly enhances our knowledge of CL transmission in Mali, poorly studied in this country to date.
| In West Africa Phlebotomus duboscqi Neveu-Lemaire is the most important vector of Leishmania major, the causative agent of cutaneous leishmaniasis (CL) [1], [2]. P. duboscqi has been incriminated as the vector of L. major in Senegal [3] and suspected as the vector of CL in Burkina Faso [4], Niger [5], [6], The Gambia [7], Ghana [8], Cameroon [9] and Mali [10], [11], [12]. The first report of P. duboscqi in Mali was from Hombori in 1906 [13] with additional reports from Timbuctu in 1913 [14] and from Bamako and Nioro in 1943 [10]. Later work by Lariviere [11] and Desjeux [1] found P. duboscqi in all regions of the country.
Cutaneous Leishmaniasis is endemic in Mali with cases historically occurring in the districts of Nioro and Segou [11], [15]. The first published report of CL in Mali concerned two cases identified from Nioro in 1944 [16]. Later studies reported leishmanin skin test positivity rates between 10 and 61%, suggesting that Leishmania is endemic in Mali [15], [17], [18], [19]. Leishmania major was first identified as the causative agent of CL in Mali by isoenzyme analysis of parasites isolated from skin samples taken from a lesion of a tourist visiting Mopti [20] and a local resident living in the same region [21].
Despite the identification of L. major as the causative agent of CL in Mali, and although suspected as the vector, no one has identified the parasite in P. duboscqi. Here, we report on a three year survey to evaluate the diversity of sand flies and the seasonal abundance of P. duboscqi in Kemena and Sougoula, two villages endemic for CL in the District of Baroueli, Region of Segou, in Central Mali. Furthermore, we report for the first time the detection and annual prevalence of L. major parasites in P. duboscqi sand flies collected from the study sites.
Sand flies were collected from two neighboring villages, Kemena (12°33′ N–6°33′ W) and Sougoula (13°05′ N, –6°53′ W), in the Baroueli Health District, Region of Segou, Mali. Both villages have a population size of approximately 1000 inhabitants. Each village is organized into a labyrinth of adjoining compounds within which a single extended family resides in several sleeping, cooking, and storage houses. Houses are constructed of clay bricks plastered with mud and straw, and with thatched or metal roofs. Domestic animals, such as goats, sheep, and chickens are kept within the confines of a family compound while cows are maintained in corrals located around the perimeter of the village. Both villages have a limited infrastructure and lack electricity and running water. The climate consists of three distinct seasons: a dry season from March to June (temperature range 27–40°C; monthly average rainfall 5.2 mm), a rainy season from June to September (temperature range 25–35°C, monthly average rainfall 82.42 mm), and a third temperate season from October to February (temperature range 20–35°C; monthly average rainfall 3.3 mm). Vegetation is sparse and is characterized by the presence of sporadically placed trees such as shea (Vitellaria paradoxa), acacia (Faidherbia albida) and neem (Azadirachta indica) and small bushes. Most of the land surrounding each village is dedicated for agricultural use.
Sand flies were collected using 1) dark activated, CDC miniature light traps fitted with double ring fine mesh collection bags (John W. Hock Company, Gainesville, FL), 2) sticky traps consisting of single sheets of A4 paper (21×29.5 cm) coated on both sides with castor oil and mounted vertically on pegs, onto which randomly impinging sand flies would adhere (used for the sand fly diversity study only), and 3) mouth aspirators (John W. Hock Company, Gainesville, FL) for collection of resting flies inside of houses used for sleeping. All sand flies were sorted by sex, species and blood meal status, and placed in tubes containing silica gel and cotton until processed. Minimum and maximum temperatures, rainfall and relative humidity for the months of July 2006 to June 2008 were collected from the nearest available weather station in Segou, Mali. Oral informed consent was obtained from head of households for indoor collection of sand flies. Households where consent was given were listed in a written log kept by the entomological team for reference.
The head and terminal segments of the abdomen containing the genitalia of each sand fly were carefully removed and placed into 96-well plates containing a solution of lacto-phenol clearing solution (Bioquip). After 24 h incubation at room temperature, the head and terminalia were fixed onto a glass slide, examined using a light microscope and identified using dichotomus keys [22].
From June 2006-July 2008, the abdomens of blood fed and non-blood fed Phlebotomus females from the same collection location were grouped in pools of no more than 20 individuals and placed in a microfuge tube containing lysis buffer (5.84 g/L NaCl, 68.5 g/L Sucrose, 12.10 g/L Tris, pH 9.1, 100 ml EDTA 0.5 M solution and 50 ml 10% SDS solution). After incubating overnight at 4°C the tissue was macerated using a pestle for 2 min then incubated for 30 min at 65°C. After the addition of 10 µl cold potassium acetate the samples were incubated for 30 min at 4°C and then centrifuged for 10 min at 14,000 RPM. The DNA was precipitated using 70% ethanol and resuspended in 100 µl water. The DNA concentration of each extraction was determined using a NanoDrop (Thermo Scientific Inc., Wilmington, DE). Samples with less than 4 ng/µl of DNA were removed from the sample set. Leishmania DNA was detected by PCR using forward and reverse primers for Leishmania sp. (Uni21/Lmj4) as described in [23]. PCR Primers targeting the sand fly tubulin gene were used as a control for template fidelity (PpTub-P24F 5′-GCG ATG ACT CCT TCA ACA C and PpTub-P24R 5′-TCA GCC AGC TTG CGA ATA C) [24].
A representation of PCR products was confirmed by DNA sequencing. Due to difficulties with direct sequencing of the PCR products using the Uni21 and Lmj4 primers, gel-purified PCR products were cloned into the pCR4-TOPO vector using the TOPO TA Cloning Kit for Sequencing (Invitrogen, Carlsbad CA) following the manufacturer's instructions. The clones were sequenced directly using the M13 forward and M13 reverse primers. Resulting sequences were analyzed using DNASTAR sequence analysis software (DNASTAR, Inc., Madison WI). Sequences were compared to published sequences of kDNA from L. major (Genbank Accession J04654), L. infantum (AF188701), L. tropica (Z32841), and L. donavani (AF167718) using BLAST (http://blast.ncbi.nlm.nih.gov/), aligned to known Leishmania minicircle kinetoplastic DNA using Clustal [25] and edited using BioEdit (http://www.mbio.ncsu.edu/BioEdit/page2.html).
To estimate the prevalence of infection in pooled samples of P. duboscqi females, we used the maximum likelihood estimate (MLE) of prevalence accounting for pooling with the confidence interval (CI) estimated by exact methods if the number of unique pool sizes was less than or equal to 3 [26], or otherwise by the skewness-corrected score confidence interval [27]; the estimates and both CIs were calculated using the binGroup R package [28].
To model the sand fly counts or infection rates we used a quasi-Poisson model and tested for significant effects using analysis of deviance and F test [29]. To test for seasonal effects, we tested the overall effect of months after controlling for previous counts and year. In testing for weather effects, we compared models with previous counts, year and months and tested to see if models that additionally added the previous month weather variables (including 4 weather variables at a time; selecting only one [minimum or maximum] of temperature or wild velocity variables) significantly improved the fit. For the models of rates, we estimated the number of infected flies of those tested by the MLE of prevalence and used those counts as responses in the quasi-Poisson model with an offset based on the number of flies tested so that the inferences describe effects on the rates [29]. The quasi-Poisson models were performed using R version 2.12 [30]. Graphs were made using GraphPad Prism 5 (Graphpad Software, California, USA).
From March 2005 to June 2006, 18,595 sand flies were collected in the two villages (9,887 in Kemena and 8,708 in Sougoula) using all three collection methods. Approximately equal numbers of male and female sand flies were collected (9,221 M, 9,374 F). Sixty-nine percent (n = 12,952) of sand flies were identified as one of 12 species in the genus Sergentomyia, none of which have been implicated in the transmission of L. major (Table 1).
Of the Sergentomyia, Sergentomyia schwetzi Adler, Theodor and Parrot represented the majority with 47.3% of collected specimens, while Sergentomyia antennata Newstead was the second most abundant at 26.4%. Ten additional Sergentomyia species were collected: Sergentomyia dubia Parrot, Mornet, and Cadenat (12.2%), Sergentomyia clydei Sinton (7.9%), Sergentomyia africana Newstead (3.2%), Sergentomyia squamipleuris Newstead (1.7%), Sergentomyia affinis vorax Parrot (0.56%), Sergentomyia bedfordi Newstead (0.49%), Sergentomyia fallax Parrot (0.02%), Sergentomyia buxtoni Theodor (0.13%), Sergentomyia darlingi Lewis and Kirk (0.06%), and Sergentomyia christophersi Sinton (0.01%). The remaining 30% of sand flies collected was identified as one of two species of Phlebotomus, the overwhelming majority of which was P. duboscqi (n = 5,643, 99.3%). Only 41 Phlebotomus rodhaini Parrot (0.7%) were collected (Table 1).
Sticky traps and light traps collected sand flies in about equal numbers (n = 8,290 vs. 8,394), yet the majority of Sergentomyia (n = 7,728, 60%) were collected using sticky traps whereas only 10% of Phlebotomus (n = 562) were collected using this method. The majority of Phlebotomus (n = 3,380, 60%) were collected using light traps. Thirty percent of Phlebotomus (n = 1,701) were collected by resting site collection compared to only 1.62% (n = 210) of Sergentomyia. Comparing sticky trap and light trap collections from inside and outside houses,the majority of Phlebotomus (92%, n = 3,641) were collected inside dwellings whereas the majority of Sergentomyia (71%, n = 9,043) were collected outside dwellings.
From July 2006 to June 2008, 7,950 P. duboscqi (3,998 female) were collected. Additionally, a total of 25 P. rodhaini were collected during the two years, 17 of which were collected during one month in Kemena (October 2006). Comparing the total number of P. duboscqi collected during year one (July 2006–June 2007) and year 2 (July 2006–June 2008), we found that 1.42 times more sand flies were collected in year 2 (p-value 0.0002, 95% CI: 1.19–1.69) (Figure 2A). We observed a similar effect when comparing the annual collections of female P. duboscqi (p-value 0.0003) (Figure 2B). Using the quasi-Poisson model, controlling for year and previous count, we observed a significant seasonal effect reflecting the month to month variation in the total number of sand flies collected (p-value <0.0001). A similar effect was observed when we considered only female sand flies (P-value <0.0001). Monthly collection trends were similar during both collection years. We modeled sand fly counts for each month using January, the lowest seasonal collection month, as a reference. An initial peak with a 3.9 fold change from January [FCJan] (95% CI: 2.6, 6.1) was observed in May and July (3.8 FCJan, 95% CI 2.4, 6.0). This was followed by a dip in collections in August (2.0 FCJan, 95% CI: 1.2, 3.3) and September (1.3 FCJan, 95% CI: 0.8, 2.2) and a second upward trend peaking in November (3.6 FCJan, 95% CI: 2.4, 5.8) (Figure 2A). By village, we found that 45% (n = 3,654) of all P. duboscqi (male and female) were collected in Kemena and 54% (n = 4,276) in Sougoula. Using the quasi-Poisson model, controlling for previous count, month, and year we observed a significant difference in total P. duboscqi counts between the two villages (p-value 0.0293) with the sand fly counts 1.188 times higher, on average, in Sougoula than in Kemena (95% CI: 1.025,1.377) (Figure 2A). Similar results hold when using only female sand fly counts (fold-change = 1.155, p = 0.1116, Figure 2B). The various weather variables (relative humidity, rainfall amount, maximum or minimum temperature, and maximum and minimum wind velocity) were not useful for predicting the observed total or female sand fly collections for either village (all models had p-value >0.53) (Figure 3B).
The majority of P. duboscqi was collected by resting site collection, particularly during the morning (10.54 and 14.04 female P. duboscqi/person/hour during morning collections in Kemena and Sougoula, respectively, compared to 5.12 and 6.10 P. duboscqi/person/hour during evening collections) (Figure 4). On average, five times more P. duboscqi were collected using light traps placed inside of dwellings than outside in the same compound (1.74 vs. 0.33 and 1.78 vs. 0.31 P. duboscqi females/trap/night in Kemena and Sougoula, respectively) (Figure 4). Virtually no P. duboscqi were collected in the light traps placed outside of the village near natural tree holes (0.03 and 0.12 P. duboscqi females/trap/night in Kemena and Sougoula, respectively).
A total of 1434 pools (3706 total flies; average 2.6 flies per pool) were examined for Leishmaina infection by PCR. Ninety-seven pools were positive for L. major (Figure 5). Assuming that the sand flies are independently distributed in the pools and the size of the pools is not related to the probability of infection of the pool, we estimate the prevalence of infection to be 2.66%, 95% CI: 2.20, 3.21 (Table 2). Infected P. duboscqi were found during each month of the year, although monthly infection estimates varied greatly year to year, being the highest during September 2006 (9.64%; CI: 4.68, 17.34) and February 2008 (9.19%; 95% CI: 5.03, 15.27) (Figure 6). After controlling for sand fly count, there was no significant difference in the rates of infection from year 1 to year 2 (p-value 0.2572) and neither was there a significant month to month difference (p = 0.2085). The estimated infection prevalence of sand flies was virtually the same for both villages (2.65%, 95% CI: 1.97, 3.51 and 2.67, 95% CI: 2.02, 3.44 for Kemena and Sougoula, respectively) with no significant difference between the two villages (p-value 0.8894) (Table 2). Of the sand flies collected by light traps and resting site collections within compounds, the majority of infected sand flies were collected using light traps versus resting site collection (4.14% of vs. 0.86%). The highest estimated prevalence of infected sand flies was collected from compound 5 in Kemena (9.42%; 95% CI: 5.50, 14.92) and compound 1 in Sougoula (5.94%; 95% CI: 2.97, 10.57). Comparing the position of the light traps, the estimated infection prevalence of sand flies collected in light traps placed directly outside dwellings was higher than those placed inside dwellings (8.15% vs. 2.07%); no infected sand flies were collected from light traps placed in trees outside of either village (Table 2). The estimated infection prevalence of flies that were non-blood fed at the time of collection was 4.01% versus 1.24% for those that were blood fed.
Eight representative PCR products from infected wild caught P. duboscqi were sequenced using primers specific to the kinetoplast minicircle DNA of L. major. Sequence analysis confirmed that all the samples were similar to published L. major sequences based on length of the product and primer region identity. Blast analysis indicated a best match to L. major kinetoplast DNA (Genbank Accession number Z32842.1, E-value 9e-48). Further alignment of the sequences obtained from this study with known Leishmania sequences of other species in Genbank confirmed that the 650 bp fragment size, observed on gel electrophoresis of the PCR products, as characteristic of L. major strains.
Killick-Kendrick [31] suggested the following criteria for incrimination of a vector sand fly: proven anthropophilic behavior and isolation and identification from the sand fly of the same species of Leishmania that infects man. Further evidence such as the demonstration that the sand fly feeds on the reservoir host (if known), concordance between the geographic distribution of the suspected sand fly and human disease, proof that the parasite develops in the fly and experimental transmission of the parasite by the bite of the fly can reinforce the incrimination. Based on monthly collections of sand flies over three years in two villages in central Mali, where CL is known to be endemic, we have demonstrated that P. duboscqi is the predominant Phlebotomus species; that it persists throughout the year; that females are primarily collected inside houses in both villages;and that it has an overall infection rate with L. major of 2.66% as demonstrated by PCR. This strongly points to P. duboscqi as the primary vector of L. major in Mali.
Species of the sub-genus Sergentomyia constitute the majority of sand flies collected in both villages with S. schwetzi being the most abundant. Members of the Sergentomyia genus are known to transmit Sauroleishmania among lizards. Sergentomyia schwetzi is the only Sergentomyia species known to be anthropophilic and was considered a possible vector by Parrot [5] in 1943. Later, Lawyer [32] concluded that despite the anthropophilic behavior of S. schwetzi, it was not a vector of Leishmania in humans. In this study, only 1.6% of all Sergentomyia sand flies were found during resting site collections and the overall majority (70%) was collected outside of dwellings, further supporting the exophilic nature of this sub-genus and the improbability that Sergentomyia sand flies are involved in transmission to humans in our two villages.
Two Phlebotomus species were found during the three collection years, P. duboscqi and P. rodhaini, both of which are known vectors of L. major elsewhere in West Africa. While fewer in number than Sergentomyia species, P. duboscqi was predominantly collected by resting site collection from sleeping dwellings and five times more P. duboscqi females were collected in light traps placed inside than outside of dwellings, supporting the anthropophilic nature of this fly. P. duboscqi was collected in similar numbers in both villages year round with two seasonal peaks, May-July and October-November. These results are consistent with Lariviere [11] who reported on the seasonality of 191 P. duboscqi collected in Mali. The collection of P. duboscqi throughout the year is probably the result of having constant monthly temperatures and a relative humidity that does not drop beyond 18%. However, non of the specific weather parameters tested could be significantly correlated with sand fly collections in either village (Figure 3). Few specimens of P. rodhaini were collected throughout the study period indicating that this species probably does not play a role, or plays a minor role, in the transmission of L. major in Central Mali.
To further incriminate P. duboscqi as the vector of Leishmania in our study villages, we tested 3706 specimens (in 1434 pools) for the presence of Leishmania DNA by PCR. We found that 97 of the pools tested positive for an overall estimated sand fly infection prevalence of 2.66%. None of the infected pools contained P. rodhaini. Since the infection rate in wild-caught sand flies is usually low [31], PCR was used to permit the efficient screening of a large number of specimens. Having established the infection rate of P. duboscqi in this region, we plan to isolate a viable culture of L. major, necessary to type the strain using traditional methods such as isoenzyme analysis. It is worth noting that in 2006 we established a colony of P. duboscqi collected from our two study villages. Subsequently, females from this colony were used successfully to transmit L. major to an animal model of CL [33] further supporting the status of this species as a competent vector of CL in Central Mali.
A recent study by our group [19] found that there is an unexplained discrepancy between the prevalence of leishmanin skin test (LST) positivity in our two study villages, Kemena (45% LST positive) and Sougoula (20% LST positive), despite the fact that the villages are geographically and demographically similar and are only 5 km apart. Furthermore, this discrepancy was consistent over two consecutive annual incidences (18% and 17% in Kemena vs. 5.7% in both years in Sougoula). We hypothesized that the sand fly density and infection prevalence may explain the dissimilar LST results. The two year seasonality study revealed that slightly more female P. duboscqi were collected in Sougoula than in Kemena, yet almost the same percentage of pools were infected in each village (2.67% vs. 2.66%, respectively), thus neither abundance nor infection prevalence can explain the disparate LST rates observed in the two villages [19].
Rodent species are well known reservoirs for L. major throughout its distribution range. The contribution of reservoirs to the observed disparity of LST positivity in the two villages remains to be evaluated. In West Africa, including Senegal where P. duboscqi has been incriminated as the vector of L. major, infected Mastomys erythroleucus, Tatera gambiana and Arvicanthis niloticus have been reported [34], [35], [36], [37]. All three species are found in Mali (T. Schwan, personal communications) and represent potential reservoirs of L. major in Kemena and Sougoula. Indeed, rodent burrows were observed in many of the houses where light traps were placed. Apart from the potential role of these rodents as reservoirs, their burrows also represent suitable sand fly breeding sites and a source of infected flies. Furthermore, all compounds in our study villages contain goats and chickens living in close proximity to houses used for sleeping which also represent good sand fly breeding sites for uninfected flies. A comprehensive study of the rodent population density and infection prevalence in the two villages is needed to fully understand the infection dynamics in both flies and people.
In summary, we have established, for the first time, the diversity of sand flies in two villages endemic for L. major in Central Mali and demonstrated by PCR that P. duboscqi is the primary vector. This work represents the most comprehensive analysis of P. duboscqi, to date, in Mali and further supports the endemic nature of CL in Central Mali. Further investigations of this nature are needed in West Africa.
|
10.1371/journal.ppat.1003129 | Kinetics of Antigen Expression and Epitope Presentation during Virus Infection | Current knowledge about the dynamics of antigen presentation to T cells during viral infection is very poor despite being of fundamental importance to our understanding of anti-viral immunity. Here we use an advanced mass spectrometry method to simultaneously quantify the presentation of eight vaccinia virus peptide-MHC complexes (epitopes) on infected cells and the amounts of their source antigens at multiple times after infection. The results show a startling 1000-fold range in abundance as well as strikingly different kinetics across the epitopes monitored. The tight correlation between onset of protein expression and epitope display for most antigens provides the strongest support to date that antigen presentation is largely linked to translation and not later degradation of antigens. Finally, we show a complete disconnect between the epitope abundance and immunodominance hierarchy of these eight epitopes. This study highlights the complexity of viral antigen presentation by the host and demonstrates the weakness of simple models that assume total protein levels are directly linked to epitope presentation and immunogenicity.
| A major mechanism for the detection of virus infection is the recognition by T cells of short peptide fragments (epitopes) derived from the degradation of intracellular proteins presented at the cell surface in a complex with class I MHC. Whilst the mechanics of antigen degradation and the loading of peptides onto MHC are now well understood, the kinetics of epitope presentation have only been studied for individual model antigens. We addressed this issue by studying vaccinia virus, best known as the smallpox vaccine, using advanced mass spectrometry. Precise and simultaneous quantification of multiple peptide-MHC complexes showed that the surface of infected cells provides a surprisingly dynamic landscape from the point of view of anti-viral T cells. Further, concurrent measurement of virus protein levels demonstrated that in most cases, peak presentation of epitopes occurs at the same time or precedes the time of maximum protein build up. Finally, we found a complete disconnect between the abundance of epitopes on infected cells and the size of the responding T cell populations. These data provide new insights into how virus infected cells are seen by T cells, which is crucial to our understanding of anti-viral immunity and development of vaccines.
| The presentation of virus peptides (epitopes) to CD8+ T cells plays a pivotal role in anti-viral immunity. Recognition of these epitopes presented on MHC class I drives CD8+ T cell priming following interactions with professional antigen presenting cells (APC) and subsequently allows control of infection through killing of infected cells and secretion of cytokines. The process of MHC class I antigen presentation is complex and multi-staged. It starts with degradation of polypeptides, typically by the proteasome, followed by transport to the ER, loading onto MHC class I and finally egress to the cell surface [1]. Along the way other proteases and chaperones refine the peptides and perform quality control functions on peptide-MHC complexes (pMHC) [2]. Surprisingly, despite the large coding capacity and therefore antigenic potential of many viruses, CD8+ T cell responses are often skewed towards a small number of peptides in a phenomenon known as immunodominance [3]. This is exemplified by studies of humans and animals infected with large, complex dsDNA viruses, such as herpes- and poxviruses, where reproducible CD8+ immunodominance hierarchies emerge. For example, up to 20% of the CD8+ T cell response following infection of C57BL/6 mice with vaccinia virus (VACV) is directed towards a single immunodominant epitope and a handful of subdominant specificities account for much of the remainder [4], [5]. Further, while MHC class I antigen presentation is well understood in principle [6] and bioinformatic predictions of MHC class I binding are often highly refined [7], prediction of antigenicity and immunogenicity have remained elusive.
In part this gap remains because kinetic studies to date have focused on single peptides [8] and broader scale studies of antigenicity have been limited to single time points [9]–[11]. This has reflected limitations of technology in that the best reagents for quantifying antigen presentation have been the few monoclonal antibodies generated to date that recognise specific pMHC complexes [8], [12]–[15]. Proteome-wide biochemical approaches have typically required prohibitively large numbers of cells (1×109 and greater) restricting experiments to single time points [16], [17] . Although we have good examples showing the diversity of native virus epitopes presented and we know the consequences of manipulating expression levels and even translation rates for presentation of model antigens [8], [18], this information remains disconnected. As a consequence, while it is clear that increasing expression of a given antigen leads to higher presentation of epitopes, it is not known whether antigen expression level per se is a useful predictor of likely antigenicity across different viral proteins. Further, whether bulk protein abundance or expression levels correlate best with production of epitopes as a general rule is not known. Indeed, several recent studies have highlighted the diversity of source for MHC class I bound peptides and have implicated both products of translational infidelity (defective ribosome initiation products (DRiPs)) [10], [19]–[22] as well as mature proteins [23]. For instance, some biochemical surveys of epitope versus transcript or steady-state antigen abundance suggest these are closely related at single time points [16], [24]. However, most epitopes studied in detail are shown to be the products of recent translation and therefore need not be related to final antigen abundance [25]–[28]. Only studies that can link the kinetics of antigen synthesis and accumulation with epitope presentation for multiple native virus proteins will allow general conclusions to be drawn. Finally, antigen expression levels can be linked to immunogenicity for model antigens, but again whether this is useful for evaluating whole viral proteomes has not been approached.
Here we present the first study that links the kinetics of virus protein build up and CD8+ T cell epitope presentation for multiple pMHC complexes. We used vaccinia virus, best known as the vaccine used to eradicate smallpox, taking advantage of its robust in vitro infections and a well characterised CD8+ T cell epitope hierarchy [4], [5]. In addition there is good evidence that anti-VACV CD8+ T cells are directly primed by infected APC making this an ideal choice to study antigen presentation in vitro [29]–[31]. The abundance of 8 VACV epitopes was quantified simultaneously at multiple times after infection using the multiple reaction monitoring approach to tandem mass spectrometry [32]. The same method was applied in parallel to determine relative abundance of the relevant virus proteins using filter assisted sample preparation and whole cell tryptic digestion [33]. Together, these data provide an unparalleled insight into the dynamic nature of antigen presentation on class I during a virus replication cycle. Further they provide the most compelling evidence to date of the direct correlation between the timing of virus antigen expression and the appearance of epitopes derived from the same protein. Finally, while we can now add kinetics to our description of epitope presentation for multiple epitopes, these biochemical data still fail to predict the hierarchy of immunodominance in responding CD8+ T cell responses.
Previous studies aimed at understanding antigen presentation kinetics have focussed on single epitopes, most commonly the model peptide SIINFEKL (presented by H-2Kb) expressed from recombinant viruses, including VACV. Whilst these experiments have yielded much useful mechanistic insight, it is not clear whether kinetic data generated are representative of virus epitopes in general. To examine this issue, we first recapitulated published data showing the rapid rise of H-2Kb-SIINFEKL complexes on cells infected with a recombinant VACV strain WR-NP-S-GFP [8], [13]. This virus expresses a chimera in which SIINFEKL is sandwiched between influenza virus nucleoprotein and enhanced green fluorescent protein [8], [34]. DC2.4 cells, a dendritic cell-like line derived from C57BL/6 mice, were infected at a multiplicity of 10 pfu per cell and presentation of Kb-SIINFEKL complexes measured using the mAb 25D1.16 and flow cytometry at various times (Figure 1A). Consistent with previous work that typically used L-Kb cells, in DC2.4 Kb-SIINFEKL complexes rose rapidly after infection and began to plateau by 6 hours post infection (hpi). To test if the kinetics observed for Kb-SIINFEKL complexes is representative of all VACV epitopes we used polyclonal T cells isolated from infected mice since monoclonal antibodies to VACV epitope-MHC complexes are not available. If all VACV antigen presentation is like Kb-SIINFEKL, the fraction of polyclonal anti-VACV CD8+ T cells that can be stimulated by infected cells should rise over time with a simple, rapid kinetic. If on the other hand, new pMHC complexes first appear on the cell surface at different times after infection, then one might expect a more complicated curve as new populations of T cells are able to be activated once their epitope appears at the cell surface. Thus using DC2.4 and the same infection protocol, global VACV epitope presentation was probed up to 12 hpi using splenocytes taken from mice seven days after VACV infection and the percent of CD8+ T cells making IFNγ determined by intracellular cytokine staining (ICS) (Figure 1B). In contrast to the simple rise of Kb-SIINFEKL presentation, the increase in number of CD8+ T cells recognising the infected cells was more complex. There were two phases of rising CD8+ T cell activation, one from 2 to 5 hours (a similar time frame to Kb-SIINFEKL presentation) followed by second, steeper rise from 5–7 hpi that continued until 12 hpi. While this reveals nothing about the kinetics of individual epitopes, it suggests that the onset of presentation differs across the native VACV epitopes. It is also consistent with published work using mono-specific T cell lines that shows presentation of some VACV epitopes is delayed for some hours after infection [35]. Together these data suggest that monitoring a single epitope does not reveal the true complexity of viral antigen presentation to T cells. We therefore sought to dissect in greater detail the presentation of individual VACV derived epitopes using mass spectrometry (MS).
Liquid chromatography coupled to multiple reaction monitoring mass spectrometry (LC-MRM) is the method of choice for detection of multiple known peptides [32], [36], [37]. LC-MRM MS affords high sensitivity and selectivity and has been recently applied to multiplexed qualitative and quantitative analyses of peptide epitopes eluted from MHC molecules [32], [37]. For this study, eight VACV epitopes restricted by murine H-2 Kb were chosen based on their well characterised immunogenicity and their expression from a variety of different VACV proteins spanning different temporal phases of the infection (Table 1) [4], [5]. In addition, SIINFEKL was included in some experiments to allow a direct comparison of this model antigen with the native VACV epitopes. Optimal MRM transition conditions (precursor ion charge, fragmentation energy and fragment ion selection) for each VACV epitope listed in Table 1 were determined using synthetic peptides (Table S1 and Figure S1 in Supporting Information). The resulting MRM method allowed for the simultaneous detection of all 8 VACV epitopes (Figure 2A) and also included transitions to measure SIINFEKL and isotopically-labelled (AQUA) SIIN*FEKL; inclusion of the SIIN*FEKL AQUA peptide was used to control for losses during processing of the MHC-bound peptides as described [32]. The unequivocal detection of peptide epitopes was achieved by several rigorous confirmatory steps in this LC-MRM workflow: firstly, RP-HPLC retention across multiple dimensions of purification (correct eluting fraction during off-line RP-HPLC and correct on-line retention time during LC-MRM MS) must be consistent with that measured for the synthetic version of each of the VACV peptides (Figure S2); secondly, they must trigger all MRM transitions concurrently and in the correct transition hierarchy; and, as a final step, each peptide sequence must be further confirmed by an MRM-triggered MS/MS sequencing scan – a modality unique to the quadrupole linear ion trap mass spectrometer used in this study [38].
In order to verify the sample workflow (Figure 2B), DC2.4 cells were incubated with a pooled mixture of the full set of 8 synthetic peptides representing VACV epitopes (Table 1). Following extensive washing to remove unbound peptides, cells were pelleted and snap-frozen and subjected to immunoaffinity purification of H-2Kb complexes, peptide elution and chromatographic separation as previously described [32], [37]. The presence of each VACV epitope in the MHC eluate was confirmed by LC-MRM (Figure 2C). The differing detection intensities across the peptide set reflects a combination of the varying ionisation efficiencies of the peptides and competition for binding to the Kb molecules during incubation.
Next, MHC elution and LC-MRM were used for the detection of SIINFEKL and native VACV epitopes generated through VACV infection with the recombinant WR-NP-S-GFP. DC2.4 cells (1×108) were infected for 6 hours with WR-NP-S-GFP to compare the levels of SIINFEKL presentation with that of the 8 native VACV epitopes (Figure 3). Capture of Kb-peptide complexes was achieved as above, including the addition of 50 fmol of isotopically-labelled AQUA SIIN*FEKL in order to control for sample preparation losses post affinity purification of the MHC-peptide complexes [32]. The quantification of each VACV epitope was achieved by comparing the area under the MRM curve to that of 100 fmol of the corresponding synthetic epitope analysed separately (Figure 2A). LC-MRM confirmed the detection of SIINFEKL and all 8 VACV peptides (Figure 3A shows representative data for SIINFEKL, B820–27 and J3289–296). Further it provides the first definitive evidence that the amino acid length and constitution of the VACV epitopes is exactly as described in the original mapping studies [4], [5]. SIINFEKL presentation on WR-NP-S-GFP-infected cells at 6 hpi was calculated to be 2.3×104 and 3.1×104 copies per cell for two independent experiments (Figure 3B). All 8 Kb-restricted VACV epitopes were detected at considerably lower estimated abundances to that of SIINFEKL. Further, abundance of the 8 VACV peptides varied over a wide range with 3 epitopes (B820–27; A47138–146 and J3289–296) being presented at levels up to 1000-fold higher than the remaining 5 VACV epitopes. When compared to CD8+ T cell response elicited in mice infected for 7 days by the same virus, there is a striking dissociation between the epitope abundance and T cell immunodominance hierarchies (Figure 3B).
Next we sought to assess the presentation kinetics of the 8 VACV epitopes during the course of infection. This was done using non-recombinant VACV, to avoid any potential competing effects from the very high levels of presentation of SIINFEKL following infection with the recombinant WR-NP-S-GFP VACV strain. DC2.4 cells were infected for 0.5, 3.5, 6.5, 9.5 and 12.5 hours, or mock infected as a negative control and epitope abundance at each time determined by LC-MRM analysis. All 8 VACV epitopes were detected and the kinetics of their presentation measured (Figure 4A). Six of 8 peptides were detected by 0.5 hpi, with the remaining 2 epitopes (A3270–277 and A1947–55) undetectable until 6 hours later. Peak expression occurred at 3.5 hpi for 5 epitopes, 6.5 hpi for two epitopes and at the final time point of 12.5 hours for a single epitope. We noted that the presentation of the immunodominant B820–27 epitope was unusual in that its onset was at 30 minutes, but instead of peaking at 3.5 hpi, like most of this group of epitopes, its peak was later at 6.5 hpi. The abundance profile spanned 3 logs, ranging from as low as an estimated 11 copies per cell for C4125–132 to as high as 32,400 copies of A47138. These basic features of presentation with some epitopes showing peak presentation around 3.5 hours after infection, while others only appear at 6.5 hours have also been observed for cells infected with the MVA strain of VACV (our unpublished observations). Thus abundance and kinetics of presentation are highly variable across different epitopes and robust presentation early after infection is not always maintained.
In order to assess how the kinetics of epitope presentation correlates with source antigen expression, a sample of the cell lysate from each infection time point was subjected to reduction, alkylation and subsequent digestion with the enzyme trypsin prior to proteomic analysis. Proteotypic tryptic fragments from each of the 8 VACV protein antigens were chosen using Skyline [39] (Table S2 and Figures S3 and S4). Following initial screening of samples, 6 of the 8 VACV proteins were detected (for A3 and J3, multiple tryptic fragments were found to be amenable to MRM analysis and so all were included). Despite rigorous testing of multiple peptides, no positive signal could be detected for proteins L2 and C4 so these were not included further. In order to achieve normalisation of protein loading across the timecourse, 12 murine tryptic peptides (corresponding to eight host proteins; Table S3 and Figure S3) were simultaneously analysed in the same LC-MRM method (Figure S3). These murine proteins were chosen as suitable candidates for normalisation based on the high copy number and long half life of their human homologues [40], with the notion that such proteins will not be grossly affected by the VACV-mediated shutdown of host protein synthesis. In addition, a good correlation between the abundance of these representative proteins and cell number recovered post-infection was found suggesting that they were appropriate for normalisation (Figure 3C). The uncorrected data is also shown in Figure S4 for comparison.
MRM peaks at each time point for the 6 VACV proteins were used to determine relative protein expression over the course of infection and these were plotted alongside the relative levels of each epitope derived from the same protein (Figure 4B). This approach allows relative expression of individual antigens to be determined at different time points but does not provide absolute quantitation of the antigen and therefore direct comparison between antigens is more qualitative. Expression profiles of the 6 proteins were consistent with their temporal expression cluster as reported by analyses of transcription and more recently defined promoters [41]–[43], which gives further confidence of the method. Translation, as determined by tryptic peptide detection, was detected at 0.5 hpi for A47, A8 and B8, corresponding with the appearance of epitopes derived from those proteins. Whilst levels of A47 peaked at 6.5 hpi, all other proteins peaked (at least within the limits of this time course) at 12.5 hours. Proteins A3 and A19, both of which are classified as late, were detected by 3.5 hours, but did not reach substantial levels until 6.5 hours and onwards; presentation of epitopes A3270–277 and A1947–55 tracked closely with the increase in protein levels. For epitopes A47138–145, A8189–196 and J3125–132, rapid and peak presentation following protein expression was followed by a sharp decline in epitope levels to almost zero by 12.5 hpi. However, epitopes B820–27 and A3270–277, although decreasing following peak levels mid-infection, maintained a more constant level around 20–40% of the maximum; for A1947–55, epitopes levels did not peak until the end of the time course, following an almost identical profile to A19 protein expression. Of note the B820–27 epitope appeared to display a lag between peak of protein expression and peak of epitope presentation.
Next, in vitro protein and epitope presentation kinetics were correlated with CD8+ T cell immunodominance in vivo. C57BL/6 mice were infected with VACV WR by the intraperitoneal route (i.p.) and 7 days after infection, the percentage of CD8+ T cells responding to ex vivo stimulation with each peptide was determined by intracellular staining for IFNγ (Figure 4C). This method of epitope detection has recently been shown to have a linear range that covers responses to all the epitopes investigated here [44]. As previously reported [4], [5], B820–27 dominated the response, A1947–55 was the weakest and the remaining 6 epitopes formed an intermediate hierarchy. Here, where the onset, peak level and longevity of epitope display were revealed (as opposed to the single time point for the WR-NP-S-GFP in Figure 3), there was still no obvious correlation between presentation and the CD8+ T cell dominance hierarchy. Although the immunodominant B820–27 was one of the most robust epitopes in peak and persistence of presentation, it is similar in this respect to the subdominant A47138–146 and J3289–296. Further, A3270–227 and A8189–196, which are the next 2 peptides in the dominance hierarchy after B820–27, have very different presentation profiles with the former only appearing later (6.5 hpi) and having better persistence but a substantially lower (approximately 10-fold) peak than the latter.
The use of liquid chromatography and mass spectrometry to detect MHC epitopes has a long heritage [e.g. [45]–[50]] yet it is only in recent years that techniques and instrumentation are beginning to surpass sensitivity and feasibility blockades to gain qualitative and quantitative insights into the immunopeptidome [24], [32], [51]–[53]. Use of LC-MRM methods to detect epitope presentation offers a large increase in sensitivity, but thus far has few precedents in the literature. LC-MRM analysis has rarely been used to examine antigen presentation with only a few examples examining melanoma epitopes [54] and measles virus epitopes [55]. We have recently further developed the methodology studying SIINFEKL presentation as a model antigen [32]. The current study is the first to comprehensively apply LC-MRM to study epitope presentation during virus infection, an inherently dynamic process. It is also the first to combine epitope and source antigen quantification from the same samples using LC-MRM. Our data provide extensions to and have implication for several aspects of antigen processing and anti-viral immunity and include:
|
10.1371/journal.pcbi.1004305 | The Sense of Confidence during Probabilistic Learning: A Normative Account | Learning in a stochastic environment consists of estimating a model from a limited amount of noisy data, and is therefore inherently uncertain. However, many classical models reduce the learning process to the updating of parameter estimates and neglect the fact that learning is also frequently accompanied by a variable “feeling of knowing” or confidence. The characteristics and the origin of these subjective confidence estimates thus remain largely unknown. Here we investigate whether, during learning, humans not only infer a model of their environment, but also derive an accurate sense of confidence from their inferences. In our experiment, humans estimated the transition probabilities between two visual or auditory stimuli in a changing environment, and reported their mean estimate and their confidence in this report. To formalize the link between both kinds of estimate and assess their accuracy in comparison to a normative reference, we derive the optimal inference strategy for our task. Our results indicate that subjects accurately track the likelihood that their inferences are correct. Learning and estimating confidence in what has been learned appear to be two intimately related abilities, suggesting that they arise from a single inference process. We show that human performance matches several properties of the optimal probabilistic inference. In particular, subjective confidence is impacted by environmental uncertainty, both at the first level (uncertainty in stimulus occurrence given the inferred stochastic characteristics) and at the second level (uncertainty due to unexpected changes in these stochastic characteristics). Confidence also increases appropriately with the number of observations within stable periods. Our results support the idea that humans possess a quantitative sense of confidence in their inferences about abstract non-sensory parameters of the environment. This ability cannot be reduced to simple heuristics, it seems instead a core property of the learning process.
| Learning is often accompanied by a “feeling of knowing”, a growing sense of confidence in having acquired the relevant information. Here, we formalize this introspective ability, and we evaluate its accuracy and its flexibility in the face of environmental changes that impose a revision of one’s mental model. We evaluate the hypothesis that the brain acts as a statistician that accurately tracks not only the most likely state of the environment, but also the uncertainty associated with its own inferences. We show that subjective confidence ratings varied across successive observations in tight parallel with a mathematical model of an ideal observer performing the optimal inference. Our results suggest that, during learning, the brain constantly keeps track of its own uncertainty, and that subjective confidence may derive from the learning process itself. Our results therefore suggest that subjective confidence, although currently under-explored, could provide key data to better understand learning.
| Many animals, human adults and even human babies possess remarkable skills to cope with the pervasive uncertainty in their environment [1,2]. Learning processes are attuned to uncertainty. They enable one to capture the stochastic characteristics of the environment, as when one learns how often a probabilistic cue leads to a reward [3]. The environmental uncertainty actually occurs at several nested levels, as the stochastic characteristics themselves may also vary suddenly and without warning. The human learning is sophisticated enough to quickly adapt to such higher-order changes: the probabilities and characteristics that subjects learn are adequately fitted by statistical models [4–7]. However, in such tasks and environments flooded with uncertainty, subjects not only estimate the characteristics of the outside world, they also evaluate the degree of certainty that their estimates are accurate. This more subjective aspect of learning, the “feeling-of-knowing”, has received little attention so far. Here, we attempt to provide a formal account of this feeling and its origin.
The feeling-of-knowing, or the sense of confidence, has been primarily demonstrated in memorization tasks [8] and in perceptual decision-making tasks in humans, monkeys and rodents [9–11]. By contrast, evidence from probabilistic learning tasks is currently limited. Many learning models actually simply do not consider feeling-of-knowing as a component of the learning process. Most share a common logic, according to which each parameter of the environment is represented at any given moment by a single numerical estimate and is continuously updated based on new observations. Rescorla and Wagner suggested a simple update rule: the point estimate should be shifted in proportion of the prediction error, i.e. the extent to which the estimate deviates from the new observation [12]. Such models therefore only provide point estimates, and they are devoid of any sense of uncertainty. It has been recognized more recently that the learning rate could actually be modulated as a function of an internal estimate of the environmental uncertainty, e.g. volatility [4,6] and that learning could even be fully reset when an environmental change is detected [7]. However, the normative Bayesian approach of learning suggests that there is a principled distinction between this environmental uncertainty and the uncertainty in the internal knowledge of what has been learned [13]. We term this second kind of uncertainty, the 'inferential uncertainty'. Despite evidence that the inferential uncertainty could affect learning in humans [5], how humans perceive this uncertainty remains largely unexplored. Here, we suggest that the feeling-of-knowing, or subjective confidence, corresponds formally to the inferential uncertainty and that it derives from the inference that underpins the learning process itself.
Indeed, the fact that humans have distinct degrees in their feeling-of-knowing suggests that they do not keep track of point estimates of environmental parameters, but instead of a set of estimates, each with its own degree of plausibility. Supporting this idea, some models assume that the brain infers full probability distributions [14]. The hypothesis was initially introduced for sensory representations, but it may be extended to higher-level tasks [15–17], possibly including the learning of any numerical parameter. Following this hypothesis, learning in an uncertain world would be underpinned by a probabilistic inference that provides, not a single parameter value, but a distribution of possible values—and therefore affords an estimation of “feeling-of-knowing” based on the concentration of this inferred distribution onto a single value [18,19].
To test this idea, we examined whether humans can provide not only accurate estimates of environmental probabilities, but also accurate confidence ratings in those estimates. Such a finding would imply that the brain not only computes a point estimate, but also, at a minimum, the uncertainty in inferring its value, and perhaps even its full distribution. We designed a challenging probabilistic learning task with two nested levels of environmental uncertainty. Fig 1 shows how we generated the random sequences of visual or auditory stimuli and Fig 2A shows an example session. First, at any given moment, the sequence depends on two parameters: P(A|B) and P(B|A), i.e. the transition probabilities between stimuli A and B. Second, these transition probabilities themselves remain stable only for a limited time, then change abruptly to a new random value, thus delineating ‘chunks’ in the sequence separated by ‘jumps’. These jumps were aimed at inducing fluctuations in the inferential uncertainty over time. Subjects were asked to detect the jumps and, occasionally, to report their estimate of the transition probability to the next stimulus and their confidence in this estimate.
Subjective estimates of transition probabilities can be compared to the true generative probabilities. However, this comparison is not completely fair because the generative parameters are not available directly to the subject, but can only be inferred from the specific stimuli received. Furthermore, confidence is simply not a characteristic of the generative process, but solely of the inference process. This highlights the need to derive both the estimates of transition probabilities and confidence levels in a principled manner from the inference itself. We therefore compare subjects' answers with the inference generated by an Ideal Observer endowed with the mathematically optimal inference process. This normative solution formalizes the link between the inference on the one hand, and the probability estimates and confidence levels on the other. Indeed, the optimal inference returns a distribution of likelihood over the transition probabilities, given the specific stimuli received. Both a point estimate and a confidence level in this estimate can be derived from this distribution. The distribution can be averaged to obtain a single best estimate of the transition probability. Confidence should reflect how precise this estimate is: whether the distribution is spread (low confidence) or concentrated (high confidence) around this estimate. We thus formalized confidence as the precision of the distribution (its inverse variance), as previously suggested [19].
The Ideal Observer being normative, it provides a reference to assess the accuracy of the single point estimates and the fluctuations in confidence levels reported by subjects. In addition, since the Ideal Observer formalizes how single point estimates and confidence levels should derive from the inference process, it affords a series of predictions serving as tests of whether the reported estimates and confidence levels indeed derive from a common inference. And last, if confidence levels derive from an accurate inference, then they should reveal several specific properties of this efficient inference system.
We first asked whether subjects could detect when the characteristics of the sequence changed suddenly. We assessed the accuracy of their detection in comparison to the actual position of jumps with a Receiver Operative Characteristic analysis. Subjects reported more jumps when transition probabilities were indeed changing (hit) than when they were stable (false alarm): the difference of hit minus false alarm rates was 0.23 (standard error = ± 0.03; t-test against 0: p<10–5). To show that this difference is positive not because of chance, but instead because the detection is based on the actual evidence provided by the observed sequence, we used a more conservative test. Comparison with surrogate data indicates that the observed difference between the hit and false alarm rates is significantly higher than expected from a random detection process (p<0.01, see Methods).
Although the detection of jumps by subjects is better than chance, it is not perfect: some jumps were missed, and some others falsely reported. However, some of these errors precisely further demonstrate that subjects based their detection on the actual level of evidence received. Indeed, in principle, not all jumps can be detected equally easily: for instance, when changes in transition probabilities are small and frequent, the sequence may not provide enough evidence for the presence of each jump. The Ideal Observer provides a principled way of quantifying the likelihood of a jump at each position in the observed sequence. We tested whether subjects are sensitive to such fluctuations in evidence by analyzing their errors (misses and false alarms) from the Ideal Observer perspective. A significant difference in jump likelihood at the time of the subjects' Hits vs. Misses (p = 0.003) indicates that subjects were more likely to miss a jump when the apparent jump likelihood was misleadingly low. Similarly, a difference between False Alarm vs. Correct Rejection (p = 0.001) reveals that the subjects' false alarms were more likely to occur when jump likelihood was high (see Fig 3).
Altogether, these results indicate that subjects partially managed to track the jumps in the objective generative process, and their responses give evidence of an efficient statistical use of the available information.
We next examined whether subjects could estimate the characteristics of the sequence despite their unpredictable changes in time. The sequence was paused every 12 to 18 stimuli and participants were asked to report the probability that the next stimulus would be A or B. Subjects’ responses were correlated across trials with the true generative probabilities (t17 = 8.8, p<10–7), indicating that subjects' probability estimates, although imperfect, consistently followed the generative probabilities. The deviations could reflect that the transition probabilities are inferred from the specific and limited amount of stimuli received. We therefore compared the subjects’ estimates of transition probabilities with the optimal values that could be inferred from the data, i.e. the parameter estimates inferred by the Ideal Observer. The subjects' responses were tightly correlated with the optimally inferred probabilities (t17 = 8.5, p<10–6, see Fig 4A). When both predictors were included in a multiple linear regression, significantly higher regression weights were found for the optimal estimates than for the generative values (paired difference of weights: t17 = 4.6, p<10–3).
Given that the Ideal Observer and the subjects are both asked to estimate a probability, we can not only test whether their estimates are correlated, but also whether they are identical. Fig 4A reveals a remarkable match, although somehow imperfect: the observed slope is actually slightly below the identity. This deviation could reflect the distortion of subjective probabilities classically reported [20]. However, this pattern could also reflect differences in accuracy across trials and subjects. Indeed, the average of ideal estimates should be perfectly aligned on the diagonal, but the average of random estimates would form a flat line at 0.5; therefore a mixture of both should result in an intermediate slope. Supporting this view, the inspection of individual data revealed that the regression slopes were significantly larger than 0 in most subjects (p>0.009 for 16 out of 18 subjects), but they were significantly equal to 1 in only 3 subjects (in these subjects, Bayes factor > 9, see [21] for the computation of this 'Bayesian t-test').
Together, these results show that subjects were able to infer the transition probabilities generating the observed sequence of stimuli despite their sudden changes in time. Not surprisingly, subjects were outperformed by the Ideal Observer endowed with the best inference scheme. However, the comparison to the optimum reveals a remarkable accuracy of the subjects' estimates.
Subjects were also asked to rate their confidence in their probability estimates. They provided confidence ratings on a bounded qualitative continuum (see Fig 1). The absolute position of a given 'feeling-of-knowing' on this continuum is a matter of subjective representation, not a property of the inference process. However, if the confidence judgment reflects the certainty of the inferred probability estimate, then distinct confidence ratings should correspond systematically to distinct levels of evidence. Therefore, we assessed the accuracy of the fluctuations in confidence judgment with a regression against a principled measure of the level of evidence. Again, we used the Ideal Observer to this end. Intuitively, confidence should be high if and only if the estimated distribution of transition probability is concentrated on the reported value. This corresponds formally to the notion of precision, the inverse variance of the estimated distribution. Thus, we defined the Ideal Observer confidence as the negative log of the variance of the distribution. We used the log scale because it is the natural space for variance [22]. Note that the log variance and log standard deviation are strictly proportional, therefore the choice of one or the other provides the exact same significance levels in the regression analyses. We found a strong positive correlation between this principled measure of confidence and the subjective confidence (t17 = 3.94, p = 0.001; see Fig 4B).
In addition, given that the experiment presented visual and auditory stimuli in separate blocks, we checked the robustness of the previous results by testing each modality separately. The regression of subjective estimates against the Ideal Observer was significant within each modality (for probability estimates: both p<10–5; for confidence: both p<0.004). Interestingly, these regression weights were positively correlated between modalities (for probability estimates, Pearson ρ16: 0.55, p = 0.017, for confidence ρ16: 0.59, p = 0.010), supporting the idea that inferential capabilities vary between observers and are not tied to one modality but instead characterize a supramodal level of processing.
Another way to evaluate the accuracy of confidence judgments is to ask whether they predict performance. Computing confidence can be useful when it serves as a proxy for the accuracy of performance—high subjective confidence should predict an objectively low rate of errors. We verified that this is true for the normative Ideal Observer: across trials, the magnitude of the error separating the Ideal Observer estimates from the true generative probabilities was negatively correlated with the Ideal Observer confidence (t17 = -8.67, p<10–6). Crucially, a similar relationship linked the subjects’ confidence with the objective error in their probability estimates (t17 = -2.27 p = 0.037). We used simulations to check that this link derives from the normative nature of the subjects' estimates and not from biases in their probability estimates or confidence ratings. We used three separate simulations to reassign randomly one variable (probability estimates, confidence ratings or true generative probabilities), while keeping the two others unaffected. Each simulation disrupts specific links between the generative probabilities and the subjective estimates to capture potential response biases. The simulations showed that the negative relationship observed between confidence and objective error is unlikely to emerge by chance on the sole basis of response biases (all p<0.019).
Altogether, the findings indicate that confidence estimation is accurate: it relates linearly to the principled inference made by the Ideal Observer, and it is also correlated with objective performance.
Our hypothesis is that estimates of transition probabilities and confidence ratings jointly derive from a single inference process. In other words, there is a common substrate for both estimates. An alternative hypothesis would be that confidence ratings are derived from the estimates of transition probabilities. Our hypothesis leads to several testable predictions that also rule out the alternative.
We predict that probability estimates and confidence ratings should be partly related: when information is scarce, the optimal default estimate for transition probability is around 0.5 and confidence is low. Extreme estimates (toward 0 or 1) are achieved only when there is substantial evidence and hence when confidence is high. Confidence should thus increase when the probability estimates depart from 0.5: this is a fundamental and inescapable property of probabilistic reasoning. The Ideal Observer estimates robustly showed this U-shape pattern (quadratic weight: t17 = 16.1, p<10–11), and so did our subjects (t17 = 9.77, p<10–7). This effect was actually significant within every subject (all p<0.0025). However, we also predict that this U-shape relationship should be only partial, since in principle, one may be more or less confident in any probability estimate, depending on the number of observations that support it. To illustrate this property, we binned the participants' confidence ratings by their subjective probability estimates, and within each bin, we then sorted trials by high and low Ideal Observer confidence with a median split (Fig 5A). Subjective confidence reflected the Ideal Observer confidence on top of the general U-shape pattern. To quantify this additional effect, we performed a multiple regression of subjective confidence, without binning, including as predictors both the subject's U-shape transformed probability estimates and the optimal confidence. The data revealed that the Ideal Observer confidence indeed captures aspects of subjective confidence (t17 = 3.12, p = 0.006) that are not accounted for solely by a quadratic effect of probability estimates.
In our experiment, subjects reported their probability and confidence estimates sequentially. We therefore ran a control experiment to check that the nested relationship between probability estimates and confidence ratings (as shown in Fig 5A) is a general property of human reasoning which cannot be attributed to sequential reporting. New subjects performed a variant of the task in which the first question about the probability estimate was omitted (see Methods). Subjective confidence rating still followed the Ideal Observer confidence (t20 = 7.00, p<10–5). As expected from a probabilistic inference, subjective confidence also showed a quadratic effect of the optimally inferred probability (t20 = 6.97, p<10–5). In addition, subjective confidence still co-varied with the Ideal Observer confidence on top of the quadratic effect of the optimal probability (multiple regression: t20 = 3.01, p = 0.007, t20 = 5.13, p<10–5 respectively).
Our main experiment enables to further test the predictions of our hypothesis concerning the common origin of probability and confidence judgments. If probability estimates and confidence ratings both derive from the same inference, then we also expect that subjects who perform the inference accurately should perform accurately in both estimating probabilities and rating confidence. We defined how accurate subjects were in estimating probabilities and rating confidence with respect to the Ideal Observer. In both cases, accuracy was summarized as the correlation coefficient between the subjects’ response and the optimal response. We found a positive correlation across subjects between the accuracies of probability estimates and confidence ratings (Pearson ρ16 = 0.67, p = 0.002, Fig 5B). We also tested whether this correlation was significant within subjects. On each trial, we computed the accuracy of probability estimates (or confidence ratings) as the distance between the Ideal Observer and the subject's responses (see Methods). Again, the two accuracies were significantly correlated across trials (t17 = 3.27, p = 0.005). Note that these correlations are also consistent with the alternative hypothesis that confidence ratings are derived from probability estimates. However, the within-subject data disprove this alternative. Indeed, we controlled that the within-subject correlation we found is not confounded by an effect of an estimation-to-confidence mapping (be it quadratic or not) by comparison with two shuffled data sets (see Methods and Fig 5C).
Altogether, these results show that probability estimates and confidence ratings are likely to derive from a common inference. In particular, the accurate confidence ratings reflect additional features of the inference that are not reflected in the probability estimates.
We now examine whether the data provide cues as to how confidence is computed. The inference should use the incoming data to constantly update an internal model of the hidden process that could have generated the observed sequence of stimuli. There are normative principles ruling this update process. Therefore, any efficient algorithm should have specific characteristics. We show that confidence ratings reveal three properties expected from an efficient information processing system.
First, whenever the probability estimates change a lot, indicating a severe revision of the internal model (for instance, after a jump), then confidence should be low; conversely, when estimates are stable, confidence in the seemingly 'good' value should be high. Questions being asked only occasionally to the subjects, the subjective model revision cannot be estimated from their reports. Instead, we estimated the degree of model revision from the Ideal Observer. This ensures in addition that subjective confidence is regressed against a normative estimate in every subject. We observed the predicted negative correlation between subjective confidence and the amount of revision in the probability estimates relative to the previous observation (t17 = 3.67, p = 0.002; Fig 6A), indicating that subjective confidence tracks the revision of an internal model.
Second, the number of data samples accumulated since the last detected jump should affect the level of confidence: more samples should lead to more precise estimations. We counted the cumulative number of samples between the optimally detected jumps. As predicted, we observed a positive correlation between subjective confidence and the number of samples since the last jump (t17 = 3.51, p = 0.003; Fig 6B), indicating that subjective confidence increases with the accumulation of evidence. Again, using the Ideal Observer to estimate the number of samples in the current chunk provides a normative comparison across subjects. Instead, using the subjects' jump detection entangles several factors, e.g. whether subjects are accurate and conservative in reporting jumps. The same analysis based on the subjects' jump detection however also revealed a positive correlation (t17 = 2.26, p = 0.037).
Third, confidence should be lower when the estimation of the model is made more difficult by decreasing the predictability of the sequence. Formally, the unpredictability of a sequence is characterized within a chunk by the entropy of the generative transition probabilities: it is maximal when the transition probability is 0.5 and it decreases as the transition probability goes toward 0 or 1. Note that we quantify here the generative environmental uncertainty, not its subjective estimate (as in Fig 5A). We therefore examined if confidence was negatively correlated with this entropy. As predicted, a negative correlation was observed (t17 = -5.58, p<10–4; Fig 6C). As a control, we examined if a similar effect occurred when computing the entropy of the other, currently irrelevant transition probability (transition from the stimulus which was not presented on the previous trial). No significant effect was found (t17 = -0.76, p = 0.50; Fig 6D). The results therefore indicate that subjects keep a distinct record of the confidence attached to each of the two transition probabilities that they are asked to estimate.
We checked that the results presented in Fig 6 survive correction for multiple comparisons and partial correlations by including the four regressors into a multiple regression of confidence levels. The three factors of interest were still significant (amount of model revision needed: p = 0.006; number of samples received: p = 0.042; entropy of the relevant transition: p = 10–5, and not the irrelevant one: p = 0.3). We also confirmed that these results coincide with the normative theory by running the same analysis on the Ideal Observer confidence (effect of the 3 factors of interest: |t17|>8.7, p<10–7, no effect of the irrelevant transition entropy: t17 = -0.3, p>0.7). These results support the idea that confidence ratings derive from a rational process that approximates the optimal probabilistic inference.
Our hypothesis is that confidence and probability estimates both derive from the probabilistic inference itself. An alternative is that subjective confidence is derived independently with a valid heuristic [23,24]. In the current experiment, the probability estimate for instance is a rational cue for confidence: as discussed with Fig 5A, there is a strong and principled correlation between confidence and how much the probability estimate departs from 0.5. However, we showed (in Fig 5A and 5C) that the accuracy of confidence judgment goes beyond this kind of mapping. This therefore precludes that subjective confidence derives only from a heuristic based on the probability estimate. An example of such heuristic would be to count the number of correctly predicted stimuli in the immediately preceding trials to determine a confidence level.
We then showed that confidence is also systematically impacted by the entropy of the generative transition probability, the amount of samples accumulated in the current chunk and the degree of revision of the probability estimates (Fig 6). At a minimum, these results imply that confidence arises from a sophisticated heuristic that combines the above factors. However, we can prove here that human confidence ratings are more accurate than such a heuristic would predict: even after regressing out the effect of the above factors, the residual subjective confidence still co-varied significantly with the Ideal Observer confidence (t17 = 2.89, p = 0.01, Fig 6E).
What additional features of the inference process could explain this finding? In deriving the “number of samples” heuristic, we assumed that subjects discretize the incoming sequence into discrete chunks separated by jumps, and that this process allows them to track how much evidence they received since the last jump. This heuristic is suboptimal, however: the optimal inference avoids any discrete decision, but computes with the full probability distribution that a jump occurred at any moment, and uses it to weight recent evidence. To evaluate whether human subjects integrate jump likelihood into their confidence estimates, we computed, on each trial, the current uncertainty on the location of the last jump. We quantified it as the variance of the current chunk length estimated by the Ideal Observer, normalized by its mean value (over similar positions in the sequence across sessions and subjects) so that values higher than 1 indicated that it was less clear than average when the last jump occurred. Subjective confidence correlated negatively with this uncertainty on jump location (t17 = -3.12, p = 0.006), exactly as expected from a normative viewpoint (same analysis with Ideal Observer instead of subjective confidence: t17 = -5.71, p<10–4). It therefore seems that subjects are able to factor an estimate of jump probability in their confidence judgments. Altogether, these results suggest that the inference underpinning learning in this task is a probabilistic computation.
We present an in-depth analysis of how humans acquire explicit knowledge and meta-knowledge of transition probabilities in an unstable environment. Our results demonstrate that subjects use the available stochastic evidence to learn about the incoming sequence: their estimates of two transition probabilities P(A|B) and P(B|A) accurately track the true generative values. Most importantly, by asking subjects to systematically rate their confidence in those estimates, we show that humans can accurately evaluate the uncertainties associated with each piece of information that they acquire. This sense of confidence, which affords a quantitative and explicit report, is available in a modality independent manner for both visual and auditory sequences, and it closely tracks the fluctuations in uncertainty that characterize an accurate probabilistic inference process.
Several classifications of uncertainties have been proposed [25]. Our distinction between environmental and inferential uncertainties is close to Kahneman and Tsersky’s [23] classical division of external uncertainty (stochastic nature of the environment) versus internal uncertainty (state of knowledge). A similar distinction is also made in recent computational works, e.g. in [13], the environmental uncertainty would correspond to the 'risk' and 'unexpected uncertainty', the inferential uncertainty to 'estimation uncertainty'; in [5] a similar distinction is made. Internal uncertainty is sometimes called ambiguity, in particular in economics, when it characterizes the absence of knowledge [25,26]. Our terminology (environmental vs. inferential uncertainties), stresses that these two kinds of uncertainties differ in their epistemic nature. By operationalizing this distinction, our study revealed how they are only partially related. We built upon previous paradigms that manipulated environmental uncertainty [4,7] in order to induce frequent variations in inferential uncertainty. We showed how a first-order environmental uncertainty (probabilistic transitions between stimuli) increases the inferential uncertainty, and how a second-order environmental uncertainty (unexpected changes in these transition probabilities) produces additional fluctuations in inferential uncertainty over time. The fact that environmental and inferential uncertainties are only partly related is particularly salient in our task when a transition probability is 0.5. Such probability produces the least predictable outcomes (high environment uncertainty) and a precise estimation of this probability needs more samples than any other probabilities (hence, a high inferential uncertainty). However, with a large number of observations, one can get quite confident that the outcomes are indeed completely unpredictable. All these effects were observed in a normative Ideal Observer model, and subjects' confidence faithfully tracked ideal-observed confidence. Thus, human adults possess sophisticated mechanisms for tracking their inferential uncertainty.
Juslin & Olson [27] made a different distinction, separating Brunswikian uncertainty, independent from us and in that sense 'external', and Thurstonian uncertainty, due to the imprecision of our information-processing systems. While Thurstonian uncertainty may have contributed to the small deviations that we observed between subjective confidence and the optimal observer, we stress here that learners are uncertain, not only because they are faulty, but primarily because inference from stochastic inputs is by essence uncertain. The Ideal Observer quantifies this irreducible level of inferential uncertainty that any learner must face in our task. It is an open question whether and how humans may combine this core inferential uncertainty with the additional uncertainty arising from their cognitive limitations.
Broadly defined, confidence indexes a degree of belief in a particular prediction, estimation or inference [19,23,25]. What confidence is about may thus vary drastically, from mere detection (feeling of visibility, e.g. [28]), to accuracy in perceptual tasks [9,10,29], in memory retrieval [8], or in response to general-knowledge questions [30,31]. Mathematical concepts clarify how the present work differs from these previous studies. In most studies, confidence can be formalized as the likelihood of some binary variable e.g. the posterior probability that a response is correct/incorrect, a stimulus is seen/unseen, etc. [9]. By contrast, here we investigated confidence in a continuous numerical quantity (the inferred transition probability), so that a principled and natural formalization for the strength of evidence is, as suggested previously [19], the precision of this variable (its inverse variance). This computational distinction, in comparison with most previous studies, entails a noticeable difference in practice. In typical binary decision tasks, the accuracy of subjective confidence is estimated by comparison with the actual performance of the subject. This estimation may be more or less susceptible to biases [32]. In our task, confidence is defined as the precision of the variable inferred, and is therefore amenable to a principled quantification with the Ideal Observer. Therefore here, the accuracy of subjective confidence can be estimated by comparison with this optimal confidence. Crucially, this estimation is independent from the performance in the primary estimation task, which may even remain unknown to the experimenter.
One could disagree with our particular formalization of confidence, and suggest alternative mathematical quantities such as the inverse variance (not its log, as we did), or the posterior probability of the mean or of the maximum of the inferred distribution, or the entropy of this posterior distribution. All these metrics roughly quantify the same notion: they are highly correlated with the one we used, and running the analyses with these other metrics led to similar (although less significant) results. The tight correlation between the ideal-observer precision and human subjective confidence therefore strongly suggests that humans possess a remarkable capacity to extract and use probabilistic information.
We assessed the accuracy of the subjective precision estimates based on their relative variations between trials. The metacognition literature however makes a classical distinction between whether the accuracy of confidence is only relative or also absolute [31]. Absolute confidence levels, and thus the identity between the subjective and the optimal levels, cannot be investigated in our design: indeed, mapping confidence onto a qualitative scale is subjective, not principled. Subjects may produce absolute confidence measures for binary variables, e.g. they may estimate the fraction of correct or seen trials, but asking them a numeric estimate of subjective precision seemed too difficult, which is why we resorted to a qualitative confidence scale. This aspect of our study leaves open the question of whether there is an internal scale for precision that could be sufficiently calibrated to be transferred between tasks [33] or even individuals [34], as previously shown for binary judgments.
Our estimation of the accuracy of subjective confidence relies on a comparison with an Ideal Observer. However, the literature on the perception of probabilities have evidenced frequent deviations from optimality, e.g. the over and under estimation of small and large probabilities [35,36], and a bias toward the detection of alternation vs. repetition [37,38]. Whether adjusting the Ideal Observer to these biases could provide a tighter fit to subjective data is an open issue and a matter for further research. Different options are available to include these biases in the ideal observer model. One possibility is that only the report of the probability is distorted. In that case, the inference, and hence the confidence levels, would remain unaffected. By contrast, the bias could affect a particular component of the inference itself. Potential targets for such distortions include (1) the likelihood of the current observation given some inferred probability estimate, which serves to update the posterior knowledge; (2) the posterior estimate itself, which serves to evaluate the likelihood of future observations; (3) the prior about the generative probabilities, which biases the inference at the beginning of each new sequence, but also at any time a jump in probabilities is suspected. These different potential sources of bias may result in quantitative differences in confidence levels, which could help to arbitrate between these scenarios.
Our results reveal some characteristics of the computation of confidence in humans. One possibility is that second-order estimates occur independently from the first-order estimates, by relying on indirect cues or heuristics such as reaction time in the first-order task [23,24]. However, several aspects of our results contradict this view. First, the sophisticated heuristics we tested did not fully account for confidence reports; similar results were reported in the perceptual domain [39]. Second, the accuracies of the first and the second-order estimates were tightly correlated across trials and subjects which contradicts that confidence levels occur independently.
The alternative view is that first and second-order processes are related, e.g. the second-order process relies on a readout of the same single-trial inferential data available to the first-order process [40–42]. Signal detection theory formalized this readout process in perceptual decisions, postulating that the second-order estimate corresponds to a statistical quantity (d-prime) characterizing the first-order process [32]. Our hypothesis extends this idea to the learning domain: learning could be supported by a probabilistic inference [17,43], resulting in a posterior distribution whose mean and precision would yield, respectively, the first-order and second-order estimates.
The terms first-order and second-order estimates may indeed be unfortunate, as they suggest a sequential process. It is in fact an open issue whether the primary response and the confidence in this response arise in parallel or serially, and from a single brain circuit or not [11,40]. Parallel extraction by distinct circuits could account for the fact that confidence and performance are often correlated, but still dissociable [44,45], for instance in situations of speeded judgment [29], overconfidence [46], or when the accuracy of confidence is impaired while performance is preserved.
By revealing some characteristics of the computation of confidence, our results may reveal some characteristics of the learning process itself. Indeed, if both the learned estimates and the assigned subjective confidence levels derive from the same inference, then investigating subjective confidence could provide critical insights on the learning process. It should be the case if subjective confidence levels reveal something more than what the learned estimates already reveal by themselves. We showed that it is the case: the accuracy of subjective confidence cannot be reduced to the accuracy of the learned estimates. This implies that the classic view of learning, exemplified by the Rescorla Wagner rule, according to which learning simply consists in updating parameter estimates, does not suffice—the brain also keeps track of the uncertainty associated with each value. Recent computational works have already started to revisit this classic learning model so as to incorporate notions of uncertainty [5,13]. Our results emphasize the need to investigate confidence as part of the learning algorithm. Future work should determine whether learning relies on simplified computations involving only summary statistics such as mean and variance [5], on sampling schemes [17,47], or on full computations over distributions [15].
The study was approved by the local Ethics Committee (CPP n°08–021 Ile de France VII) and participants gave their informed written consent prior to participating.
18 participants (9 females, mean age 23, sem: 0.74) were recruited by public advertisement. The task was delivered on a laptop using Matlab (Version R2013a) and PsychToolBox (Version 3.0.11). The experiment was divided into 4 blocks, each presenting a sequence of 380 stimuli (denoted A and B). On alternated blocks, A and B were either auditory or visual stimuli perceived without ambiguity, see Fig 1 for a description and the timing. A fixation dot separated the visual stimuli and remained present during the auditory blocks. The modality used in the first block was counterbalanced over subjects.
The sequence was generated randomly based on predefined transition probabilities between stimuli, e.g. an 80% chance that A is followed by A and a 30% chance that B is followed by A. These values are thus called 'generative transition probabilities'. The sequence was structured into chunks: transition probabilities were constant within chunks and changed from one chunk to the next at so-called 'jumps'. Chunk lengths were sampled from a geometric distribution, with an average chunk length of 75 stimuli. To avoid blocks without jumps, chunks longer than 300 stimuli were discarded. In each chunk, transition probabilities were sampled independently and uniformly in the 0.1–0.9 interval, with the constraint that, for at least one of the two transition probabilities, the change in odd ratio p/(1-p) relatively to the previous chunk should be at least 4. The sequence was paused occasionally (every 15 stimuli, with a jitter of ± 1, 2 or 3 stimuli) to ask subjects about their probability estimates and confidence (see Fig 1). Probing subjects more often would have provided more information on their internal estimates; however it would also have disrupted more their effort to integrate serial observations, which is critical to estimate transition probabilities. Asking every 15 stimuli is thus a compromise. The raw data are provided as Supporting Information (S1 Dataset, see S1 Text for a description).
20 participants (12 females, mean age 25, sem: 0.76) were recruited for the control experiment. The key difference compared to the main task was that subjects were only asked the confidence question. The other task parameters were identical, excepted a minor modification: subjects used a four-step scale instead of a continuous scale to report their confidence level. Subjects first performed one session of the main experiment which served as training. Then, they performed four sessions of the modified task.
All participants received detailed explanation about how the sequences are generated. An interactive display made intuitive the notions of transition probabilities, jumps and randomness. Transition probabilities were framed as state-dependent probabilities: e.g. if the current stimulus is A, there is an 80% chance that it is repeated and a 20% chance that it changes for B. For each state ('after A' and 'after B') these contingencies were presented as pie-charts. Random sampling from these contingencies was illustrated as a 'wheel of fortune': a ball moved around the pie chart, with decreasing speed, and the final position of the ball determined the next stimulus (A or B). Participants could repeat this process and simulate a sequence of stimuli until they felt familiar with the generative process. To introduce the concept of jump, a dedicated key press triggered a change in the pie-chart (hence, in transition probabilities).
During the task, subjects were instructed to report jumps. They could press a key at any moment to pause the sequence and access the bottom right-hand screen shown in Fig 1. By adjusting the counter displayed, they specified when the jump occurred (e.g. '13 stimuli ago'). It was made clear that 1) the estimation and confidence questions would be prompted automatically, 2) the occurrence of questions and jumps was predefined and independent so that it was unlikely that a question prompt would coincide with a jump and 3) answers in the task had no impact on the actual generative transition probabilities.
We used two methods to analyze the accuracy of jump detection. The first is the classic approach of the Receiver Operating Characteristic (ROC): the reported jumps were compared to the actual, generative jumps. The second approach is a follow-up of the ROC analysis, benefiting from the Ideal Observer perspective: the binary subjective reports (there is a jump vs. there is not) were compared with the continuous, normative posterior probability of a jump.
For both approaches, we sorted the subjects' responses into hits and false alarms. Given the stochastic nature of the task, it is difficult to detect exactly when a jump occurred. Consider for instance the sequence:
A1 B2 A3 A4 A5 A6 B7 A8 A9 A10 B11 B12 B13 B14 A15 B16 B17 B18
Subscripts indicate stimulus position and the italic font indicates the second chunk. These chunks were generated from the following transition probabilities: low for AB and high for BA from stimulus 1 to 9; high for AB and low for BA from stimulus 10 to 17. The true generative jump occurred at stimulus 10, yet it seems more likely to have occurred at stimulus 11: A9A10 better fits in the first chunk in which the AA transition rate is high. To circumvent this issue, we tolerated some approximations in the jump detection by counting a hit when there was a true generative jump within a window of ±5 stimuli around the reported jump location, and a false alarm otherwise. This same window size was used throughout our data analysis, and other choices did not change the qualitative findings.
In line with the ROC approach, we computed, for each subject, the difference in hit rate minus false alarm rate, known as the informedness index. Informedness is bounded between -1 and 1, with values higher than 0 denoting a detection better than chance; and lower than 0 a detection worse than chance. A t-test on informedness revealed that the mean value was significantly larger than zero. However, to make sure that such a result was unlikely to emerge by chance from the detection characteristics of our subjects and the generative structure of our sequences, we adopted a more conservative permutation-based approach. We computed a null (chance-level) t-value distribution for informedness by keeping subject reports unchanged but randomly regenerating (10000 times) the stimulus sequence. The p-value reported in the text corresponds to the probability of observing a t-value equal or higher under the null distribution, indicating how likely it is that the result is due to chance.
We followed up the results of the ROC analysis by inspecting the posterior probability of jump estimated by the Ideal Observer in trials corresponding to the subjects' hits, misses, false alarms and correct rejections. More precisely, since we tolerated a margin of ±5 stimuli in the subjects' jump detection, we compared the subjects' report with the posterior probability that a jump occurred in a window of ±5 stimuli around each observation, see Fig 2E for an example session. For hits and false alarms, we took the posterior probability of a jump at the position reported by subjects, given the sequence they had observed when they reported it. It is less straightforward for misses and correct rejections since, precisely, jumps were never reported at these positions. We thus estimated for each subject the typical latencies of jump report and we averaged over this list of latencies to compute the posterior probability of jump at each position corresponding to a correct rejection or miss.
To assess the accuracy of the subjects' probability estimates and confidence ratings, we used several regressions against predictor variables. The significance of these regression analyses was estimated by computing regression coefficients at the subject-level as a summary statistic and then comparing these coefficients against zero with a two-tailed t-test at the group level (t and p-values are reported in the text). All regression models included a constant and the z-scored regressors of interest.
The multiple regressions corresponding to Fig 6 deserves more details. In Fig 6A the estimation revision is the absolute difference of the Ideal Observer probability estimates between two consecutive similar transitions. Consecutive transitions are not necessarily consecutive stimuli (e.g. the transition 'from A' in ABBBBAA). In Fig 6B, the jump-wise count of samples was also made per transition type. For this count, a log-scale was used since it is an analytical result that, on average, confidence (the Ideal Observer log-precision) should increase linearly with the log-number of samples. We based this count on the Ideal Observer. However, the Ideal Observer does not estimate a binary variable (there is a jump vs. there is not), instead it computes the continuous posterior probability that a jump occurs at each position of the observed sequence, and it revises this estimate each time a new observation is made. We therefore transformed the posterior probability estimates (a two-dimensional matrix, see Fig 2E for an example) into discrete jumps. The thresholded (two-dimensional) posterior probability serves to identify when the sequence should be interrupted to report a jump and what should be the location of the reported jump, e.g, report at trial W that a jump occurred at position Z (thus W-Z trials ago). The posterior jump probability being relatively smooth (e.g. in Fig 2E), the thresholding forms patches. Each of these patches corresponds to a jump; the reported W and Z corresponds to the coordinates of the upper limit of each patch. We used a Receiver Operating Characteristic to identify the threshold (posterior probability = 0.25) that maximized the accuracy of this discretization, with respect to the actual generative jumps: we searched the threshold that resulted in the maximal difference between hit and false alarm rates.
We took as an estimate of single-trial accuracy, the un-signed error (i.e. the distance) between the subject estimate and the Ideal Observer estimate. The probability estimates in both the subjects and the Ideal Observer are expressed on the same probability scale: they can be compared directly. This is not the case for confidence: the scale for the Ideal Observer is normative, it is the log-precision which can be potentially infinite; by contrast for subjects the scale was bounded and qualitative, the mapping between confidence levels and the scale is thus highly subjective. To express the Ideal Observer and the subject confidence on a common scale, we adjusted their offset and scaling based on a linear fit.
For each subject, the single-trial accuracies in probability estimates and confidence ratings were taken into a Pearson correlation over trials. The resulting correlation coefficients could then have been taken into a classical t-test; however, we wanted to estimate to what extent the correlation would be positive due to a systematic mapping between probability estimates and confidence ratings. We thus devised two permutation-based estimations, each corresponding to a null-hypothesis distribution of the correlation of accuracies between probability estimates and confidence ratings. Shuffling #1 (Fig 5C, middle) preserved the mapping but disrupted the sequence, by keeping pairs of probability estimates—confidence ratings and shuffling their order in the sequence separately for the Ideal Observer and the subjects. Shuffling #2 (Fig 5C, right) disrupted both the mapping and the sequence by shuffling the trials independently for probability estimates and confidence ratings, thus removing any correlation between them. 10000 distinct permutations were used to estimate each null distribution. Given that the shuffling was applied within-subject, we computed the null t-distribution for the paired differences between 'Observed data' and 'Shuffling keeping pairs'. The 'Full shuffling' resulted in values close to 0 for all participants so that the estimated null t-distribution was equivalent to the parametric t-distribution; tests against the 'Full shuffling' null were thus classical t-tests against 0. P-values in Fig 5C correspond to one-tailed t-test.
We derived mathematically the optimal observation-driven estimates of the transition probabilities and jump locations: the so-called Ideal Observer. This optimal inference relies on Bayesian principles and returns a distribution of estimates p(θ | y), i.e. the posterior distribution of the transition probability, θ, at each time step in the experiment, given the observed sequence of stimuli, y. From this distribution, we derive the expected value of the inferred transition probability: μ = ∫ θp(θ | y)dθ and the confidence in that estimation, which we defined as its log-precision: -log(∫(θ − μ)2 p(θ | y)dθ).
We designed two algorithms for this Ideal Observer: a sampling approach and an iterative approach. The iterative approach was used to double check the sampling approach: both provided numerically similar values of probability estimates, confidence levels and jump location. The sampling approach explicitly computes the likelihood of possible decompositions of the sequence into chunks, whereas the iterative approach computes the likelihood that a jump occurred at any given position, independently from the other potential positions. The sampling approach is computationally slower but it allows a straightforward estimation of jump-related statistics used here: 1) The likelihood that a jump occurred around a given position, e.g. within a window of ±5 stimuli; 2) The variance of the estimated length of the current chunk, which reflects the precision of the knowledge of the observer about the last jump location. The derivation of each algorithm is presented in detail below. Computations were performed numerically in Matlab using regular grids.
If we assume that the transition probabilities generating the sequence are stable over time, then the inference can be computed analytically: the posterior distribution is a function of the number of transitions observed in the sequence. The formula is derived in the first sub-section below. However, sequences in the task were generated with jumps. For a given partition, the inference of transition probabilities can be made chunk-wise using the above-mentioned formula. Such an inference is conditional in the sense that it is computed given a particular partition. However, the partition itself is unknown and must be inferred from the sequence observed. The estimation of the transition probabilities must therefore factor out the uncertainty in the partition, which is achieved by marginalizing the conditional inference over all partitions:
p(θ|y1,…,yt)=∑πp(θ|y1,…,yt,π)p(π|y1,…,yt)
(1)
Where y is the sequence of A and B stimuli, θ = [θA|B, θB|A] are the transition probabilities 'from B to A' and 'from A to B' respectively, and π is a partition describing the location of jumps. The 1st term of the sum is thus the conditional posterior distribution of transition probabilities given a particular partition of the sequence; the second term is the posterior probability of this partition.
The sequence length being 380, there are 2380 possible partitions of the data. The exact inference would require that we compute the sum over these 2380 partitions. It is computationally intractable and actually not necessary: most partitions are very unlikely and contribute little to the sum. The posterior distribution of transition probabilities can thus be approximated numerically by averaging the conditional posterior distributions of transition probabilities over a subset of partitions sampled uniformly [22]. The second subsection below shows how to sample uniformly from the posterior distribution of partitions.
It is not necessary to decompose the sequence explicitly into a partition to compute the posterior θ distribution given the stimuli observed. Indeed, if we know θ at position t in the sequence, then at position t+1, θ should remain the same if no jump occurred, or be different if a jump occurred. In case a jump occurred, the new θ is sampled from the prior distribution and the likelihood can be assessed given the (t+1)-th stimulus. In that case, the observations made before t become no longer needed to estimate θ after t. This so-called Markov property makes it possible to estimate θ iteratively, by going forward: at stimulus t+1, we update the estimate made at time t, based on the new observation.
In the following we derive the forward algorithm to estimate θ, the transition probabilities. We also derive a backward algorithm to estimate the likelihood of jumps in the observed sequence. Note that both algorithms are provided with the exact same observations as those presented to the subject. In particular, the backward sweep does not benefit from extra stimuli not yet observed by the subject, but simply processes the information received by moving backward in time.
|
10.1371/journal.pntd.0004334 | Schistosoma mansoni Egg, Adult Male and Female Comparative Gene Expression Analysis and Identification of Novel Genes by RNA-Seq | Schistosomiasis is one of the most prevalent parasitic diseases worldwide and is a public health problem. Schistosoma mansoni is the most widespread species responsible for schistosomiasis in the Americas, Middle East and Africa. Adult female worms (mated to males) release eggs in the hepatic portal vasculature and are the principal cause of morbidity. Comparative separate transcriptomes of female and male adult worms were previously assessed with using microarrays and Serial Analysis of Gene Expression (SAGE), thus limiting the possibility of finding novel genes. Moreover, the egg transcriptome was analyzed only once with limited bacterially cloned cDNA libraries.
To compare the gene expression of S. mansoni eggs, females, and males, we performed RNA-Seq on these three parasite forms using 454/Roche technology and reconstructed the transcriptome using Trinity de novo assembly. The resulting contigs were mapped to the genome and were cross-referenced with predicted Smp genes and H3K4me3 ChIP-Seq public data. For the first time, we obtained separate, unbiased gene expression profiles for S. mansoni eggs and female and male adult worms, identifying enriched biological processes and specific enriched functions for each of the three parasite forms. Transcripts with no match to predicted genes were analyzed for their protein-coding potential and the presence of an encoded conserved protein domain. A set of 232 novel protein-coding genes with putative functions related to reproduction, metabolism, and cell biogenesis was detected, which contributes to the understanding of parasite biology.
Large-scale RNA-Seq analysis using de novo assembly associated with genome-wide information for histone marks in the vicinity of gene models constitutes a new approach to transcriptome analysis that has not yet been explored in schistosomes. Importantly, all data have been consolidated into a UCSC Genome Browser search- and download-tool (http://schistosoma.usp.br/). This database provides new ways to explore the schistosome genome and transcriptome and will facilitate molecular research on this important parasite.
| Schistosomiasis is a public health problem caused by parasites of the genus Schistosoma, of which S. mansoni is the primary causative agent. The parasite has a complex life cycle; their sexual reproductive stage is dependent on female and male adult worms mating inside the mesenteric circulation of the human host, with the female releasing hundreds of eggs daily. This phase of the life cycle is responsible for the development of pathology, which is proportional to the number of eggs accumulating in the liver and intestine of the human host. Genome and transcriptome sequencing of this parasite represent important advances in schistosome research, but there is still a need for integrated analyses to better understand the biology of this parasite. In this study, we describe the first large-scale transcriptomes of eggs, and female and male adult worms, the parasite forms that are mainly responsible for the pathology of schistosomiasis. We were able to cross-reference the gene transcription regions with promoter regions, thus improving the gene annotations. Moreover, we identified the expression of novel protein-coding genes not yet described in the current genome annotation, advancing the biological knowledge regarding this parasite.
| Schistosomiasis is a parasitic disease caused by blood-dwelling worms of the genus Schistosoma. It is an important public health problem, with high morbidity and mortality in endemic countries. Over 230 million people worldwide are infected by Schistosoma spp., comprising three main species [1]; Schistosoma mansoni is the species responsible for infecting people in the Americas, Middle East and Africa [1]. The parasite has a complex life cycle that includes several morphological phenotypes in the intermediate Biomphalaria spp. snail host and in the human definitive host, with adult worms having separate sexes, and the mated female worms releasing hundreds of eggs daily in the mesenteric circulation of the human host [2].
In the past decade, genomic tools have helped to reveal relevant molecular players in parasite biology. Thus, the S. mansoni genome was sequenced [3], followed by a systematically improved high-quality version of the genome [4]; however, the latter still includes over 800 genome sequence fragments and a number of incomplete gene annotations. The first S. mansoni transcriptome analyses from six different life cycle stages (cercaria, schistosomulum, adult worm, egg, miracidium and germ ball) were performed using first-generation EST bacterial cloning and sequencing technology [5], with limited sequencing depth. Later, with the new second-generation, cloning-independent RNA-Seq techniques, mixed-sex adult worms [4] or male adult worms [6] were studied; however, no separate male and female gene expression was assessed by RNA-Seq.
The comparison of S. mansoni male and female adult worm transcriptomes was performed only using oligonucleotide microarrays, which use short (60 nt) probes to detect known genes [7], and using Serial Analysis of Gene Expression (SAGE), which sequences short (10–14 bp) SAGE tags [8] that need to be matched to previously known full-length gene sequences to unequivocally assign their identity. In addition, the egg transcriptome was analyzed only once using Sanger sequencing [5], and no additional studies have been performed on this life-cycle stage using unbiased, large-scale sequencing.
In this study we compared egg, female and male transcriptomes of S. mansoni using large-scale RNA-Seq, which enabled the identification of genes functionally related to these specific developmental forms of the parasite. In addition, we documented the extension of the 5’-ends of hundreds of transcripts over the previous S. mansoni sequence predictions [4]. We then cross-referenced the new genomic coordinates of these transcript start sites (TSSs) with the genomic coordinates of the publicly available dataset of promoter-associated H3K4me3 histone marks obtained by ChIP-Seq [9], and the coincident coordinates for the TSS and the H3K4me3 provided genome-wide evidence of a possible involvement of this histone modification in the transcriptional regulation of these genes. Moreover, an in silico analysis of the de novo-assembled transcriptome revealed new protein coding genes with conserved domains not previously predicted in the parasite genome.
Approximately 200 S. mansoni (BH strain) adult worm pairs were freshly obtained through the periportal perfusion [10] of hamsters infected 6–8 weeks earlier with 200–300 mixed-sex cercariae. After perfusion, the adult males were separated from the adult females by keeping the parasites in 150 mm culture dishes at room temperature for a period of 15 to 30 min in Advanced RPMI Medium 1640 (Gibco, #12633–012, Thermo Scientific, USA) supplemented with 10% heat-inactivated calf serum (freshly added), 12 mm HEPES (4-(2-hydroxyethyl) piperazine-1-ethanesulfonic acid) pH 7.4, and 1% antibiotic/antimycotic solution (Gibco, #15240–096).
Eggs of S. mansoni were isolated from the livers of five hamsters, each infected 6–8 weeks earlier with 200–300 cercariae, using the method described by the group of Brindley and collaborators [11]. Immediately before the hamster liver extraction, the S. mansoni adult worms were collected by periportal perfusion [10].
PolyA+ RNA was extracted from eggs, from 200 adult male worms and from 200 adult female worms using two rounds of the FastTrack MAG Maxi mRNA Isolation Kit (Invitrogen, Thermo Scientific, USA), as described [6], with the following modifications: 1,400 units of RNase Out (Invitrogen) and 20 mM of Vanadyl Ribonucleoside Complexes (VRC) were added to Lysis Buffer L4 at the first step of the sample preparation; at the final binding step, an additional 1,400 units of RNase Out (Invitrogen) was added to the sample while the tube remained in the rotator; treatment with 30 units of DNase I Amplification Grade (Invitrogen) was performed for 45 min at room temperature; and six washings were performed at the final washing step before elution. These modifications resulted in PolyA+ RNA samples with small percentages of rRNA contamination (3 to 7%), as estimated with a Bioanalyzer using the RNA 6000 Pico Kit (Agilent Technologies, Santa Clara, CA, USA).
Because the Roche 454 sequencing platform did not provide a protocol for the construction of strand-oriented RNA-Seq libraries, we previously developed a method for generating strand-oriented cDNA libraries for 454 sequencing [6], and here, we have improved the protocol, aiming to correct the tendency toward preferential sampling of the 3′-end of the transcripts verified in the previously described method [6]. In the new strand-oriented 5′-end-first cDNA libraries method, the design of the primers was modified so that sequencing initiates from the 5′-end of the transcripts. The resulting single-stranded cDNA (sscDNA) libraries were PCR amplified (10 cycles of PCR for male and female libraries, and 17 cycles of PCR for egg libraries), and the resulting directional double-stranded cDNA (dscDNA) libraries were purified using AMPure beads (Agencourt, Beckman Coulter, Indianapolis, IN, USA) and quantified using PicoGreen (Invitrogen). A detailed description of the library construction is given in the Supplementary Methods in S1 Text, along with a summary scheme of the procedure (see Fig A in S1 Text).
Directional dscDNA libraries generated as described above were sequenced from the 5′-end using the emPCR Titanium Kit and the Titanium Sequencing Kit on a Roche 454 Genome Sequencer FLX instrument, following the manufacturer’s instructions [12]. Data processing used standard 454 software procedures to generate nucleotide sequences and quality scores for all reads. High-quality (minimum Phred score 20) trimmed reads were generated, and the sequence complexity filtering criteria were applied using PRINSEQ [13]. Sequencing data were deposited at the NCBI Sequence Read Archive (SRA) under the accession number SRP063353.
To exclude reads of ribosomal, mitochondrial or transposable element origin, we performed an alignment of reads against rRNA, mitochondrial sequences and 29 S. mansoni transposon sequences for which there were published curated full-length sequences [14–16] utilizing BLASTn [17], and reads displaying an alignment with an e-value ≤ 10−15 were deleted. To evaluate the coverage of the reads, the filtered reads were mapped to the genome with Tophat2 [18], and the distribution along the 5′ to 3′ gene body was assessed using the RSeQC package [19].
The S. mansoni transcriptome was reconstructed without a reference genome, using all reads from egg, female and male RNA-Seq with a Trinity assembler (Version from April 2014) [20] and a k-mer size of 25. This project has been deposited at the Transcriptome Shotgun Assembly (TSA) division of DDBJ/EMBL/GenBank under the accession numbers GDQY00000000 (putative novel genes) and GDUI00000000 (all other contigs). The version described in this paper is the first version, GDQY01000000 and GDUI01000000.
For annotation, the contigs resulting from the Trinity assembly were aligned to the predicted Smp gene sequences from the S. mansoni genome version 5.2, utilizing BLASTn [17] and applying as threshold an e-value = 10−5, coverage ≥ 20% and strand specificity. In the case of the alignment of a contig to multiple Smps, the hit with higher identity was considered as the correct Smp alignment. For each Smp with multiple contigs aligned to it, we considered the contig with the highest coverage of the Smp as the representative of evidence of expression of that Smp.
All Trinity-assembled contigs were aligned to the genome sequence using BLAT [21]. Contigs with genomic coordinates that did not intersect the genomic coordinates of any Smp-predicted gene were analyzed with InterProScan [22] using the Pfam database [23], Batch CD-search tool [24] and Coding Potential Calculator (CPC) [25] to assess the protein-coding potential of these transcripts. Contigs with a Pfam conserved domain were clustered into categories using domain-centric Gene Ontology (dcGO) [26] and the online tool CateGOrizer (http://www.animalgenome.org/tools/catego/) with the GO_slim2 option.
Multiple alignments of Lifeguard proteins from S. mansoni, from five other schistosome species (obtained from ftp://ftp.sanger.ac.uk/pub/pathogens/HGI/) and from various invertebrate and vertebrate species were restricted to the BAX1-I domain and were performed utilizing the ClustalX2 program [27]. Complete sequences for the orthologs with high identity to SmDLFG1 and 2 proteins were obtained from the five other schistosome species by analyzing their preliminary genome sequence with the program Spaln [28]. Except for the latter Schistosoma sequences, the accession numbers of the sequences used in the analysis are indicated in Fig G in S1 Text. Phylogenetic analyses were performed using Bayesian inference methodology using MrBayes program v3.2.2 x64 [29]. The analysis was performed using default parameters, except for the use of the command “prset aamodelpr = mixed,” which enables sampling across all fixed amino acid rate matrices (models for amino acid evolution) implemented in the program. Analyses were stopped after 1,000,000 generations, with samplings every 100th generation. Tree information was summarized utilizing “sumt burnin = 2500”, which discards the first 250,000 generations. A measured potential scale reduction factor (PSRF) parameter equal to 1 was obtained using the “sump burnin = 2500” command, indicating a convergence of the analysis. The resulting tree was visualized using the TreeView program [30].
Transcript abundance in eggs, females and males was quantified using the Trinity assembly output and the reads from each form as input for the Sailfish tool [31]. The number of reads was normalized using the upper quartile, correcting for the different sequencing depths of the libraries. Significant differential expression between two conditions (egg versus female; egg versus male; female versus male) was computed using the NOISeq program [32] with the NOISeq-sim option and the following parameters: nss = 5, to simulate five technical replicates, each comprising 20% of the reads in the dataset (pnr = 0.2), allowing a small variability (v = 0.02). To identify contigs with significant differential expression, a probability P ≥ 95% was used as the cutoff. Next, with the list of contigs representative of each Smp, we searched for the most highly expressed genes in the egg, the female or the male. For this purpose, we identified genes that simultaneously had a significantly higher expression in eggs in the NOISeq comparison with both males and females, and we repeated the procedure, identifying the genes that simultaneously had a significantly higher expression in females than in both eggs and males, as well as the genes that simultaneously had a significantly higher expression in males than in both eggs and females; these genes were flagged on the full list of significantly differentially expressed genes as Egg_High, Female_High or Male_High, respectively.
The most highly expressed genes in eggs, females or males were categorized using Gene Ontology (GO) terms and the Ontologizer tool [33], with all genes detected in the transcriptome as background. GO terms for S. mansoni genes were obtained from the Metazoa Mart database (http://metazoa.ensembl.org/biomart/martview/), and p-values were calculated by the parent-child union method.
To refine the S. mansoni predicted gene model structures, we mapped our Trinity-assembled contigs to the genome using BLAT [21]. We then used Bedtools utilities [34] to cross-reference the mapped coordinates of our RNA-Seq transcripts with the coordinates of the coding sequences of the Smp predicted genes [4] and flagged the Smp predicted genes that were extended at the 3′-end and/or the 5′-end. Moreover, we flagged genes in the vicinity of one another that were merged by our RNA-Seq data. Next, we compared our list of extended genes obtained above with the 5′- and 3′-UTR annotations available for 2,160 Smp genes [4] (excluding the UTR annotations of another 617 Smp genes for which the UTR coordinates are inconsistent with the coordinates of the coding sequence), and we flagged the Smps for which our assembled contigs predicted a different or a longer UTR.
To assess further evidence for the transcription start site (TSS) of S. mansoni genes, we downloaded the Chromatin Immunoprecipitation Sequencing (ChIP-Seq) dataset of adult worm Histone H3K4me3 (SRR1107840) [9], and we used the HOMER pipeline [35] to map the ChIP-Seq sequences to the genome and to find the genomic coordinates of H3K4me3 enriched peaks. HOMER found enriched peaks by calculating the density of the reads at the peaks that should be at least 4-fold higher than the peaks in the surrounding 10 kb region [35]. We then used Bedtools with a window of ± 500 bp to search for an H3K4me3 peak around the Smp gene 5′-end or around the RNA-Seq transcript 5′-end. We flagged the RNA-Seq transcripts that extended the 5′-end of Smp genes and had an H3K4me3 peak around this new 5′-end. Using the Bedtools closest function, we compared all expressed Smp genes of males and females having the H3K4me3 mark near (within +/- 500 bp of) the 5´-end against the dataset of expressed Smp genes without the histone mark to evaluate the prediction of gene TSS.
Our gene models (Trinity-assembled contigs) and all the histone mark data [9] mapped to the genome are accessible through a local installation of the UCSC Genome Browser in a Box (GBiB) [36] at http://schistosoma.usp.br/.
Total RNA was extracted from eggs using Trizol reagent (Invitrogen), according to the manufacturer's instructions, and treated with DNAse I using the RNeasy Micro Kit (QIAGEN, Germantown, MD, USA). RNA from adult female and male worms was extracted and treated with DNase I using the RNeasy Mini Kit (QIAGEN). Three biological replicates were assessed for each life cycle stage and form. The Reverse Transcriptase (RT) reaction was performed with 1.0 μg of each total RNA sample using the SuperScript III First-Strand Synthesis SuperMix (Invitrogen). Each real-time qPCR was run in three technical replicates with Sybr Green PCR Master Mix (Applied Biosystems, Thermo Scientific, USA), 160 nmol of each primer (Forward and Reverse) and cDNA from the reverse transcription, using the 7500 Real-Time PCR System (Applied Biosystems) with the default cycling parameters. Specific pairs of primers (S1 Table) for selected genes were designed using the Primer 3 tool (http://biotools.umassmed.edu/bioapps/primer3_www.cgi) with default parameters, anchoring each primer of a pair on a different exon. The housekeeping gene PAI1 (Smp_009310) was chosen from nine genes that showed no differential expression in the RNA-Seq data, and the real-time qPCR data for the nine genes is shown in Fig B in S1 Text. Data were analyzed using the RefFinder tool [37] to determine a geometric mean of ranking values for each gene among the three stages and to choose the most stable gene for qPCR normalization (lowest ranking gene). Real-time data were normalized in relation to the level of expression of the PAI1 gene, and p-values were determined by one-way analysis of variance (ANOVA) and Tukey post-hoc tests. The statistical significance of the correlation between the RNA-Seq and qPCR data was calculated using the Spearman test.
Infected hamsters were maintained at the Instituto Adolfo Lutz, and the Comissão de Ética no Uso de Animais do Instituto Adolfo Lutz (CEUA-IAL) reviewed and approved the animal care and use protocol, license number 07/2013. The experimental procedures were conducted according to the Brazilian national ethical guidelines for animal husbandry (Lei 11794/2008).
We performed separate sequencing of dscDNA strand-oriented libraries generated using RNA transcripts of seven-week-old S. mansoni adult male or female worms recovered from hamster portal vein perfusions and of S. mansoni eggs recovered from hamster livers. A total of ~2.6 million RNA-Seq reads was obtained, and reads matching transposon, mitochondrial and ribosomal genes were filtered out, resulting in ~1.5 million high-quality, strand-oriented, long reads, with an average length of 278 nt (ranging from 40 to 1,026 nt) (Table 1). Each of the three parasite forms was sampled (250 to 730 thousand reads each, Table 1); however, the egg sequencing depth was hampered because the RNA yield, purity and stability were lower than those of the adult worms.
Using the Trinity de novo assembler [20] without a reference genome, we obtained 23,967 contigs representing the S. mansoni transcriptome from eggs and adult worms, with an average contig length of 669 nt (ranging from 201 to 6,508 nt) (Table 2). Of this total, 3,799 contigs represented different isoforms (Table 2) assembled by Trinity that belonged to 1,676 putative alternatively spliced transcript fragments, i.e., an average of 2.3 isoforms per contig. The remaining 20,168 contigs corresponded to unique transcript fragments, with no evidence of alternative splicing (Table 2).
We then mapped the contigs to the genome to cross-reference the genomic coordinates with the coordinates of predicted Smp genes [4] and found that 13,268 contigs (55% of total) matched 6,760 known predicted Smp genes (Table 2), an average of two contigs per targeted Smp gene. Notably, a large number of contigs (10,472, 44% of total) had no overlap with any predicted Smp gene (Table 2), and these contigs were found to map to intergenic regions of the genome, to intronic gene regions or to the antisense strands of Smp genes. Less than 1% of contigs (227 contigs) did not map to the genome (Table 2). Some of these contigs might belong to genome sections that have not been sequenced yet.
We subsequently cross-referenced the genomic coordinates of known S. mansoni retrotransposons [38] with the coordinates of the contigs that mapped outside of Smp genes (intergenic and intronic antisense regions), searching for the intersection of coordinates between them. We found that despite having filtered out individual transposon reads from the input file, the assembly still contained a few contigs (98 out of the 10,472, i.e., 0.9%) that overlapped with retrotransposon regions.
We detected evidence of expression for 6,760 predicted Smp genes (Table 3) by cross-referencing the Smp gene coordinates with the coordinates of contigs that matched those Smps (see S2 Table for the list of genes and their expression levels). Of this total, 4,610 genes were expressed in the egg stage, 6,288 were expressed in females and 4,947 genes were expressed in males (Fig C in S1 Text). Interestingly, 3,443 genes were expressed in all three parasite forms (Fig C in S1 Text).
The quantitative expression level of Smp genes for each parasite form (egg, female or male) was assessed by counting the number of reads from the respective library that matched to each contig, using normalized expression data to correct for the differences in sequencing depth for each dataset. Subsequently, pairwise comparisons between the stages identified the contigs with significant (P ≥ 95%) differential expression among the three conditions; a total of 4,364 contigs were detected as significantly differentially expressed in one stage compared with at least one other stage, and these differentially expressed genes are shown in Fig 1A (see S2 Table for the list of genes and their corresponding differential expression significance).
The number of Smp genes that were most highly expressed in one given form compared with both the other two forms was identified (Table 3), and these genes were flagged as Male_High, Female_High or Egg_High in S2 Table. By this approach, we detected 510 genes most highly expressed in males, 672 genes most highly expressed in females, and 615 genes most highly expressed in eggs (S2 Table). This set of genes corresponds to a representative differential gene expression profile for each parasite form.
Interestingly, we found five micro-exon genes (MEGs) most highly expressed in males, namely two MEG-4 genes (Smp_085840 and Smp_163630), MEG-8 (Smp_171190), MEG-11 (Smp_176020) and MEG-14 (Smp_124000) (S2 Table). In addition, MEG-1 (Smp_122630) was most highly expressed in females, while MEG-5 (Smp_152580) was detected as highly expressed both in females and males. Three MEGs were most highly expressed in eggs, namely two MEG-2 genes (Smp_159810 and Smp_180310) and MEG-3 (Smp_138080).
To identify significantly enriched gene categories among the genes most highly expressed in eggs, females or males, we performed GO analyses. Significantly enriched GO categories (p-value ≤ 0.01) identified gene groups related to a number of different parasite development and maintenance biological processes (Fig 1B and S3 Table). Some categories were present in more than one parasite form but with a significant enrichment p-value in only one. Among the Molecular Function and Biological Process ontologies, the three most significantly enriched GO categories in eggs were carbohydrate phosphatase activity (GO:0019203), lipid transport (GO:0006869) and response to stress (GO:0006950). In females, the enriched categories were cellular protein modification process (GO:0006464), DNA metabolic process (GO:0006259) and catalytic activity (GO:0003824). In males, the three most significantly enriched GO categories were calcium ion binding (GO:0005509), potassium ion transport (GO:0006813) and protein tyrosine kinase activity (GO:0004713).
Taken together, these results point, for the first time, to relevant biological processes enriched in the S. mansoni egg stage compared with adult worms. In addition, these results indicate a set of genes involved in biological processes enriched in either S. mansoni male or female worms.
In the literature, there is an absence of qPCR data comparing egg, female and male gene expression; similarly, a control housekeeping gene for S. mansoni has been previously evaluated in the literature only for mixed sex adult worms [7,39]. Additionally, the α-tubulin gene has been used in many studies as a housekeeping gene for normalization among all life cycle stages; however, in our RNA-Seq and qPCR data, α-tubulin is highly differentially expressed in eggs, with approximately 25 times higher expression in this stage when compared with its expression in male and female adults, as detected by qPCR (Fig 2A). In this context, we chose the Plasminogen Activator Inhibitor PAI1 gene (Smp_009310.1) as a housekeeping gene from 9 possible candidates identified by searching the set of genes expressed by all three parasite stages in the RNA-Seq data for the genes that were not differentially expressed in the three RNA-Seq pairwise comparisons of this study. We confirmed PAI1 by qPCR as an adequate housekeeping gene for use in the normalizations by performing qPCR for all 9 candidate normalizer genes and analyzing the data with RefFinder [37], as shown in Fig B in S1 Text.
Using PAI1 as a housekeeping gene, we selected and tested by qPCR six differentially expressed genes for each parasite form to provide for an independent validation of the RNA-Seq data with different biological samples. These genes were selected for their association with important biological functions in each of the parasite forms, as noted later in the Discussion.
We selected glycolipid transfer protein (GLTP), tubulin, translocase of outer membrane 70 (TOM70), RNA polymerase I (PolI), DEAD box RNA helicase (DDX) and nuclear receptor SmE78 as genes highly expressed in eggs. All genes displayed higher expression in eggs when assayed by qPCR, but only four genes were statistically significant when compared with their expression in males and females; in one additional case, only the difference between eggs and males was statistically significant (Fig 2A). The egg RNA-Seq data for the selected genes correlated well with the qPCR data, with a Spearman’s correlation of 0.77 and p-value < 0.001.
We selected the tyrosinases Tyr1 and Tyr2, p14, eggshell protein, trematode eggshell protein and ATP-binding cassette transporter as genes highly expressed in females, based on the RNA-Seq data. The qPCR confirmed the higher expression in females compared with that in eggs and males for all six selected genes (Fig 2B), with a Spearman’s correlation between RNA-Seq and qPCR of 0.80 and p-value < 0.0001.
We selected Na/K ATPase, calcium binding protein (CaBP), Discoidin domain receptor (DDR), serotonin receptor (5HTR), Wnt5 and scavenger receptor CD36 as genes highly expressed in males, based on the RNA-Seq data. All genes displayed higher expression in males, but this difference was statistically significant when compared with the expression in eggs and females in only four of the genes. In the two remaining cases, the difference was statistically significant only when compared with either females or eggs (Fig 2C). The Spearman’s correlation between the qPCR and RNA-Seq data for males was 0.50, and p-value = 0.034.
RNA-Seq contigs were used to improve the Smp gene model predictions from the 5.2 version of the genome [4], which includes a total of 10,852 Smp gene models. RNA-Seq Trinity-assembled contigs that mapped to the genome with genomic coordinates intersecting the coordinates of gene model exons of any predicted Smp transcribed in the same strand were considered as evidence of an mRNA transcribed from that gene. Using this approach, it was possible to document the extension of Smp genes, and these extensions could occur either at the 5′- or the 3′-end of genes. Our transcriptome data extend the 5′-ends of 3,337 genes and the 3′-ends of 2,417 genes, of which 747 were extended at both ends (S4 Table). It is evident that the majority of genes were extended at the 5′-end, and this bias resulted from using RNA-Seq libraries that predominantly covered the 5′ region of RNAs, as documented by mapping the individual reads along the S. mansoni complement of genes (Fig D in S1 Text).
Protasio and collaborators [4] had previously improved the original Smp predicted gene models [3] by updating 731 predicted genes that were merged or split. Our transcriptome data showed that despite this gene annotation improvement, there are still an additional 589 gene models that should be altered by merging each predicted sequence with the sequence of the neighboring predicted gene (S5 Table); in each case, we found a single transcript contig that overlaps both neighboring gene predictions. A typical case involves Smp_101670 and Smp_124310, which map to adjacent regions in the genome (Chr_1:5,968,156–6,004,959) and are transcribed in the same strand; with our transcriptome data (contig c7177_g7_i1), it was possible to determine that the two predicted genes are actually part of the same transcript (S5 Table). These new gene models contribute to an improved annotation of the S. mansoni gene complement.
In eukaryotes, it is known that the histone H3 lysine 4 trimethylation (H3K4me3) mark is present in the promoter region of expressed genes, in the vicinity of their Transcription Start Sites (TSSs), and in the 5′-UTR region of genes [40]. Roquis and collaborators performed H3K4me3 chromatin immunoprecipitation followed by sequencing (ChIP-Seq) of mixed-sex adult worms [9], and we used this dataset to assess the presence of enriched peaks of the H3K4me3 mark that could indicate the TSS regions in the S. mansoni genome. The H3K4me3 peaks were found near (within +/- 500 bp of) 3,084 Smp genes, showing that these gene models have their predicted 5′-end close to the promoter region (S4 Table). Among these genes, 2,454 genes have a contig from our dataset confirming the 5′-end of the Smp gene. With the extension of the 5′-end of another 3,337 Smp genes (that were re-structured using our RNA-Seq data), the H3K4me3 TSS histone mark was identified for a further 673 re-structured Smp genes (S4 Table).
Using all expressed genes found in our RNA-Seq data, we searched for the closest H3K4me3 histone mark around the predicted Smp annotation TSS regions or the extended Smp TSS regions. Unsurprisingly, for those expressed genes for which we found an H3K4me3 histone mark near the TSS (+/- 500 bp from the TSS), we detected a peak centralized at the midpoint of the TSS, but also a tail extending downstream from the peak (Fig E in S1 Text, blue line). For those expressed genes without the histone mark at the gene TSS (no H3K4me3 within +/- 500 bp from the TSS), the closest H3K4me3 was found approximately 1–2 kbp upstream of the TSS or 1 kbp downstream of the TSS (Fig E in S1 Text, red line). These distance distribution patterns are different from the distribution pattern expected by chance, obtained in a random-position-generated TSS set (Fig E in S1 Text, gray line).
To exemplify the scenario of improved Smp gene models, we selected a specific S. mansoni locus (Chr_ZW: 2,310,000–2,350,000), which hosts the Smp_142960 gene (Fig 3, blue boxes and arrowheads) and two additional short mono-exonic genes, namely Smp_177650 and Smp_188430 (Fig 3, short blue boxes within intron 6 of Smp_142960). Our RNA-Seq assembly generated three contigs that map to the locus with a completely different architecture (Fig 3, red boxes and arrowheads); all three contigs new 5’-ends coincided with H3K4me3 histone marks (Fig 3, green peaks), confirming that they represent new genes with novel TSSs. The new gene on the upstream side of the locus (contig c3145_g1_i1) encodes an isoprenoid biosynthesis enzyme, Class 1 domain (cl00210) (e-value = 1.92∙10−6), and is not part of the glutathione synthase gene originally annotated as Smp_142960. In fact, the full domain of glutathione synthase is encoded by the new gene (contig c807_g1_i1) on the downstream side of the locus. It should be noted that the 3′-end of both these new genes was not covered by our RNA-Seq data, which is consistent with the fact that our new RNA-Seq method covered predominantly the 5′-ends of genes (Fig D in S1 Text).
This analysis cross-references ChIP-Seq information on TSS histone marks with RNA-Seq contig data and looks for evidence of an extended Smp gene prediction, where the 5′-end of the gene overlaps the TSS histone mark, and is shown here for the first time for S. mansoni, adding important information regarding the parasite’s gene regulatory regions. Using these data, it was possible to observe that the histone H3K4me3 deposition in S. mansoni frequently overlaps the first exon of the gene and extends into the first intron (Fig E in S1 Text).
The large number of contigs (10,472 contigs) with no match to Smp genes raised the hypothesis of potential new genes not yet described in S. mansoni, and transcripts with open reading frames (ORFs) longer than 150 nt were considered as candidates for further analyses. Subsequently, using the Conserved Domains Database (CDD), we investigated the similarity of the proteins encoded by these ORFs with protein-conserved domains present in proteins from other species, which identified 232 contigs encoding protein-conserved domains, thus indicating potential new genes (S6 Table), among which 159 contain full-length ORFs. These contigs were deposited at NCBI TSA under accession number GDQY01000000. For the entire set of potentially new protein-coding genes, we assessed the presence of an H3K4me3 TSS histone mark close to the 5′-end of the contigs and found 79 with evidence of this mark in the gene promoter region (S6 Table).
In this set of potential new S. mansoni genes, we searched for contigs encoding protein-conserved domains not yet reported among the Smp genes. Closer inspection revealed contigs encoding significantly conserved protein domains (score < 1 x 10−5, covering approximately 70 to 100% of the domain), none of which were present in any Smp gene (S6 Table).
Specifically, our RNA-Seq data identified an interesting new putative gene with a meiosis expressed gene (MEIG) protein-conserved domain, which was not predicted among Smp genes. This gene plays an essential role in the regulation of spermiogenesis in mammals [41] and was also detected in ovaries of mouse embryos when the oocytes reached the prophase I meiotic stage [42]. The SmMEIG gene encodes a protein with 85 amino acids and is orthologous to the S. japonicum MEIG gene (accession number CAX73271.1, identity = 91% and e-value = 6∙10−50, covering 100% of the target gene), as determined using BLASTx with the nr NCBI protein database. This putative gene was only detected in male RNA-Seq data, but the qPCR results (Fig 4) showed that this new gene is expressed in all three parasite stages; the expression in eggs is two-fold higher than that in males. This gene could play an important role in schistosome female oocyte and male sperm production, and the higher expression in eggs suggests a function not yet characterized in this parasite stage.
Another new putative gene encodes the enzyme VKOR, which is responsible for the recycling of vitamin K cofactor in eukaryotes, reducing vitamin K that is oxidized upon the carboxylation of glutamic acid residues of proteins in apoptosis, signal transduction and growth control pathways [43]. This new putative gene is expressed in eggs, females and males, as detected by RNA-Seq data and qPCR (Fig 4), showing that the parasite might use this cofactor in the metabolic pathways in all three parasite forms. This putative VKOR gene is found in many mammal species, and using the BLASTx tool, we found that the S. mansoni VKOR gene is orthologous to the H. sapiens gene with an identity of 30% and e-value of 10−12, covering 41% of the target gene.
We selected other new putative genes detected in S. mansoni by RNA-Seq for testing with qPCR, such as Reticulon and BolA. These new putative genes were detected in all three parasite forms (Fig 4), confirming their expression. Reticulon proteins in eukaryotes are localized to the endoplasmic reticulum, and there is evidence that they influence endoplasmic reticulum-Golgi trafficking, vesicle formation and membrane morphogenesis [44]. The BolA gene is widely conserved from prokaryotes to eukaryotes and seems to be involved in cell proliferation or cell-cycle regulation [45]. We confirmed the presence of these putative new genes in the three parasite forms.
We also detected a new S. mansoni gene from the transmembrane BAX inhibitor motif (TMBIM) family containing the Bax Inhibitor 1 domain (BI-1) [46]. This new protein-coding gene (contig c17331_g1_i1) encodes a 259-amino-acid-long protein that shares similarity with the Lifeguard members of this family, with an expected value of 0.003 (27% identity and 43% similarity) for the alignment with a Drosophila wilistoni protein containing a BAX1_i domain. Several other hits with similar proteins containing this domain are obtained, confirming the consistency of this result. Indeed, we constructed a PSSM matrix with the new protein utilizing PSI-blast and considering the proteins aligned with e-value cutoff of 0.01; after a single round of iteration using this PSSM matrix, at least one hundred different Lifeguard proteins were aligned with e-values lower than 10−30. This result indicates that although this new protein displays a relatively divergent sequence from known Lifeguard proteins, it retains residues that are conserved among members of this family. Moreover, a comparison between the transmembrane helix profile of this protein and of a known Lifeguard protein revealed a very high similarity, providing further evidence of homology between the sequences (Fig F in S1 Text). Therefore, we named the new gene SmDLFG1 (Schistosoma mansoni Divergent Lifeguard 1).
The RNA-Seq contig c17331_g1_i1 encoding the SmDLFG1 gene does map to two different and adjacent loci on Chromosome 1, in a region where no gene had been predicted. The first higher-score match (BLAT score = 888) is at the Chr_1:623,131–636,930 locus. The second match has some mismatches at the 5′-end of the contig and a lower matching score (BLAT score = 790). Using the genome sequence at the locus of the SmDLFG1 second match (Chr_1:606,203–612,641) and RNA-Seq data from NCBI SRA, it was possible to detect reads with a 100% match, which confirmed the expression of an isoform of SmDLFG1 that we named SmDLFG2, mapping to Chr_1:605,936–612,714.
The five Smp paralog genes from this protein family with complete BAX1_ domain (Smp_044000 or G4VD38, Smp_181470 or G4VEN1, Smp_150500 or G4VGF8, Smp_026160.1 or G4VKQ6, Smp_210790 or G4V7Q6) and the newly identified SmDLFG1 and SmDLFG2 protein-coding genes were used in a phylogenetic analysis that included representative proteins of this family from five other schistosome species as well as from various invertebrate and vertebrate species (Fig G in S1 Text).
RNA-Seq detected the expression, in the three parasite forms, of all six Smp predicted gene paralogs of the TMBIM family, but apart from the SmDLFGs, only Smp_026160 (annotated as Putative growth hormone inducible transmembrane protein) exhibited a significantly higher expression in eggs compared with males or females. The new SmDLFGs are highly expressed in females compared with males and eggs (Fig 4) (the qPCR primers did not distinguish between the two isoforms).
GO categorization of 198 out of the 232 protein-coding putative new genes, encoding conserved-protein domains, identified GO biological processes related to development, metabolism, cell organization and biogenesis (Fig 5). This dataset of putative new genes encoding conserved-protein domains provides an opportunity of identifying new important genes in metabolic pathways where steps remain missing.
The remaining 10,444 contigs not encoding conserved protein domains were classified according to their protein-coding potential with the CPC tool, which identified only 703 contigs with protein-coding potential (however with no identifiable conserved domain), whereas 9,741 contigs were classified as non-coding and therefore represented putatively expressed long non-coding RNAs (lncRNAs) (S6 Table).
The S. mansoni genome and transcriptome have been explored for the past decade [47], providing information related to the gene expression profiling and transcription regulation of certain life-cycle stages of the parasite. In this study we have obtained, for the first time, large-scale RNA-Seq separate profiles of S. mansoni females and males, as well as the egg-derived expression profile. We also used, for the first time, a combination of de novo transcriptome assembly with existing genome coordinates of predicted genes, along with newly mapped public ChIP-Seq data, which permitted the identification of novel putative S. mansoni genes. Because we have, for the first time, generated an individual gene profile for each of these three parasite forms, we explored this additional information by searching the literature for the possible functions of a selected set of genes most highly expressed in each form, as described below.
A set of 6 genes for each of the three parasite forms was selected for validation by RT-qPCR based on the fact that most of these genes were highlighted by the functional analyses mentioned above. Twenty-nine of the thirty-six comparisons (81%) performed confirmed, using RT-qPCR, the significant expression enrichment previously determined by RNA-Seq (Fig 2), a fraction of the confirmation similarly found in the literature [48]. The Spearman correlations between RNA-Seq data and RT-qPCR were in the range 0.50 to 0.80 with p-values in the range 0.034 to < 0.0001. We consider this result to be a successful validation, especially if we note that none of the directions of enrichment in the RNA-Seq and RT-qPCR data showed a conflicting opposite result; it is already known that all transcriptome techniques, including microarray, RNA-Seq and qPCR, have inherent pitfalls that affect quantification and cannot be fully controlled [49].
First, we analyzed the differentially expressed genes in the egg stage in light of the known biological characteristics of the eggs. The glycolipid transfer protein (GLTP) (Smp_076390) from the lipid transport GO category is highly expressed in eggs compared with males and females. GLTP is responsible for the intermembrane transfer of lipids linked to sugars, such as glycosphingolipids [50]. As schistosomes are unable to synthesize fatty acids, the uptake of these compounds from the host is essential. We suggest that the enrichment of lipid transport genes in the egg stage might be related to an important uptake of lipids from the host, possibly related to embryo development. In fact, it is known that the schistosome egg uptakes cholesteryl ester from HDL vesicles, mediated by the CD36 receptor, and that this uptake is important for egg maturation [51]. In this context, we speculate that because the CD36 receptor has such an important function in egg development, the enrichment of lipid transport genes among the highly expressed egg genes could be related to the transport function. Thus, GLTP could be an important gene for egg development and maintenance inside the host circulation until the eggs reach the host intestine lumen, where they are released.
Egg formation is dependent on structural components such as microtubules, an enriched GO category. An important gene from the microtubules category is tubulin (Smp_027920), whose expression was 17-fold greater in eggs compared with that in females and 34-fold greater than that in males. Microtubule genes enriched in eggs will integrate the layer between the eggshell and the developing miracidium, known as Reynolds’ layer, comprising microfibrils in a granular matrix [52].
Genes related to stress response were also increased in eggs, and in this group, we highlight the mitochondrial genes from the translocase of outer membrane (TOM) machinery that serve as the main entry gateway of preprotein into the mitochondria [53]. The mRNA expression of TOM70 (Smp_010930) and of HSP70 (Smp_065980) was detected in eggs, and the proteins encoded by these genes are key to the mitochondrial import pathway [54] and contribute to the folding and refolding of proteins after stress denaturation [55].
The mRNA level of Pol I (Smp_129500), which encodes the enzyme responsible for the transcription of ribosomal DNA (rDNA) [56], is significantly higher in eggs than in males, suggesting ribosome production and cellular proliferation inside the egg. Another gene with a high expression level in eggs encodes a protein related to transcription, namely the DEAD box RNA helicase (Smp_013790). DEAD box proteins unwind short duplex RNA and remodel RNA-protein complexes, and they are important players in RNA metabolism from transcription and translation to mRNA decay [57]. The nuclear receptor SmE78 gene (Smp_000340) also showed an increased expression in eggs. Wu et al. had previously shown that nuclear receptors are important transcriptional modulators, and in S. mansoni, SmE78 may be involved in growth and vitellogenesis [39].
To further validate the gene enrichment analysis of the egg stage, which has the lowest RNA-Seq coverage, we have manually inspected three genes known to be highly expressed in eggs, namely omega-1 [58], IPSE/alpha-1 [59] and kappa-5 [60]. Only the omega-1 and IPSE/alpha-1 genes are annotated as Smp genes, namely Smp_193860 and Smp_112110, respectively. Indeed, they are listed in S2 Table as highly expressed genes in the egg stage with the flag Egg_High, as these two genes were detected in our RNA-Seq data as significantly differentially expressed in eggs compared with males and females.
Interestingly, the omega-1 gene (Smp_193860), which encodes a protein with 127 amino acids, appears in our RNA-Seq data as contig c7716_g1_i1 with an additional five new exons at the 5’-end of the gene, and the new longer ORF encodes an omega-1 protein with 236 amino acids, which is now compatible with its described molecular weight of 31 kDa [58].
The kappa-5 gene with accession AY903301.1 [60] is not yet annotated as an Smp gene; only a paralog, Smp_150240 is annotated in the genome, with an identity of 89% and query coverage of 74% compared to the kappa-5 gene. Because the kappa-5 gene is not present in the Sanger Genome annotation, we did not detect it in the automated analysis, but by a curated manual investigation, using the NCBI accession AY903301.1, it was possible to align the sequence to the genome with genomic coordinates Chr_3:13873256–13879312 minus strand; we found that contig c6481_g1_i1 is the transcript corresponding to the kappa-5 gene. This contig is highly expressed in the egg stage, with normalized egg read counts of 1,909.35 and no reads detected in the adult male and female forms.
Subsequently, we analyzed the genes differentially expressed in females in light of the known biological characteristics of the females. Their transcriptome profile is highly linked with egg production, as the S. mansoni female produces approximately 350 eggs daily, and proteins that are important to the eggshell structure should be expressed. The hardened and tanned structure of the eggshell is derived from tyrosinase activity, which catalyzes the cross-linking of proteins known as quinone tanning [61]. The mRNA expression levels of Tyrosinase 1 (Smp_050270) and Tyrosinase 2 (Smp_013540) in females are higher than in eggs and males. Interestingly, Tyr1 has a much higher expression level than Tyr2, although the two genes encode highly similar proteins. The higher abundance in females is consistent with the fact that tyrosinase originates from the female vitelline cells inside the vitellaria, acting on eggshell formation.
The most studied S. mansoni eggshell proteins are p14 and p48. Chen et al. showed that the p14 gene is expressed only in mature female vitelline cells and is undetectable in the RNA obtained from eggs [62]. Our RNA-Seq results show that the p14 (Smp_131110) transcript was detected at low levels in eggs extracted from hamster liver, and the low abundance of this transcript in eggs was confirmed by qPCR. A high expression level of p14 was detected in females compared with males or eggs, as expected. Two other proteins that were not yet shown as part of eggshell synthesis are eggshell protein chorion (Smp_000430) and Trematode eggshell protein (Smp_000390), both with the Trematode eggshell domain (Pfam:08034). We found that the genes encoding these two proteins are highly expressed in females, and compared with p14, the expression of eggshell protein chorion gene is twice as high, pointing to a new protein for possible exploration as a candidate antigen in liver granuloma formation.
Female reproduction requires a high bioenergetic supply for the production of hundreds of eggs per day. Consistently, genes from the ATPase activity category (Fig 1B) are enriched in our analyses of female highly expressed genes, including genes related to the production of ATP through mitochondrial oxidative phosphorylation. Additionally, the ATP-binding cassette transporter gene (Smp_040540) exhibited higher expression in females than in males, but interestingly, the expression level in eggs was almost the same as that in females. In addition to the energy supply for egg production, it was recently proposed that enzymes present in the schistosome tegument could act similarly to human cell-surface ATPases, inhibiting platelet activation and modulating the host coagulation mechanism [63,64].
Finally, we analyzed the differentially expressed genes in adult males in light of the known biological characteristics of the males. Their transcriptome was already individually explored by RNA-Seq [6]; however, the different transcriptome profiles of the two sexes that reflect their distinct biological characteristics were not previously compared. We found that a number of genes most highly expressed in males are involved in the regulation of transmembrane transport (GO:0034762), membrane (GO:0016020) and heparin sulfate proteoglycan binding (GO:0043395), probably belonging to the male tegument. This result is consistent with the fact that the male body is more highly exposed to the host immune cells in the circulation than the female, and consequently, the tegument renewal rate is higher in the male than in the female [65].
Among the most highly expressed genes in males, we found an enrichment of the potassium ion transport category (GO:0006813), and the high expression of the sodium-potassium ATPase gene (Smp_124240) in males compared with in females was confirmed by qPCR. The surface plasma membrane is involved in nutrient uptake, involving several amino acid and sugar transporters, aquaporins, anion selective channels and Na+/K+ and Ca2+ ATPases [66,67].
We detected enriched GO categories among male genes associated with the schistosome muscle layer, such as the troponin complex and the calcium binding protein Sm20 gene (Smp_005350), whose expressions were higher in males than in females, according to both the RNA-Seq data and qPCR. Our RNA-Seq data showed that Smp_005350 is fully transcribed; this gene encodes a 58-kDa calcium binding protein with distinct Ca2+ binding motifs. It should be noted that Mohamed et al. studied a distinct 20.8 kDa antigenic S. mansoni calcium binding protein encoded by the U91941 cDNA clone and named Sm20.8 [68], subsequently renamed Smp_086530, which is distinct from the Sm20 gene (Smp_005350) detected here as abundant in the male.
Receptor activity and protein tyrosine kinase activity were GO categories enriched in male-expressed genes, such as Discoidin domain receptor DDR (Smp_133250), Serotonin receptor 5HTR (Smp_126730) and Wnt5 (Smp_145140), encoding a signaling molecule. This finding suggests that a characteristic cell signaling process might operate in the male.
The scavenger receptor CD36 antigen transcript (Smp_011680), which is a mediator of cholesteryl ester uptake by adult worms [51], was detected by qPCR in all three forms, with males and females having the same level of CD36 expression, in disagreement with Fitzpatrick et al., who detected the CD36 antigen transcript only in female worms [7]. The divergent result from Fitzpatrick et al. could be explained by the different parasite strain (Puerto Rican strain in the Fitzpatrick et al. study, BH strain in this study).
Micro-exon genes (MEGs) constitute a large family characterized in the parasite genome, each gene with multiple symmetrical exons, arranged in tandem, with lengths that are a multiple of three nucleotides (from 6 to 36 nucleotides for each exon) [3,69]. This arrangement leads to protein variation through alternative splicing [69], and may have an impact on the escape of parasites from host defenses. Differential expression of MEGs between males and females has only been observed with microarrays in a comparison of male esophagus with female gastrodermis [70]. Here we compared eggs, female and male adult worms and identified stage-specific MEGs, most highly expressed in each of the three forms, supporting the idea that MEGs might be specifically modulated in response to defenses of the host. Interestingly, MEG-4 and MEG-14, most highly expressed in males, have been previously described using whole mount in situ hybridization to be specifically expressed in the esophagus of adult worms from both sexes [71]. Male and female highly expressed MEG-5 had its protein product previously detected in tegumental preparations [69]. Egg highly expressed MEG-2 and MEG-3 protein products were detected in egg secretions and only the latter was also detected in schistosomula secretions, being produced in the head gland [69]. These observations suggest that either the expression of esophageal MEGs is more robust in males than in females or that in the males this organ contributes with a higher amount of mRNA for the total pool.
We used the transcriptome data as a guide to improve Smp gene predictions, adding UTRs or new coding exons to hundreds of Smp genes. Moreover, we cross-referenced the genomic coordinates of the genes with the genome-wide coordinates that we obtained by genome-mapping the publicly available H3K4me3 adult worms ChIP-Seq dataset [9], and we confirmed the recently described presence in adult schistosomes of the H3K4me3 mark at the TSSs of thousands of Smp predicted genes [72]. More importantly, we identified the presence of the H3K4me3 mark at the TSSs of 79 out of 232 putative novel protein-coding genes identified by our RNA-Seq, which provided additional evidence of the regulation of these novel putative genes by histone modification. We also found the H3K4me3 TSS histone mark for another 525 re-structured Smp gene models (out of 2,083 Smp gene models with their 5′-ends extended by our RNA-Seq data), thus providing additional confirmation that the original gene model predictions for those 525 genes were improved by using this new 5′-end information. For those genes whose predicted TSS genome positions do not intersect with H3K4me3 peaks, the distance to the closest H3K4me3 peak could indicate the most probable TSS positions [40]; for the Smp predictions that are distant from a H3K4me3 peak, we found that this distance was between 1–2 kbp upstream from the predicted TSS.
Finally, the novel protein-coding genes detected in our transcriptome data are complementary to the existing S. mansoni gene predictions and annotations, allowing the discovery of genes with possible biological relevance to the parasite. We suggest that among other genes, the SmMEIG gene should be investigated as a potential regulator of parasite sexual reproduction and egg laying. Attention could also be given to the divergent Lifeguard genes SmDLFG1 and 2, two possible inhibitor regulators of cell apoptosis. Interestingly, examining the genomes and gene predictions of other platyhelminths did not provide any evidence of close orthologs, suggesting that this protein would be specific to schistosome species among the platyhelminths. By examining the phylogenetic tree (Fig G in S1 Text), it is possible to observe that the branch that contains the two new SmDLFGs from S. mansoni contains no other protein from species outside the Schistosoma genus, while its branch length is very long. Therefore, it appears to be an isoform specific to this genus that was probably subjected to a positive selection process at some point during its evolution. A possible explanation for this phenomenon would be the co-optation of this protein for some process related to host-parasite interaction, as the new SmDLFGs are transmembrane proteins (see Fig F in S1 Text); in fact, it was previously demonstrated in the Rickettsiaceae model that genes coding for parasite membrane proteins tend to display positive selection in relation to genes for membrane proteins of non-parasites [73]. The finding that the new SmDLFGs are highly expressed in females compared with males and eggs (Fig 4) is interesting because the genes from this family are described in humans as encoding an apoptosis inhibitor protein; females are known to be more resistant to drug treatments than males, and the high expression of a possible apoptosis inhibitor protein should be considered when repurposing apoptosis-inducing cancer drugs to treat schistosomiasis [74].
Other putative new genes and gene fragments detected here are likely involved in different biological processes, such as metabolism, development, morphogenesis and cell communication, which raises the possibility of finding and confirming missing genes in the parasite’s molecular pathways.
We also confirmed in this study that thousands of lncRNAs (> 200 nt) are transcribed in S. mansoni [6,75], and many of these non-coding RNAs may possibly exert regulatory functions that have yet to be explored [75]. In fact, Guttman and co-workers [76] showed that in mammals, there is a class of conserved lncRNAs (i.e., the large intervening non-coding RNAs, lincRNAs) whose transcription and processing appear to be similar to protein-coding genes, with Pol II transcription, 5´-capping and poly-adenylation. Since then, a number of other studies have confirmed that most lncRNAs are poly-adenylated, although there are also non-poly-adenylated ones. Our RNA-Seq cDNA libraries have captured polyA+ RNAs, and the informatics analysis of the sequenced RNAs has identified thousands of long transcripts with non-coding potential, thus providing additional evidence for the existence of such poly-adenylated long noncoding transcripts in S. mansoni. Cross-reference with known transposon sequences showed that these transcripts do not originate from transcribed repeats. In fact, these lncRNAs are part of the polyA+ RNA pool of a normal eukaryotic cell, and in human cancer cells one of the best characterized consequences of an altered expression of lncRNAs is an important change in deposition of regulatory histone marks at the promoter regions of oncogenes and tumor-suppressor genes [77]. The lncRNA sub-population of a cell, which has long been overlooked, is now being studied by a wide variety of methods [78] that are revealing lncRNAs as an important layer of information that in many instances controls transcription, translation, imprinting, histone modifications, among other functions in the eukaryotic cells.
Additionally, the data discussed in this work are available in a new S. mansoni Genome Browser Database (http://schistosoma.usp.br/), which uses the UCSC Genome Browser in a Box platform [36] and provides a S. mansoni public data-searching tool (using the option “Tools > Blat” with a query sequence, or entering a gene name in the enter position or search term window), a convenient data-downloading tool (using the option “Tools > Table Browser”) and a visualization tool in a user-friendly format.
|
10.1371/journal.pcbi.1003988 | Mapping the Pareto Optimal Design Space for a Functionally Deimmunized Biotherapeutic Candidate | The immunogenicity of biotherapeutics can bottleneck development pipelines and poses a barrier to widespread clinical application. As a result, there is a growing need for improved deimmunization technologies. We have recently described algorithms that simultaneously optimize proteins for both reduced T cell epitope content and high-level function. In silico analysis of this dual objective design space reveals that there is no single global optimum with respect to protein deimmunization. Instead, mutagenic epitope deletion yields a spectrum of designs that exhibit tradeoffs between immunogenic potential and molecular function. The leading edge of this design space is the Pareto frontier, i.e. the undominated variants for which no other single design exhibits better performance in both criteria. Here, the Pareto frontier of a therapeutic enzyme has been designed, constructed, and evaluated experimentally. Various measures of protein performance were found to map a functional sequence space that correlated well with computational predictions. These results represent the first systematic and rigorous assessment of the functional penalty that must be paid for pursuing progressively more deimmunized biotherapeutic candidates. Given this capacity to rapidly assess and design for tradeoffs between protein immunogenicity and functionality, these algorithms may prove useful in augmenting, accelerating, and de-risking experimental deimmunization efforts.
| Protein therapeutics have created a revolution in disease therapy, providing improved outcomes for prevalent illnesses and conditions while at the same time yielding treatments for diseases that were previously intractable. However, this powerful class of drugs is subject to their own unique challenges and risk factors. In particular, the biological origins of therapeutic proteins predispose them towards eliciting a detrimental immune response from the patient's own body. Therefore, fully capitalizing on the medicinal reservoir of natural and engineered proteins will require efficient, effective, and broadly applicable deimmunization technologies. We have developed deimmunization algorithms that simultaneously optimize therapeutic candidates for both low immunogenicity and high-level activity and stability. Here, we combine computational modeling and experimental analysis to show that the process of protein deimmunization manifests inherent tradeoffs between immunogenic potential and biomolecular function. Our experimental results demonstrate that dual objective optimization allows us to assess and design for these tradeoffs, thereby enabling facile construction of deimmunized variants that span a broad range of immunogenicity and functionality performance parameters. Thus, we can rapidly map the design space for deimmunized drug candidates, and we can use this information to guide selection of engineered proteins that are most likely to meet performance benchmarks for a given clinical application.
| Therapeutic proteins are revolutionizing disease therapy across a broad range of indications and illnesses, and biotherapeutic sales are an increasingly important part of the pharmaceuticals market [1], [2]. However, these powerful drugs suffer from their own limitations, which, if addressed, could accelerate the pace of biotherapeutic development and approval. A relatively unique risk factor for protein therapeutics is their inherent potential to induce anti-drug immune responses in human patients [3], [4], [5]. These undesirable immune reactions can compromise drug efficacy or cause more serious adverse events [6], [7].
In a healthy human immune system, all extracellular proteins are sampled by antigen presenting cells (APCs). Once internalized by APCs, a protein is cleaved into small peptide fragments, putative immunogenic segments are loaded into the groove of class II major histocompatibility complex proteins (MHC II), and the complexes are trafficked to the APC surface. True immunogenic peptides, termed T cell epitopes, facilitate the formation of ternary MHC II-peptide-T cell receptor complexes with surface receptors of cognate CD4+ T cells [8]. This critical molecular recognition event initiates a signaling cascade that drives stimulation of helper T cells, maturation of B cells, and ultimately production of circulating antibodies that bind to and clear the foreign therapeutic protein. Detailed knowledge of this process enables protein deimmunization via mutation of key residues in immunogenic epitopes, a methodology commonly known as T cell epitope deletion.
There exist in the literature numerous examples of successful T cell epitope deletion projects. To date, the majority of these efforts have relied on time, labor, and resource intensive experimental strategies. Experimentally driven approaches entail dividing the target protein's primary sequence into a large panel of overlapping peptides, synthesizing those peptides, and using them for exhaustive epitope mapping with human peripheral blood mononuclear cells and/or purified human MHC II proteins [9], [10], [11], [12], [13], [14]. To circumvent the considerable effort and expense required for experimental epitope mapping, a wide range of epitope prediction tools may be accessed [15], [16], [17], [18], [19], [20], [21], [22], [23]. In some recent cases, such tools have been leveraged to good effect in deimmunizing therapeutic candidates [24], [25].
Identification and mapping of T cell epitopes is a relatively mature field, but selection of mutations that simultaneously delete epitopes and maintain protein function remains a challenging task. Experimentally driven deimmunization typically relies on alanine scanning or similar empirical strategies to select epitope deleting mutations. While Cantor et al. employed epitope prediction to rapidly identify immunogenic regions of asparaginase, their selection of deimmunizing yet function-preserving mutations required construction of large combinatorial protein libraries and implementation of a sophisticated ultra-high throughput screen [24]. As an alternative to the above methods, bioinformatics tools are increasingly used to filter prospective deimmunizing mutations for those least likely to disrupt protein structure and function [12], [26], however these in silico analyses of a mutation's structural and functional consequences have historically been applied post hoc. Any such sequential application of computational tools fails to consider the combined effects of all mutations on immunogenicity and function, thereby precluding a global approach to protein deimmunization. Thus, while T cell epitope deletion is a well validated methodology, the success, efficiency, and general utility of the approach would be enhanced by bringing to bear more advanced protein engineering and design technologies.
The next generation of protein deimmunization tools have seamlessly integrated immunoinformatic epitope prediction with in silico analysis of the functional consequences associated with prospective deimmunizing mutations [27], [28], [29], [30]. By packaging both design objectives in a single optimization algorithm, these technologies enable global protein design and deimmunization on a highly compressed time scale. The first two iterations of these novel algorithms, Dynamic Programming for Deimmunizing Proteins (DP2) and Integer Programming for Immunogenic Proteins (IP2), have undergone preliminary experimental validation with Enterobacter cloacae P99 Beta-lactamase (P99βL) [25], [31], a biotherapeutic that has been deimmunized previously using conventional experimentally-driven techniques [9]. Here we deimmunize the P99βL target using a more advanced extension of IP2, embodied in the protein design algorithm “Protein Engineering Pareto Frontier” (Pepfr) [32]. Whereas IP2 samples a subset of the designs that optimally balance the immunogenicity and functionality objectives, Pepfr generates the entire set of Pareto optimal variants, i.e. all enzymes whose predicted immunogenicity and functionality are not simultaneously dominated by any other single design. Eighteen of these Pareto optimal variants have been produced and subjected to a rigorous experimental analysis of key performance parameters. This combined computational and experimental analysis of increasingly aggressive plans has provided new insights into the inherent tradeoffs linking the target enzyme's sequence, function, and immunogenic potential. As a whole, this work outlines a design-based approach to functional deimmunization of biotherapeutic candidates.
Whereas previous computationally-driven deimmunization of P99βL had targeted eight common MHC II alleles [25], [31], here we optimized against only four alleles (DRB1*0101, 0401, 0701, and 1501) for which MHC II-peptide binding experiments had been fully optimized [33]. Analysis with the ProPred epitope prediction tool [20] revealed that putative immunogenic peptides were broadly distributed throughout the sequence, with several discrete regions exhibiting numerous overlapping and promiscuous epitopes, e.g. proximal to residues 14, 105, 210, 235, and 334 (Fig. 1). While our prior P99βL deimmunization efforts had focused on validation of our protein optimization algorithms, the objective of the current study was a systematic analysis of the sequence-function-immunoreactivity tradeoffs that are inherent to the deimmunization process.
In pursuit of this goal, we applied the Pepfr protein design algorithm [32] to optimize the two objective functions derived from the IP2 deimmunization formulation [28]: a sequence score (Sseq, materials and methods equation 1), capturing the predicted effects of mutations on protein function, and an epitope score (Sepi, materials and methods equation 2), capturing the predicted effects of mutations on protein immunogenicity. Both scores are defined such that lower is better (less perturbation to function and reduced immunogenicity, respectively). Pepfr identifies the “Pareto frontier” of the deimmunized design space, comprised of those designs whose sequence and epitope scores are not simultaneously dominated by any other variant (i.e., those designs making the best tradeoffs between the scores). Separate Pepfr runs were performed to identify designs at mutational loads ranging from 1 to 8.
The resulting output was a panel of 18 P99βL designs that exhibited a range of mutational loads and extents of epitope disruption. A plot of Sseq vs. Sepi for the 18 protein plans enabled visualization of the objective functions' competing nature (Fig. 2). The overarching goal was reduction of P99βL epitope score via mutagenic deletion of predicted epitopes; however each deimmunizing mutation incurs an Sseq penalty. Any increase above the wild type Sseq reflects a putative risk of reduced protein stability and/or function, and therefore mutagenic deimmunization must carefully balance the opposing objective functions. The Pareto frontier analysis (Fig. 2) highlights the relative tradeoffs between predictions of epitope content and biological activity, but the practical relationship between these mathematical functions is an unknown quantity. Thus, experimental analysis is ultimately required to understand the magnitude of biological activity that is sacrificed per unit immunogenicity.
Incrementally enhanced deimmunization, moving from right to left on the Pareto curve (Fig. 2), was realized by three complementary mechanisms. First, increasing mutational loads allowed for simultaneous disruption of multiple, distributed epitope clusters. Compare, for example, design 1I, which targets a single epitope with one mutation, to design 8Z, which targets seven distinct immunogenic regions with eight mutations (Table 1). Second, in some instances accrued mutations were combined in close proximity to better target one particularly immunogenic region. For example, designs 4M through 7S as well as plan 8U encoded the R105S mutation, which was predicted to disrupt three of seven epitopes in a dense cluster centered on position 105 (Fig. 3). The more ambitious designs 8V through 8Z deleted six of these same seven epitopes with the combined G103D+R105S double mutation. The mutational combinations M235Q+V243L and Q333D+I334L were likewise predicted to yield enhanced epitope deletion relative to their single mutation counterparts (Fig. 3). In parallel to escalating mutational loads, a third mechanism for improved epitope deletion was the use of increasingly aggressive individual mutations. In particular, mutation N14R was associated with three designs possessing moderate sequence scores (4N, 5R, and 8V; Sseq range of 17.1 to 41.4; Table 1), but it deleted only three of six epitopes in the dense cluster centered on residue 14 (Fig. 3). Mutation A13E, employed by six designs having a Sseq range of 25.3 to 107.6, disrupted five of the six epitopes in this cluster. Finally, A13D deleted all six predicted epitopes, but this aggressive substitution contributed to particularly poor overall sequence scores (Sseq = 98.8 and 144.8 for designs 4P and 8Z, respectively). In aggregate, incremental increases in mutational load and mutational stringency produced a systematic series of deimmunized designs ranging from the wild type Sepi = 60 to that of variant 8Z (Sepi = 32), in which almost half of the predicted epitopes were targeted for disruption.
Engineered gene constructs were assembled by recursive PCR from overlapping synthetic oligonucleotides, and each gene was modified with a 5′-coding sequence for the OmpA leader peptide and a 3′-coding sequence for a C-terminal hexa-histidine tag. Genes were cloned behind the strong T7 promoter of vector pET26b, and proteins were expressed in the E. coli host BL21(DE3) [F– ompT hsdSB (rB- mB-) gal dcm (DE3)]. Recombinant enzymes were released from the periplasm by osmotic shock and subsequently purified to>95% by immobilized metal affinity chromatography. Yields were 1–30 mg/liter of cell culture, depending on the enzyme variant.
The relative structural stabilities of the eighteen engineered enzymes were assessed as apparent melting temperatures (Tm), quantified by differential scanning fluorimetry [34]. The Tm's of the eighteen variants ranged from 47.09–56.27°C, or 83–99% of the wild-type value (56.61°C) (Table 1). While the observed 9.5°C range in Tm should not be interpreted as insubstantial, it bears noting that none of the engineered variants exhibited significant unfolding at 37°C (S1 Fig.), which is the temperature of ultimate therapeutic relevance.
The incremental manner in which the design series progressively targeted epitopes resulted in extensive mutational overlap between adjacent designs, and insights regarding the destabilizing effects of specific substitutions were obtained by deconvoluting the mutational composition of various constructs. Consider for example the adjacent series 1J, 2K, and 4L. Design 1J encoded only M235Q, which resulted in a negligible 0.64°C reduction in Tm (Table 1). In contrast design 2K, which encoded both M235Q and R210H, exhibited a 2.83°C drop in Tm, indicating that R210H has a significant destabilizing effect, either by itself or in the context of M235Q. The next variant, 4L, revealed that neither V25I nor T342K were destabilizing substitutions, as 4L differed from 2K by only these mutations yet exhibited an equivalent Tm. The permissible natures of V25I and T342K were further corroborated by comparison of Tm's for 5Q vs. 4M and 1I vs. WT, which differed by the respective single mutations and again possessed essentially the same Tm values.
Separately, the data indicated that substitutions N14R and A13E were interchangeable with respect to structural integrity. In particular, the 4-mutation designs 4N and 4O differed only by N14R and A13E, respectively, and the 8-mutation variants 8V and 8W exhibited the same distinguishing feature. In both cases, the two alternative substitutions yielded essentially equivalent Tm values (49.47≈49.74°C and 50.01≈50.25°C, respectively). Moreover, there was evidence that these N-terminal mutations did not further compromise designs already exhibiting moderately decreased stability. For example, 5R differed from 4M by the simple addition of N14R, yet both enzymes showed similar stability (Tm = 50.15 and 49.95°C, respectively).
The most striking observation, however, was the clear bifurcation in Tm values between designs that encoded R105S and those that did not. Without exception, plans bearing R105S possessed Tm's below 52°C (average Tm = 49.53°C), while variants bearing wild type R105 uniformly exhibited Tm's above 53°C (average Tm = 55.09°C). This suggested that R105S was the single most destabilizing mutation from the study, and separate experiments on the R105S point mutant showed that this single mutation substantially reduced protein stability (Tm = 52.64°C, S1 Table). It seems likely that this effect stems from the fact that R105 resides in the center of a pocket defined in part by D86, D87, D108, and E300 (S2 Fig.). Presumably, R105 electrostatically stabilizes the adjacent acidic residues, and removal of this positive charge by the R105S mutation renders the protein less stable. Interestingly, while the isolated R105S mutation caused a reduction in thermostability, it manifested no substantial effect on catalytic activity (S1 Table), and it provided for a net reduction in peptide interaction with human MHC II proteins (see results below). Thus, the unfavorable consequences of R105S appeared to be confined to structural stability.
Finally, it should be noted that the sequence potential was intended, in part, to quantify the likelihood that mutations or combinations of mutations would maintain P99βL structural integrity. A plot of Sseq vs. apparent Tm yielded the expected inverse relationship, and a linear regression showed that the correlation was highly significant (non-zero slope, P = 0.0019) (Fig. 4A). While the sequence potential was not an accurate predictor of individual Tm values (linear R2 = 0.44), from a global perspective it did effectively capture this aspect of experimental performance.
A second goal of the sequence potential was to select mutations least likely to disrupt P99βL activity. To assess mutational effects on molecular function, Michaelis-Menten kinetic parameters were quantified using the beta lactam substrate nitrocefin (Table 1). Linear regression of Sseq vs. turnover number (kcat) or catalytic efficiency (kcat/Km) revealed a highly significant inverse correlation (non-zero slope, P = 0.0098 and 0.0005, respectively), whereas there was not a strong correlation with Km (Fig. 4B, C, D). Similar to the relationship with Tm, the sequence potential could not be used to predict catalytic proficiency for individual enzymes, but it did accurately reflect the overall trends for measured reaction rates and catalytic efficiency.
The kcat values for individual designs ranged from 65–206% that of wild type P99βL (average over all variants = 114%). Notably, 11 of 18 variants exhibited faster than wild type maximum reaction velocities. However, the majority of variants (15/18) also experienced an increase in Km, and as a result the average kcat/Km for all variants was 88% that of wild type (ranging from 56–121%). In general, wild type or better reaction rates and catalytic efficiencies were maintained up through the two least aggressive 8-mutation plans, 8T and 8U. The two variants with the highest overall kcat/Km values were designs 5Q and 5R (109% and 121% of wild type, respectively), and in both cases these enhanced efficiencies were driven exclusively by substantial increases in the kcat parameter (206% and 164% wild type, respectively). Together, these observations highlight the fact that the wild type sequence does not represent a global optimum with respect to catalytic conversion of nitrocefin. The functional deimmunization process identified numerous performance enhanced variants, and the high activity observed across the full spectrum of mutational loads underscores the practical utility of the statistical sequence potential. Indeed, even the five most aggressive 8-mutation designs (8V through 8Z) proved to be highly active enzymes, with kcat and kcat/Km values that averaged 77% and 71%, respectively, of wild type. The single most deimmunized variant, 8Z, maintained well above 50% wild type rate acceleration and efficiency. Thus, all 18 deimmunized enzymes exhibited activity comparable to naturally evolved biocatalysts.
The immunoreactivity of various constructs was assessed by measuring the MHC II binding affinity of their corresponding peptide fragments. These competition immunoassays are a widely recognized metric for assessing immunogenic potential and validating computational predictions [22], [25], [33], [35], [36], [37], [38], [39]. Synthetic fragments of wild type P99βL were designed so as to encompass each of the epitopes targeted by the deimmunization algorithm, and corresponding variant peptides were synthesized to represent the deimmunized designs (S2 Table). The affinity of each peptide for human MHC II molecules DRB1*0101, 0401, 0701, and 1501 was measured by competition with known peptide immunogens for each allele. A quantitative comparison of wild type versus variant MHC II binding affinity was used as a proxy measure for the success of epitope deletion (Fig. 5). Peptide affinities are reported as IC50 values, and putative epitopes were classified as strong (IC50<1 µM), moderate (1 µM≤IC50<10 µM), weak (10 µM≤IC50<100 µM), or non-binders (IC50≥100 µM).
High affinity interaction between peptide antigens and class II MHC is a key determinant of subsequent T cell immunogenicity [40], [41], [42], and a total of four wild type P99βL peptides were found to possess sub-micromolar IC50's for one or more of the tested alleles. The wild type A13+N14 peptide was a high affinity binder of DRB1*0701 (IC50 = 800 nM), and both A13D and A13E successfully converted this to a weak binding interaction with N14R yielding a moderate binding interaction (Fig. 6). As found in prior studies [25], wild type peptide L149 was bound by all four alleles, and here it was a particularly strong binder of 1501 (IC50 = 300 nM). The L149Q mutation reduced 1501 affinity by 40-fold, converting this strong binding interaction to a weak interaction. Wild type peptide I262 also bound all four alleles, and it possessed sub-micromolar affinity for both 0401 and 1501. The I262V mutation yielded a 6-fold reduction in 1501 affinity, thereby converting a strong binder to a moderate binder. In contrast, I262V did not substantially alter affinity for allele 0401, although this outcome was predicted during the design process (Fig. 3). The only other high affinity binding of a wild type peptide was 0101 binding of I48, which, contrary to predictions, was unaffected by the I48V mutation.
A total of 29 peptides, 11 wild type and 18 engineered, were analyzed to produce 116 affinity measurements. Of the 72 pairs of wild type and cognate deimmunized affinities (Fig. 5), there were 16 cases in which the designed mutation reduced MHC II affinity by more than an order of magnitude. There were an additional 11 instances wherein the designed mutation reduced affinity by 5- to 10-fold, and 10 examples of more modest 2- to 5-fold reductions. In aggregate the engineered mutations showed a 37.5% success rate in reducing MHC II binding by 5-fold or more. In contrast, there were only nine total instances in which the designed mutation enhanced MHC II affinity by any measurable degree. In five of those cases, the increase was a modest 2- to 5-fold, and there were no quantified examples of 10-fold or greater increases in affinity.
To correlate the experimentally measured MHC II affinities with the algorithm's binary prediction of peptide binding/non-binding, a threshold value for experimental “binding” was arbitrarily selected. So as to maintain consistency with our prior work on P99βL, we set the cutoff for experimental binders at an IC50<100 µM, i.e. counting all strong, moderate, and weak binders as defined above. Given this experimental threshold and a ProPred prediction threshold of 5%, the protein design process yielded a 65% positive prediction rate for binders across all four alleles (Fig. 6). Predictions were most accurate for DRB1*0401 (76%) and least accurate for allele 0701 (59%). Overall, we observed a 13% false positive rate and a 22% false negative rate, similar to those we have reported previously [25], [31]. Comparable analyses using the newer IEDB consensus [22] and NNAlign [34] prediction methods revealed that, in this instance, no single predictor exhibited dominant accuracy across all four alleles (S3, S4 and S5 Tables). In particular, the ProPred predictor was comparable to the others for the peptides assessed here. As a whole the results show that the IP2 deimmunization formulation, implemented through the Pepfr protein design algorithm and using the ProPred epitope predictor, proved to be proficient at identifying high affinity MHC-binding peptides and selecting corresponding disruptive mutations.
To enable comparison of whole protein immunoreactivity, MHC II binding data for individual peptides was integrated across the full length of each enzyme design. For each protein, a categorical immunoreactivity score was obtained by summing the number of strong, moderate, and weak MHC binders across the protein's component peptides (11 peptides • 4 MHC alleles = 44 possible interactions). Consistent with the predicted Sseq epitope parameter (Table 1), the design series showed a general trend of decreasing experimental immunoreactivity moving from variant 1I to 8Z (Fig. 7). Of the 18 engineered designs, 11 had a net deletion of one or more high affinity interactions, 17 deleted one or more moderate affinity interactions, and 16 deleted one or more weak interactions. Importantly, none of the engineered enzymes suffered a net increase in total experimental epitopes. In only one case was a design found to have a net addition of epitopes in any single binding category. Namely, 4P possessed one additional weak binder, but at the same time it deleted four moderate and two strong binders, the latter two being most prone to drive a T cell mediated immune response [40]. Indeed, with respect to deleting moderate and strong binders, design 4P was bested only by 8Z. The latter was the single most aggressive design, and it was in fact found to have the lowest categorical immunoreactivity. Specifically, 8Z yielded a net deletion of three strong binders, four moderate binders, and one weak binder, thereby eliminating a full quarter of all experimentally identified MHC II binders. Considering only the higher affinity peptides (IC50<10 µM), 8Z benefitted from a 39% reduction in epitope content, similar to the 47% reduction predicted by the deimmunization algorithm. Thus, prediction of epitope disruption was borne out in the overall experimental analysis.
As a second measure of whole-protein immunogenic potential, a global quantitative immunoreactivity value was calculated by averaging the numerical IC50's for each enzyme's component peptides. Importantly, the dynamic range for our MHC II binding assay is 10 nM to 250 µM, and many of the binding affinities, particularly for engineered peptides, were found to be too weak for precise quantitation (values>250 µM, Fig. 6). Because these non-binding peptides are key indicators of reduced immunoreactivity, it was critical that they be factored into the quantitative, whole-protein score. To do so, we employed equation 4 (see materials and methods). Each enzyme's global immunoreactivity, normalized to 100% for wild type P99βL, is reported in Table 1. Similar to the categorical analysis, there was a general trend towards decreased global immunoreactivity with increasing mutational load and aggressiveness. On this normalized scale, designs 8T and 8U are the most immunotolerant variants, both exhibiting a 65% reduction relative to wild type P99βL immunoreactivity. The most extensively engineered design, 8Z, is also highly immunoevasive, with a 63% reduction compared to wild type. Overall, the global, quantitative immunoreactivity was found to have a highly significant and surprisingly close correlation with the predicted Sepi parameter (linear R2 = 0.64; non-zero slope P<0.0001; Fig.4E). Thus, Sepi offered reasonable predictive power even for individual P99βL designs. In total, the algorithm successfully incorporated compatible and increasingly effective deimmunizing mutations so as to achieve a systematic reduction in immunogenic potential.
Mutagenic deletion of T cell epitopes, which has been successfully implemented with diverse proteins, is a powerful means for deimmunizing biotherapeutics. With very few exceptions, however, published studies of T cell epitope deletion, in full length proteins, have focused on disrupting one or two immunogenic regions [9], [10], [11], [12], [14], [25], [43]. Indeed, there is debate regarding the feasibility of broad, protein-wide epitope deletion, which is complicated by the high degree of MHC II polymorphism in human populations [44]. Thus, while there are many reports of limited but successful T cell epitope deletion, one is left to wonder how many projects might have failed due to the presence of numerous and dispersed immunogenic regions that could not be targeted simultaneously using conventional strategies. To more fully understand this challenge, we have conducted a combined computational and experimental analysis of the immunogenicity and functionality tradeoffs that are inherent to the deimmunization problem.
The studies described here were enabled by an advanced deimmunization algorithm that seamlessly integrates immunogenic epitope prediction with in silico analysis of the functional consequences associated with deimmunizing mutations. We combined the IP2 deimmunization formulation with the Pepfr optimization algorithm [28], [32] to design a suite of 18 Pareto optimal P99βL variants. Each of these designs optimally balances two objective functions – one modeling immunogenicity and the other functionality – such that no other single variant is predicted to outperform with respect to both design objectives. Inspection of the Pareto optimal designs reveals that, in the context of the mathematical model, there is an inverse relationship wherein ever greater deimmunization is achieved at the expense of progressively reduced function (Fig. 8A). To assess the practical implications of these predicted tradeoffs, we have recombinantly produced all 18 designed enzymes and rigorously characterized their stability, activity, and immunoreactivity with human MHC II proteins. The results of this analysis represent the first systematic assessment of the functional penalty that is paid for pursuing progressively more deimmunized drug candidates.
Our previous work had demonstrated the capacity to design P99βL variants bearing 1–5 deimmunizing mutations yet retaining wild type or better activity and near wild type stability [25], [31]. By more thoroughly mapping the Pareto optimal design space, we show here that up to seven immunogenic regions can be simultaneously targeted while incurring essentially no loss in molecular stability and function (see design 8T, Table 1). Considering only the most aggressive designs (8T to 8Z), the computational Pareto curve suggested that there would be an accelerating loss of molecular function throughout the series, ultimately resulting in a 6-fold deterioration relative to 8T (Fig. 8A). The experimental analysis verified the predicted trend, revealing that the 8-mutation series did in fact exhibit a systematic reduction in stability and catalytic efficiency. Importantly, however, design 8Z showed a mere 40% reduction in catalytic proficiency relative to 8T. This dramatic difference in the magnitude of ΔSseq versus measured change in molecular function suggests a non-linear relationship between the statistical sequence potential and actual experimental performance. Indeed, a non-linear correlation is suggested by the above graphical analysis (Fig. 4A, C, D). Thus, it seems likely that more aggressive designs exhibiting even higher Sseq penalties might be realized experimentally before reaching the point of diminishing returns. In other words, it appears we have yet to reach the practical limit of epitope depletion for P99βL.
While the quantified losses in activity and stability were not as sharp as predicted by the Sseq design parameter, there was in fact a general trend towards escalating loss of function with more aggressive deimmunization. To better visualize these real world tradeoffs, we constructed the experimental analog of the Pareto curve (Fig. 8B). This analysis entailed plotting an integrated experimental performance score (averaging the normalized, reciprocal values for Tm, kcat, and kcat/Km; equation 3) vs. the quantitative global immunoreactivity score (equation 4). The graphical analysis clearly shows that more aggressively deimmunized enzymes sustained progressively greater losses of molecular function. Analogous plots were constructed for various individual performance parameters as well as alternative combinations of these parameters (Fig. 8C-H). It is interesting that the computationally generated Pareto plot most effectively captures the general trends observed with integrated, as opposed to individual, experimental performance measures (compare Fig. 8 panels B, G and H to panels C, D, E and F). This is a notable and advantageous outcome, as biotherapeutic researchers will typically be interested in overall molecular performance as opposed to any single metric (e.g. thermostability, binding affinity, rate acceleration, or catalytic efficiency). As a whole, the parallels between the computational and experimental Pareto plots are striking, and this observation underscores the Pepfr algorithm's capacity to effectively factor in the inherent tradeoffs between immunogenicity and molecular function.
As a final note, previous engineering of P99βL with the IP2 algorithm had produced higher activity variants than those designed with the earlier DP2 algorithm [31]. However, the IP2 designs from the former study were generated after locking down all residues in close proximity the active site. This raised the question of whether or not the combined 1-body + 2-body sequence potential of IP2 was in fact more effective than the 1-body potential implemented in DP2, where mutations to active site residues had been allowed [25]. The Pareto optimal designs from the present study did not benefit from locked active site residues, yet the 4-mutation and 5-mutation designs from this study substantially outperformed previous 2-mutation DP2 designs and were largely equivalent to previous 4-mutation and 5-mutation IP2 designs in which the active site had been held invariant (S3 Fig.). This result shows that the more advanced sequence potential of IP2 can in fact generate highly mutated and yet highly active proteins in the absence of detailed structure-function information. Moreover, when residues need not be locked down during the design process, there is greater inherent capacity for epitope deletion.
In conclusion, we have computationally and experimentally mapped the deimmunized Pareto frontier of P99βL. The predictions underlying the design process correlated well with experimental analyses of protein function. In particular, we observed that incremental deletion of progressively more T cell epitopes lead to a relative escalation in concomitant loss of function. Thus, the predicted tradeoffs underlying protein deimmunization were borne out in real world analyses. Nonetheless, all 18 of the computationally designed enzymes proved to exhibit reasonable thermostability and impressive activity; not a single design failed to express or function. The most highly engineered enzyme, which incorporated eight mutations targeting seven distinct epitope clusters, was found to have a 39% reduction in high affinity MHC II binding interactions while maintaining well over 50% of the wild type enzyme's catalytic activity. It is therefore evident that we have additional capacity for designing even more extensively deimmunized yet functional P99βL variants. If these trends translate to other therapeutic proteins, as anticipated, the integrated design algorithms evaluated here will accelerate identification of engineered variants spanning a broad spectrum of immunogenic potential and biological function. These panels of deimmunized proteins should prove a rich resource from which to select therapeutic candidates that meet diverse clinical needs.
Oligonucleotides for sequencing and standard PCR methods (25 nmol scale, standard desalting) and oligonucleotides for gene synthesis (100 nmol scale, PAGE Purified) were purchased from Integrated DNA Technology (San Diego, CA). Nitrocefin was purchased from Oxoid (Cambridge, UK). Human lysozyme and SYPRO Orange 5000× Protein Stain were purchased from Sigma (St. Louis, MO). MicroAmp Fast Optical 0.1 ml 96-Well Plates and MicroAmp Optical Adhesive Film were from Applied Biosystems (Bedford, MA). Restriction enzymes and PCR reagents were purchased from New England BioLabs (Ipswich, MA). Growth media was purchased from Becton Dickinson (Franklin Lakes, NJ). Plasmid purification kits and Ni-NTA resin were purchased from Qiagen (Valencia, CA). PCR cleanup and gel extraction kits were from Zymo Research (Irvine, CA). Peptides derived from P99βL were ordered from GenScript (Piscataway, NJ), and were greater than 85% pure. Biotinylated tracer peptides were purchased from 21st Century Biochemicals (Marlborough, MA). MHC II DR molecules were purchased from Benaroya Research Institute (Seattle, WA), anti-MHC II-DR antibody from Biolegend (San Diego, CA), and DELFIA Eu-labeled Streptavidin was from PerkinElmer (Boston, MA). Unless noted, all other chemicals and reagents were from VWR (Radnor, PA).
Functionally permissible mutations were identified using an IP2 sequence potential, generated essentially as described [28]. Briefly, a multiple sequence alignment (MSA) of 94 homologs from Pfam 00144, including the wild type, was constructed by filtering for ≥30% sequence identity to wild type, ≤90% sequence identity to each other, and ≤25% gaps. The negative log frequency of each amino acid a at each position i was used to compute position-specific one-body terms φi(a). Allowed substitutions were constrained to those appearing at or above background amino acid frequencies [45]. Two-body terms φi,j (a,b) for pairs of amino acids (a,b) at coupled positions (i,j) were computed as the negative log amino acid frequency of the pair, minus the corresponding one-body terms, which avoids double counting. Only pairs of positions with significant coupling according to a χ2-based test were included in the sequence potential. In addition to mutational constraints based on the evolutionary sequence record, prolines and cysteines were neither mutated out of nor substituted into the engineered enzyme variants.
The impact of functionally acceptable mutations on putative T cell epitope content was analyzed with the ProPred epitope predictor set to a 5% threshold. ProPred has been shown to be one of the most accurate MHC II prediction tools in the public space, and readers are referred to the following references for a detailed comparison of different methods [22], [46]. The analysis considered MHC II alleles DRB1*0101, 0401, 0701, and 1501, which are common alleles [42] and for which binding assays had been fully optimized [33]. Each nonamer peptide X considered in the optimization (i.e., incorporating a contiguous combination of wild type residues and allowed substitutions) was classified as either a binder or non-binder of the four target MHC II alleles. The number of binders was summed to generate the nonamer's epitope score e(X). As a comparison, putative epitopes from ProPred predictions were subsequently analyzed using the IEDB consensus [22] and NNAlign [34] prediction methods with binding cutoffs (IEDB 5% or 10%; NNAlign 50 nM or 1000 nM) set to values commonly used in protein immunogenicity prediction [47]. Ultimately, we found that the alternative epitope predictors, when applied to the ProPred based designs, yielded the expected trends of immunogenicity-functionality tradeoffs across the set of 18 P99βL variants (S4 Fig.).
Given a wild-type sequence and a mutational load, Pepfr identifies each Pareto optimal variant s with the specified number of mutations making undominated tradeoffs between the competing objectives of total sequence potential Sseq and total epitope score Sepi:
(1)(2)where bracketed expressions indicate selection of the amino acid at the position or the substring of amino acids at the contiguous positions.
Briefly, Pepfr identifies the Pareto frontier of this two-objective space (Sepi vs. Sseq; see Fig. 2) by employing a divide-and-conquer algorithm wrapped around an IP2-based variant optimization. Given a region in the objective space (min/max values for the objectives), Pepfr uses a constrained version of IP2 to optimize an undominated design in that region. Pepfr then further divides the region into four quadrants around the design and recurses only for the upper-left and lower-right quadrants, as the upper-right is dominated and the lower-left is empty. It thereby finds all and only the Pareto optimal designs, and does so efficiently in that the number of calls to the integer programming optimizer is proportional to the number of Pareto optimal designs. We used Pepfr as described for deimmunization (He et al. 2012), with the underlying integer programming instances optimized by calls to the IBM CPLEX package.
Gene synthesis was performed using a two-step process. First, an assembly reaction was performed using fifty-two synthetic oligonucleotides encoding each design with an appended 5′- ompA leader sequence and 3′ hexa-His coding sequence (sequence GGGSAETVEHHHHHH). The assembled genes were then amplified in a second PCR using external primers. The constructs were then digested using Xba1 and HindIII, ligated into similarly digested pET26b, and electroporated into BL21(DE) E. coli cells [F– ompT hsdSB (rB- mB-) gal dcm (DE3)].
Expression was performed in 200–500 ml of LB medium containing 30 µg/ml kanamycin (LB-Kan). Expression cultures, from a 1∶100 subculture of saturated overnight cultures, were grown with aeration at 37°C in 2 L baffled flasks for an hour and forty five minutes. The temperature was then shifted to 16°C, equilibrated for 15 minutes, and expression was induced with 1 mM IPTG. Following 12–20 hours of induction at 16°C, osmotic shocktates were prepared using the protocol described in the Epicentre PeriPreps Periplasting Kit with slight modifications. Briefly, cells were pelleted at 6000g for 10 minutes and resuspended in PeriPreps Periplasting Buffer containing 1.5 µg/ml human lysozyme. Cells were quenched after a five minute incubation period with ice-cold water, and then incubated on ice for 10 minutes. The periplasmic fraction was collected by spinning the shocktate at 14,000g for 10 minutes and collecting the supernatant.
Proteins were purified from clarified periplasmic fraction using Ni-NTA resin (400 µl bed volume). After the clarified periplasmic fraction was flowed through the resin by gravity, the column was washed 2 times with 1 mL of PBS (137 mM NaCl, 2.6 mM KCl, 10 mM Na2HPO4, 1.7 mM KH2PO4, pH 7.4) containing 20 mM imidazole, and the enzyme was eluted with 2 ml of PBS containing 200 mM imidazole. The elution fraction was either dialyzed (10,000 MW cutoff) against 3 changes of 4 L PBS or concentrated and buffer exchanged by centrifugation (10,000 MW cutoff) against 3 washes of 15 mL PBS to a final concentration of 0.5–2 mg/ml protein. Purified protein was stored at 4°C prior to further analysis. All protein preparations were>95% pure, as determined by reverse-phase HPLC analysis (Agilent 1200 Series HPLC) on a Vydac 214TP 180mm C4 column, eluted at 65°C with a gradient of [90% acetonitrile/9.9% water/0.1% trifluoroacetic acid] in [99.9% water/0.1% trifluoroacetic acid] at a flow rate of 1 ml/min.
Nitrocefin substrate stock was prepared immediately prior to the experiments by dissolving nitrocefin powder in DMSO to a concentration of 20 mM. Triplicate assays were run in 96-well plate format at 30°C measuring absorbance at 490 nm (Molecular Devices SpectraMax 190 plate reader). Absorbance measurements were converted to micromolar product concentrations using the appropriate molar absorptivity (εM = 20,500 M−1 cm−1). The assay buffer was PBS, and each well contained a final enzyme concentration of 50 ng/µl, 0.04% BSA, and nitrocefin at concentrations ranging from 10 µM to 200 µM. Initial reaction rates were plotted against substrate concentration, and Michaelis-Menten kinetic parameters were determined by nonlinear regression using GraphPad Prism v.5 software (La Jolla, CA). Measurements were made in triplicate, and enzymes were purified and assayed in biological duplicate.
Differential scanning fluorimetry was performed essentially as described (Niesen, Berglund et al. 2007) using an ABI 7500 Fast Real-Time PCR System from Applied Biosystems (Bedford, MA). Proteins and SYPRO Orange were diluted in PBS. Final protein concentrations were 100 µg/ml and final dye concentrations were 5×. Twenty µl reactions were performed in 12 replicates. The PCR gradient was run from 25–94°C with a 1 minute equilibration at each degree centigrade. Fluorescence was quantified using the preset TAMRA parameters. Melting temperatures were determined by data analysis with the “DSF Analysis v3.0.xlsx” Excel sheet (ftp://ftp.sgc.ox.ac.uk/pub/biophysics/) and GraphPad Prism v.5 software.
To construct the experimental equivalent of the Pareto optimal plot (Fig. 8), a global molecular performance score was calculated for each individual enzyme using equation 3:(3)
where “” is the normalized reciprocal Tm value, “” is the normalized reciprocal kcat value, and “” is the normalized reciprocal kcat/Km value. This integrated performance score effectively averages the normalized reciprocal values for thermal stability, rate acceleration, and catalytic efficiency, yielding the experimental analog of Sseq.
MHC II competition binding assays were performed as described [33]. All data were fit to the one-site log(IC50) model by non-linear regression in GraphPad Prism v.5 software. Global immunoreactivity values were computed for each variant by (i) averaging the IC50 values for all component peptides, (ii) multiplying this figure by the number of peptides with IC50>250 µM, (iii) taking the reciprocal of the resulting product, and (iv) taking the ratio of this computed value for a variant to that of wild type P99βL. See equation 4:(4)where “” is the mean IC50 value averaged over all component peptides having affinities <250 µM, “#Nonbinders” is the total count of component peptides with affinities ≥250 µM, and the subscripts “mut” and “wt” indicate calculations for mutant and wild type proteins, respectively. This calculation accounted for both the affinity of any quantified binders and the equally important metric of total count for non-binders.
Linear regressions of experimental performance vs. computational predictions employed an F test for statistical significance of non-zero slopes. Significance was determined at the α = 0.05 level.
|
10.1371/journal.ppat.1003176 | Isolation of a Novel Swine Influenza Virus from Oklahoma in 2011 Which Is Distantly Related to Human Influenza C Viruses | Of the Orthomyxoviridae family of viruses, only influenza A viruses are thought to exist as multiple subtypes and has non-human maintenance hosts. In April 2011, nasal swabs were collected for virus isolation from pigs exhibiting influenza-like illness. Subsequent electron microscopic, biochemical, and genetic studies identified an orthomyxovirus with seven RNA segments exhibiting approximately 50% overall amino acid identity to human influenza C virus. Based on its genetic organizational similarities to influenza C viruses this virus has been provisionally designated C/Oklahoma/1334/2011 (C/OK). Phylogenetic analysis of the predicted viral proteins found that the divergence between C/OK and human influenza C viruses was similar to that observed between influenza A and B viruses. No cross reactivity was observed between C/OK and human influenza C viruses using hemagglutination inhibition (HI) assays. Additionally, screening of pig and human serum samples found that 9.5% and 1.3%, respectively, of individuals had measurable HI antibody titers to C/OK virus. C/OK virus was able to infect both ferrets and pigs and transmit to naive animals by direct contact. Cell culture studies showed that C/OK virus displayed a broader cellular tropism than a human influenza C virus. The observed difference in cellular tropism was further supported by structural analysis showing that hemagglutinin esterase (HE) proteins between two viruses have conserved enzymatic but divergent receptor-binding sites. These results suggest that C/OK virus represents a new subtype of influenza C viruses that currently circulates in pigs that has not been recognized previously. The presence of multiple subtypes of co-circulating influenza C viruses raises the possibility of reassortment and antigenic shift as mechanisms of influenza C virus evolution.
| Influenza C viruses infect most humans during childhood. Unlike influenza A viruses, influenza C viruses exhibit little genetic variability and evolve at a comparably slower rate. Influenza A viruses exist as multiple subtypes and cause disease in numerous mammals. In contrast, influenza C viruses are comprised of a single subtype in its primary human host. Here we characterize a novel swine influenza virus, C/swine/Oklahoma/1334/2011 (C/OK), having only modest genetic similarity to human influenza C viruses. No cross-reaction was observed between C/OK and human influenza C viruses. Antibodies that cross react with C/OK were identified in a significant number of swine but not human sera samples, suggesting that C/OK circulates in pigs. Additionally, we show that C/OK is capable of infecting and transmitting by direct contact in both pigs and ferrets. These results suggest that C/OK represents a new subtype of influenza C viruses. This is significant, as co-circulation of multiple subtypes of influenza allows for rapid viral evolution through antigenic shift, a property previously only shown for influenza A viruses. The ability of C/OK to infect ferrets along with the absence of antibodies to C/OK in humans, suggests that such viruses may become a potential threat to human health.
| Influenza A, B and C viruses are members of the Orthomyxoviridae family that can cause influenza in humans [1]. Influenza A viruses exist in humans, various other mammal species, and birds; migratory or domestic waterfowl are their largest reservoir. Humans are thought to be the primary hosts and reservoir of influenza B and C viruses, although both have been identified in other hosts after reverse zoonotic transmission from humans. While influenza B virus is a common seasonal human pathogen similar to influenza A virus in its clinical presentation, influenza C virus causes primarily upper respiratory tract infections in children [2]. Clinical manifestations (cough, fever, and malaise) are typically mild, but infants are susceptible to serious lower respiratory tract infections [3]. Influenza C viruses co-circulate with influenza A and B viruses and causes local epidemics [4], [5]. Six genetic and antigenic lineages of influenza C viruses have been described, and as in influenza B viruses, are considered monsubtypic [6], [7]. Co-circulation of multiple subtypes of influenza allows for rapid viral evolution through the process of antigenic shift, a property previously only shown for influenza A viruses. Thus, both influenza B and C viruses do not have pandemic potential. In contrast, the Influenza A genus includes 17 hemagglutinin and 9 neuraminidase subtypes, and reassortment among different subtypes has repeatedly generated pandemic viruses to which the human population is naïve [8]–[10]. It is the animal reservoirs of diverse influenza A viruses that give them the unique property within orthomyxoviruses of causing human pandemics.
Aside from humans, influenza C virus has been isolated only from swine in China (in 1981) [11]. Genetic analysis showed a close relation between Japanese human and Chinese swine influenza C isolates [12], [13]. Serological surveys in Japan and the United Kingdom found 9.9% and 19% of swine, respectively, to have positive HI antibody titers to human influenza C viruses, suggesting that the virus is not uncommon in swine [14], [15]. Swine inoculated with influenza C virus had mild respiratory disease and transmitted the virus to naive swine by direct contact [11]. Here we characterize an orthomyxovirus isolated from a clinically ill pig and show that the virus is distantly related to human influenza C virus and readily infects and is transmissible in both ferrets and pigs. Genetic and antigenic analysis suggest that this virus represents a new subtype of influenza C virus, raising the possibility of reassortment and antigenic shift as mechanisms for influenza C virus evolution which could pose a potential threat to human health.
In April 2011, nasal swabs from 15-week old swine exhibiting influenza-like illness were submitted to Newport Laboratories, Worthington, Minnesota, for virus isolation. Real-time reverse transcription PCR (rt-RT-PCR) was negative for influenza A virus [16]. In swine testicle (ST) cells, the viruses caused influenza-like cytopathic effects (CPE) by day 3. The cell culture harvests were again negative for influenza A virus by rt-RT-PCR. Electron microscopic (EM) studies of the cell cultures demonstrated features characteristic of an Orthomyxovirus (Fig. 1). Negative-staining EM showed enveloped spherical to pleomorphic viral particles approximately 100–120 nm in diameter (Fig. 1A). The virion surface contained dense projections 10–13 nm in length and 4–6 nm in diameter. Thin-section EM studies of infected cells revealed filamentous budding of virions from the plasma membrane (Fig. 1B). These data strongly suggested the virus to be a member of the family Orthomyxoviridae. Enzymatic assays revealed that the virus had negligible neuraminidase but detectable O-acetylesterase activity using 4-nitrophenyl acetate, suggesting it to be a member of the influenza C genus. However, further RT-PCR analysis was negative for influenza B and C viruses [17]. RT-PCR or PCR assays to detect porcine reproductive and respiratory syndrome virus, porcine coronavirus, and porcine circovirus were also negative (data not shown).
The virus was purified by ultracentrifugation and sequenced on an Ion Torrent Personal Genome Machine. De novo genome assembly found that most of the sequence reads mapped to seven contigs of approximately 1000–2400 bp. Open reading frame (ORF) analysis of the contigs found a single ORF for all segments, with the exception of two ORFs for the smallest contig. BlastP searches of the putative proteins identified modest homology to human influenza C virus, suggesting that this virus was distantly related to human influenza C virus (Table 1). Consequently, the virus was provisionally designated C/swine/Oklahoma/1334/2011 (C/OK). The genomic coding sequences of all segments were determined and used for subsequent genetic and phylogenetic analyses.
Because PB1 is reported to be the most conserved influenza virus protein, it is frequently used to evaluate the evolutionary relationship among influenza viruses [18]. We firstly performed ClustalW alignment of predicted polymerase basic 1 (PB1) amino acid sequences of influenza A, B, and C viruses. C/OK shared approximately 69%–72% mean pairwise identity to influenza C viruses and 39%–41% identity to influenza A and B viruses. Homology between the influenza A and B PB1 proteins was approximately 61% and intrasubtype PB1 proteins of influenza A viruses are extremely conserved reaching up to 90% homology. PB1 protein alignments indicated that C/OK was more closely related to influenza C viruses than to influenza A and B viruses but more distant from individual members of human influenza C type. Pairwise identity between C/OK and influenza C viruses was considerably lower for polymerase basic 2 (PB2) and polymerase 3 (P3) (53% and 50%, respectively). In influenza A and B viruses segment 3 is referred to as polymerase acidic (PA) protein because of its pKa of approximately 5.2. Conversely, segment 3 of influenza C viruses encodes a polymerase with a neutral pH (pKa ∼7.2) and is referred to as P3 [18]. Interestingly, the predicted pKa of C/OK P3 is 6.2, which is between those of the influenza A/B and influenza C viruses.
In influenza C virus, a hemagglutinin esterase (HE) protein is responsible for receptor binding, receptor destroying (acetylesterase), and membrane fusion activities, whereas in influenza A and B viruses, separate hemagglutinin (HA) and neuraminidase (NA) proteins perform these functions in a cooperative fashion. The pairwise sequence identity of the C/OK and human influenza C virus HE proteins was 53%, similar to the 49% observed across the influenza A HA subtypes [10] but clearly higher than the HA homology (approximately 25–30%) between influenza A and B viruses. NS1 of C/OK virus had the lowest homology to its counterpart in human influenza C viruses (29%–33% identity), similar to the less conserved influenza A and B NS1 proteins (22% identity).
Like PB1, the nucleoprotein (NP) and matrix (M) proteins are highly conserved among members of each genus of influenza viruses. Despite the high intragenic homology (>85%), NP and M1 are highly variable among the three influenza virus genera and their intergenic homologies are only about 20–30%, which serve as genus-specific antigens that distinguish between the influenza A, B, and C viruses [19], [20]. The amino acid sequence of the C/OK NP had 38%–41% identity to influenza C viruses. Unspliced mRNA from the C/OK M segment 6 encodes the polyprotein P42, which is cleaved by a signal peptidase to yield M1′ and CM2 [21]. P42 was analyzed due to unknown mRNA splice and protein cleavage sites used by C/OK virus to generate M1 and CM2, respectively. The C/OK P42 had 38% identity to influenza C viruses. Relative low homologies of NP and M proteins between C/OK and human influenza C viruses are interesting but seem to be consistent with pairwise protein homology analysis for polymerase and non-structural proteins.
In addition to the coding region, each RNA segment of influenza viruses also contains noncoding (NC) regions at its 5′ and 3′ ends. These NC regions are highly conserved, particularly those at the terminal ends, among the genome segments of each species. These regions form panhandle structures by partial inverted complementarity between the 5′ and 3′NC regions and play a critical role in genome replication and packaging [22]–[25]. Using 5′ and 3′ RACE coupled with direct PCR sequencing by the Sanger method, we determined the complete sequences of the 3′ and 5′ NC regions of the seven segments of the C/OK virus (Table S1). The 3′ and 5′ NC region sequences of C/OK genome segment were similar to those of human influenza C viruses with the exception of one nucleotide (position 5 from the 3′-terminus) and polymorphism at position 1 of the 3′terminus.
Viral RNA packaging sequences are composed of the 5′ and 3′NC regions and the terminal coding sequences of each segment. Incompatibility between homologous segment packaging sequences has been shown to prevent segment reassortment [26]. Nearly identical NC sequences at the proximal ends of seven RNA segments observed between C/OK and human influenza C viruses suggest a potential for viral segment reassortment in nature. Significant variability was observed in the NC regions immediately adjacent to each coding region for C/OK as compared to human influenza C; however previous work demonstrated that the highly conserved NC region at the proximal ends of the segment plays a key role in transcription and replication [27].
Our phylogenetic analysis used representative influenza A, B, and C viruses (Fig. 2). The segments encoding the C/OK virus PB2, PB1, P3, NP, M and NS clustered most closely with influenza C viruses, suggesting that these C/OK genes diverged from known human influenza C viruses after they diverged from influenza A and B viruses but before they diverged from previously sequenced influenza C viruses. As HE does not occur in influenza A and B viruses, only influenza C viruses were included in that analysis. Previous studies have found that multiple genetically and antigenically distinct but related lineages of influenza C virus co-circulate and frequently reassort [6], [28]–[30]. Given this evidence, it is puzzling that the seven segments of C/OK are only slightly to moderately homologous to characterized influenza C viruses.
An HI assay was performed to determine the antigenic cross-reactivity and seroprevalence of C/OK virus in humans and swine. The assay included reference strains of influenza A, B, and C genera and their matched antisera (Table S2). No cross-reactivity was observed between C/OK virus and heterologous antisera. For the human cohort, we used a set of 316 serum samples.
These sera originated from patients recruited in the Greater Vancouver area of British Columbia, Canada, or in the vicinity of the Greater Hartford area of Connecticut during the 2007–2008 and 2008–2009 influenza seasons as described in Marcelin et al. [31]. All but four of these sera had undetectable C/OK HI titers (≤10). Three had HI titers of 20, but each of these also had high titers (160, 320, and 1280) to the human influenza C isolate C/Yamagata/10/1981. The remaining positive sample had a HI titer to C/OK of 40 with no corresponding titer to the human influenza C isolates tested. The low titers and number of positive samples (1.3%) obtained are inconclusive in determining circulation of C/OK in the human population, especially as thirty-four percent of the serum samples had HI titers ≥20 to C/Yamagata/10/1981 which is consistent with previous studies showing approximately 60% of elderly humans retain influenza C virus antibody titers [32]. Swine serum samples (n = 220) submitted to Newport Laboratories for unrelated diagnostic testing by commercial swine production facilities nationwide were similarly analyzed. Sera were collected from pigs aged 3–20 weeks from March through September 2011. HI titers (range, 10–80) were detected in 9.5% of samples, with a GMT of 20.7. To assess the specificity of the HI titers to C/OK in swine sera, we performed HI assays using the human influenza C virus C/Taylor/1233/47 (C/Taylor). Only 2.8% of the swine sera had measurable titers (range, 10–20). Taken together, these results suggest that C/OK virus circulates in swine populations but is not widespread in humans. Further serologic studies focusing on individuals occupationally exposed to swine are required.
To better understand the pathogenesis and epidemiology of C/OK, we performed infection studies with ferrets and swine. We first addressed the zoonotic potential of C/OK virus by conducting a pathogenesis and transmission study in the ferret model. After intranasal inoculation of ferrets, C/OK virus was first detected in nasal washes on day 3 (mean titer, 3.3 log10 TCID50/mL) (Fig. 3). C/OK virus was first detected in ferrets exposed by direct contact to inoculated ferrets on day 7, reaching a mean titer of 4.3 log10 TCID50/mL by day 10. Virus was not detected in ferrets exposed to respiratory droplets. No clinical signs of disease were observed. In the tissues of ferrets on day 5 post-inoculation (p.i.), a mean titer of 3.9 log10 TCID50/mL was observed in the nasal turbinates, but no virus was detected in the upper and lower trachea, lung, small intestine, liver, or spleen. Histopathological examination of lung tissues showed no typical influenza lesions. These results are consistent with a previous study that investigated human influenza C replication in ferret alveolar macrophage cells where viral replication with titers >104 egg infectious dose 50 from days 4 to 9 were measured with no cytopathic effects [33]. All ferrets that were inoculated or exposed by direct contact and 1/3 of the ferrets exposed to respiratory droplets seroconverted 3 weeks after exposure as measured by HI assay (GMT = 780). To assess the pathogenicity and transmissibility of the virus in swine, we similarly challenged swine intranasally with C/OK (Fig. S1). Virus was first detected in nasal swabs on day 3 p.i. by using an rt-RT-PCR method specifically developed for C/OK virus. Virus shedding peaked at day 8 p.i. and remained detectable on day 10. Virus was detected in swine exposed by direct contact on days 7 and 9 after exposure. No clinical signs of illness were observed. Lung samples collected from inoculated swine on day 7 p.i. showed no evidence of the virus by rt-RT-PCR. Histopathological examination of lung tissues showed no typical influenza lesions. Sera collected on day 14 p.i. from donor pigs were positive for antibodies to C/OK virus in an HI assay (GMT = 30.3). All 5 pigs were positive for antibodies to C/OK. Additionally, 2 of the 5 direct contact pigs seroconverted by day 13 post exposure. These data suggest that in animals the replication kinetics is slower for C/OK virus than for influenza A viruses and infection in both swine and ferrets was limited to the upper respiratory tract. The ability of C/OK to readily transmit to contact ferrets suggests that a level of transmission potential to humans is possible. Zoonotic H5N1 and H9N2 influenza A viruses are typically unable to transmit in ferrets.
We compared in vitro cellular tropism between C/OK and human influenza C viruses by rt-RT-PCR in several cell lines, including ST, adenocarcinomic human alveolar basal epithelial (A549), Madin-Darby canine kidney (MDCK), Green African monkey kidney (Marc-145), human rectal tumor (HRT-18G), baby hamster kidney (BHK-21) and porcine kidney (PK-15) cells. C/OK replicated, in order of highest replication, in ST, MDCK, Marc-145, HRT-18G and A549 cells (Fig. S2A). Minimal replication was observed in BHK-21 and PK-15 cells. In marked contrast, the human influenza C virus C/Taylor showed poor growth and replicated only in ST and HRT-18G and not in other cells tested (Fig. S2B). It should be noted however that cultivation of influenza C/Taylor virus was performed at 33°C, the optimal temperature that is typically used to propagate human influenza C virus [34]. This virus failed to replicate at 37°C in our hands which differs from the C/OK virus that can replicate efficiently at this temperature (data not shown). These results suggest that C/OK has a broader cellular tropism than C/Taylor and is also not restricted at elevated temperatures for replication.
The influenza C virus utilizes the 9-O-acetyl-N-acetylneuraminic acid (Neu5,9Ac2) as the primary receptor for attachment to the cell surface to initiate infection [35]. The receptor binding specificity and affinity are mainly determined by the 9-O-acetyl group of the Neu5,9Ac2 [36]. As a result, the virus encodes a sialate-O-acetylesterase, not neuraminidase, in order to release virions from infected cells by cleavage of the 9-O-acetyl group [37]. To provide structural insights to the observed different tropism, we conducted structural modeling of the C/OK HE protein in complex with the receptor based on the solved X-ray crystallographic structure of a human influenza C virus (C/Johannesburg/1/66) HE protein [36]. The overall 53% sequence identity between the two HE proteins allows us to predict important structural features such as the receptor-binding pocket and the enzymatic active site of the C/OK HE protein. Based on the assumption that the HE protein uses similar sites for function, our structural modeling analysis identified a conserved enzymatic active site but revealed a variable receptor-binding pocket between two HE proteins (Figs. 4 and S3). These results are consistent with the observed difference in cellular tropism. For example, both HE proteins possess an identical catalytic triad: S71/H369/D365 for C/Johannesburg and S73/H375/D372 for C/OK (Fig. 4A). The other two substrate-interacting residues for optimal enzymatic function are also completely conserved in the two HE proteins (G99/N131 for C/Johannesburg and G101/N133 for C/OK). In addition, both HE proteins utilize two conserved arginine residues for substrate binding (R72/R332 for C/Johannesburg and R74/R342 for C/OK). The conserved enzymatic site between C/Johannesburg and C/OK suggests that C/OK utilizes 9-O-acetyl sialic acid as the cellular receptor for infection.
Analysis of the receptor-binding pocket formed by a cluster of noncontiguous amino acid residues revealed some interesting similarities and differences between C/Johannesburg and C/OK. The influenza C HE protein uses two binding pockets for recognizing the receptor: one binds to the 9-O-acetyl group while the other engages the 5-N-acetyl group [36]. As shown in Fig. 4, the 5-N-acetyl binding pocket of C/OK HE becomes smaller as compared to that of C/Johannesburg due to L→W substitution, i.e. L198 in C/Johannesburg is replaced by W201 at C/OK HE position 201; the large rigid aromatic side-chain of W201 extends into the binding pocket (Fig. 4B, 4C, and 4D). The binding pocket for the 9-O-acetyl group is nearly identical in C/Johannesburg and C/OK, implying the utility of the 9-O-acetyl-neuraminic acid for C/OK virus infection. For interacting with the 9-O-acetyl group, human influenza C viruses utilize a cluster of amino acid residues Y141, F239, Y241, R250, and R302, which is also used by the C/OK virus except for two phenylalanine residues replacing tyrosine (Y141) and arginine (R250), respectively. We hypothesize that these amino acid differences may alter the binding specificity and affinity of the HE protein to the receptor that in turn result in the observed difference in cellular tropism between two viruses.
Previous studies have indicated the presence of multiple lineages and antigenic groups in influenza virus type C virus [28]–[30]. Despite such variation, it has been long viewed that the influenza C viruses consist of a single subtype [1]. In contrast to this conventional wisdom, here we describe the characterization of a novel influenza C virus from swine with influenza-like illness. The phylogenetic analyses, together with the observations of genomic structure, indicated that this novel virus is more closely related to influenza C than to other members of the Orthomyxoviridae family including influenza A and B viruses, and suggested that the virus could be considered a new subtype of influenza C virus, despite its divergence from human influenza C viruses, which is similar to the divergence between influenza A and B viruses. Identification of this virus was not an isolated case, as we have identified four additional swab samples from swine showing influenza-like symptoms that were positive for this virus in a RT-PCR assay (data not shown). These samples were collected from different pig farms across the U.S. between 2010 and 2012.
The finding that influenza C virus, like influenza A, harbors multiple subtypes is significant. It suggests the possibility of reassortment between subtypes, which could potentially generate viruses with phenotypes that may pose a threat to public health. Nine and a half percent of surveyed swine possessed antibodies specific to C/OK indicating that this previously unidentified virus circulates in U.S. swine. A causal relationship is evidently supported by the experimental infection of pigs. As such, reassortment may occur between C/OK and human influenza C viruses in pigs because both viruses can infect and transmit among pigs and because swine has been documented in serving as a mixing vessel for reassortment of influenza A viruses [38], [39].
Despite a lack of compelling evidence of C/OK virus infection in humans, we suspect that the virus may infect and replicate in the human population because of the following reasons: First of all, the virus infects and transmits in ferrets, a surrogate for human influenza pathogenesis studies. The ability of C/OK virus to readily transmit to contact ferrets suggests that a level of transmission potential to humans is possible. Second, C/OK virus displays a broader cellular tropism compared to human influenza C virus, and this virus seems quite plastic in terms of propagation because high temperatures such as 37°C do not restrict its replication at least in cell culture. Human influenza C virus causes a mild respiratory disease in humans and the infection is normally confined to the upper respiratory tract, although occasionally it can also cause lower respiratory infection [5]. Considering that C/OK virus varies significantly from currently circulating human influenza C viruses in the amino acid sequences of predicted proteins, assessment of clinical disease in humans, particularly in children, caused potentially by C/OK is justified. Knowledge of its presence in human clinical settings is important to any future attempt to manage and control the disease outbreak. By using the nucleotide sequence of C/OK virus reported in this work, sensitive and specific diagnostic methods can be developed to investigate the pathogenesis and epidemiology of this novel virus in humans.
Influenza C virus is not readily isolated and cultured, and primary isolation can be challenged. The obstacle is largely due to a lack of suitable cell lines for influenza C virus isolation as suggested previously [5], [34]. In contrast, identification and cultivation of the C/OK in the ST cell line is relatively straightforward. The emerging but puzzling question then is why this virus has not been identified until now. We suspect that several factors including use of less susceptible cell lines and complicated co-infection often involving influenza A viruses and other viruses having the capacity to agglutinate red blood cells may account for the previous failures in identifying this virus. Alternatively, the C/OK virus may have spread to swine in recent years from an unknown animal reservoir. On-going retrospective seroepidemiological analyses will help to address this question.
The difference in cellular tropism between C/OK and human influenza C may be a result of differences in the receptor recognition of the HE protein. To explore the possible structural origins of this difference, we created a homology model for C/OK using the crystal structure of the influenza C HE protein (Fig. S3). The high sequence identity (53%, Fig. S3) between C/OK HE and influenza C suggest that the quality of this model will be high, given previous work in homology modeling and structure validation for influenza A and B HA proteins in which similar sequence similarities gave RMSD errors of ∼1 Å [40], [41]. The predicted receptor binding for C/OK HE appears on the top face of the receptor binding domain similar to human influenza C HE protein and remote homolog HE proteins of other viruses such as coronavirus and torovirus (∼30% sequence similarity to influenza C HE) [42]. Here, four out of nine residues of the receptor binding site and residues around the binding pocket of HE of human influenza C are retained in that of C/OK (Fig. S4). Previous work has shown that receptor binding specificity and affinity are sensitive to substitutions in the receptor binding site of HE and HA [43], [44]. Examining the predicted receptor-binding site of C/OK HE revealed changes in the receptor binding site relative to human influenza C HE. The most notable difference is a reduction in the size of the 5-N-acetyl binding pocket of C/OK HE relative to C HE due to the L198 in C HE being replaced by W in C/OK HE (Fig. 4). Alternatively, the binding pocket for the 9-O-acetyl group is similar between the two HE proteins (Fig. 4). We speculate that these differences may indicate that C/OK HE utilizes a different substrate than influenza C HE. Further experimental verifications of are required to test this prediction.
It has been suggested that influenza A, B, C viruses have a common precursor, and of the three virus types, influenza A and B viruses are much more similar to each other in genome organization and protein homology than to C viruses, which suggests that influenza C virus diverged well before the split between A and B viruses [9]. Numerous studies have shown that influenza C viruses have the slowest evolutionary rate among influenza viruses [13], [19], [45]–[47]. One theory is that influenza C viruses, like influenza B viruses, are close or at an evolutionary equilibrium in humans, whereas influenza A viruses have not yet reached an equilibrium [48]. Consistent with this hypothesis is that only a single subtype is thought to exist for influenza B and C viruses and humans, not other mammals, are the primary hosts of influenza B and C viruses. The discovery of C/OK in pigs, being distantly related to human C viruses, seems to challenge these accepted views and warrant future studies of influenza C virus evolution.
Virus nomenclature is the subject of discussion and there is still a possibility that C/OK virus can be assigned as the prototype of a new genus of the Orthomyxoviridae family. Most compelling evidence in support of the tentative designation of the C/OK virus are (i) seven genomic segments, (ii) 3′ and 5's NC regions similar to those of human influenza C viruses, and (iii) HE protein sharing approximately 53% homology with that of influenza C viruses. The last parameter is the primary determinant to classify subtypes of influenza A virus. However, the overall divergence between C/OK and human influenza C viruses is similar to that observed between influenza A and B viruses and argues for classification of C/OK into a potential new virus genus. Influenza A, B, and C viruses are classified on the basis of antigenic differences between their nucleoprotein (NP) and matrix (M) proteins [1]. Intriguingly, only modest homologies of these structural proteins between C/OK and human influenza C further confound the provisional classification for this newly discovered virus. Recently the family Orthomyxoviridae was expanded by including two novel genera, Thogotovirus, consisting of three viruses that infect birds and ticks, as well as the genus Isavirus, consisting of infectious salmon anemia virus [49], [50]. For the novel swine influenza virus reported here, perhaps not until attaining more detailed serological, virological, and molecular data, a final classification of this virus can be made. Of particular importance are reassortment experiments between C/OK and human influenza C viruses. While similar, two discrepancies were found in the NCR's of C/OK as compared to human influenza C viruses. It is not known whether these mutations prevent reassortment between C/OK and human influenza C viruses. Preliminary reassortment experiments between C/OK and C/Taylor have been performed and have failed to identify reassortant viruses. More detailed studies are underway.
In summary, we identified a novel influenza C virus that infects and spread among pigs or ferrets by direct contact. The ability of this novel pathogen to infect ferrets; a surrogate for human influenza infection suggests that such viruses may become a potential threat to human health. Our finding reported in the present study raises several interesting questions. Does this influenza C-like virus have the capability of generating a viable reassortant with currently circulating human C viruses? If so, could such a reassortment allow influenza C virus to diverge and to have greater pathogenicity? When and where did this novel virus emerge? What is its animal reservoir in nature? Future elucidation of these questions will provide insights into the ecology, virology, and pathobiology of influenza C virus.
Ferret experiments were conducted in an Animal Biosafety Level 2+ (level 2 with enhanced biocontainment for pandemic H1N1 influenza A virus) facility at St. Jude Children's Research Hospital, in compliance with the policies of the National Institutes of Health and the Animal Welfare Act and with the approval of the St. Jude Children's Research Hospital Animal Care and Use Committee (IACUC No. 428). Pig experiments were performed at Newport Laboratories under biosafety level 2 conditions in accordance with the Guide for the Care and Use of Agricultural Animals in Research and Teaching and were approved by the Institutional Animal Care and Use Committees at Newport Laboratories (IACUC No. 02-2012).
Nasal swabs were collected from 15-week-old pigs exhibiting influenza-like illness at a commercial swine production facility in Oklahoma, USA, in April, 2011. Viral isolation was performed on swine testicle cells and cytopathic effects were evident by day 3 post inoculation. Detailed information on cell culture conditions is available in SI Appendix, Supplementary Materials and Methods. Hemagglutination assays were performed using chicken red blood cells.
C/OK was concentrated from ST cell supernatants by ultracentrifugation and subsequently purified through a 20% sucrose cushion by ultracentrifugation. Viral RNA was isolated using the Qiagen Viral RNA Isolation Kit and converted to cDNA using random primers included in the GoScript Reverse Transcription Kit (Promega). The cDNA was made double stranded with DNA polymerase and used to construct a library for Ion Torrent Sequencing. Detailed sequencing methodology is available in SI Appendix, Supplementary Materials and Methods.
Contigs were assembled de novo by using SeqMan NGen software (DNAStar). Contigs encoding proteins with homology to influenza C proteins were identified by BlastP analysis. The genome sequence of C/OK was submitted to Genbank under accession no. JQ922305-JQ922311, relating to segments 1–7, respectively. Phylogenetic analyses were performed by using Mega 5 software [51]. Evolutionary analyses were conducted by using the Maximum Likelihood algorithm, and the tree topology was verified by performing 1000 bootstrap replicates.
The PB1 sequence was used to design primers and a Taqman probe for detection of C/OK (position of primers and probe in PB1 gene: forward, nucleotides 1420–1439; reverse, nucleotides 1555–1535; probe, 1482–1460). Viral RNA was extracted by using the MagMAX-96 viral RNA isolation kit (Life Technologies) according to the manufacturer's instructions. rt-RT-PCR was performed by using QIAGEN Quantitect RT-PCR with the C/OK primers and probe. Method specificity was assessed by using influenza A, B, and C reference viruses, and no cross-reaction was observed.
Swine testicle (ST) cells were grown in DMEM containing 5% fetal bovine serum. Influenza C/Taylor/1233/47 virus was provided by BEI Resources (NIAID). For infection, cell medium was replaced with DMEM and viral inoculum was added at a multiplicity of infection of 0.001. Viral growth studies were performed on a monolayer of ST, A549, Marc145, HRT-18G, BHK-1, or PK-15 cells using an inoculum of 1.0–3.0 TCID50/ml in duplicate (multiplicity of infection 1×10−5–1×10−3). C/OK virus replication was performed at 37°C while C/Taylor replication was done at 33°C. Samples were removed at 0, 24, 48, and 72 hours p.i. and virus was titrated by rt-RT-PCR. Experiments were performed three times in duplicate.
Human sera were treated with receptor-destroying enzyme (Denka Seiken Co., Tokyo, Japan) overnight at 37°C, heat-inactivated at 56°C for 30 min, diluted 1∶10 with PBS, and tested by hemagglutination inhibition (HI) assay with 0.5% packed chicken red blood cells (cRBCs) as described in the WHO Manual on Animal Influenza Diagnosis and Surveillance [52].
The pathogenicity and transmission of the virus was tested in 3- to 4 month-old male ferrets. Detailed pathogenicity and transmission methodology is available in SI Appendix, Supplementary Materials and Methods. Three donors were inoculated intranasally under light isoflurane anesthesia with 106 TCID50 of swine/Oklahoma/1334/2011 virus in 1 ml of sterile PBS. Two additional ferrets were similarly inoculated and were housed separately for virus titration and histopathology in organs. At 23 h p.i., each of the three remaining donor ferrets was housed in a cage with one naïve direct-contact ferret (n = 3). An additional ferret (n = 3) was placed in an adjacent cage separated from the donor's cage by a two layers of wire mesh (∼5 cm apart) that prevented physical contact but allowed the passage of respiratory droplets. Clinical signs of infection, relative inactivity index [53], weight, and temperature were recorded on days 0, 3, 5, 7, and 10 p.i.. Nasal washes were collected from ferrets 3, 5, 7, and 10 days p.i. Two donor animals were euthanized 5 dpi, and tissue samples were collected. Samples were homogenized and virus was titrated (log10 TCID50 per gram of tissue) in ST cells. Tissues were also subjected to histopathologic analysis.
Swine challenge studies were performed at Newport Laboratories under biosafety level 2 conditions. Twenty-eight swine approximately 10 weeks of age were obtained from a commercial high-health herd. Eleven swine were placed in a single room and inoculated intranasally with 6.0 log10 TCID50 of C/OK. On day 1 p.i., 11 naïve direct-contact swine were introduced into the room. Temperatures were recorded and nasal swabs were collected on days 0, 2, 3, 6, 8, and 10 p.i. Six inoculated swine and three mock-inoculated swine were euthanized on day 7 p.i. and lung specimens were fixed in 10% neutral buffered formalin and submitted for histopathological analysis. The remaining swine were euthanized on day 14 p.i. Nasal swabs and lung tissue were analyzed by rt-RT-PCR as described above.
The structure of C/OK HE (aa 17–620) was modeled using Modeller 9.10 [54]. The structure of C/Johannesburg/1/66 HE was used as template (PDB id: 1FLC) [36] because the two HE proteins show both high sequence and secondary structure similarity. The quality of the modeled structure was estimated by Verify_3D and 95.47% of the modeled residues are compatible with the structure (averaged 3D-1D score >0.2) [55], suggesting that the quality of the modeling is good. The distance between the modeled structure and template is 0.49 angstrom. Electrostatic surface maps were generated using APBS [56].
|
10.1371/journal.pntd.0006225 | Failure of fluconazole in treating cutaneous leishmaniasis caused by Leishmania guyanensis in the Brazilian Amazon: An open, nonrandomized phase 2 trial | The treatment of Leishmaniasis caused by Leishmania (Viannia) guyanensis is based on a weak strength of evidence from very few clinical trials and some case series reports. Current treatment guidelines recommend pentamidine isethionate or meglumine antimoniate (Glucantime) as the first-line choices. Both are parenteral drugs with a low therapeutic indexes leading to a high risk of undesired effects. Imidazole derivatives interfere with the production of leishmanial ergosterol, an essential component of their membrane structure. One drug that has been studied in different clinical presentations of Leishmania is fluconazole, a hydrophilic bis-triazole, which is easily absorbed through the oral route with a low toxicity profile and is considered safe for children. This drug is readily available in poor countries with a reasonable cost making it a potential option for treating leishmaniasis.
An adaptive nonrandomized clinical trial with sequential groups with dose escalation of oral fluconazole was designed to treat adult men with localized cutaneous leishmaniasis (LCL) in Manaus, Brazil. Eligible participants were patients with LCL with confirmed Leishmania guyanensis infection.
Twenty adult male patients were treated with 450 mg of fluconazole daily for 30 days. One patient (5%) was cured within 30 days of treatment. Of the 19 failures (95%), 13 developed a worsening of ulcers and six evolved lymphatic spreading of the disease. Planned dose escalation was suspended after the disappointing failure rate during the first stage of the trial.
Oral fluconazole, at the dose of 450mg per day, was not efficacious against LCL caused by Leishmania guyanensis in adult men.
Brazilian Clinical Trial Registration (ReBec)—RBR-8w292w; UTN number—1158-2421
| The main agent causing localized cutaneous leishmaniasis in the Guianan Ecoregion Complex, a region that covers Guiana, Suriname, French Guiana, the southern portion of Venezuela and the northern portion of the Amazon Basin, is Leishmania (Viannia) guyanensis. The current treatment regimens for treating this parasite follow a weak strength of recommendation based on a low-quality evidence. All the drugs used are parenteral drugs with low therapeutic indexes and inadequate cure rates. Azoles are drugs that interfere with the production of leishmanial ergosterol and have been studied in other Leishmania species with promising results. Fluconazole is a hydrophilic bis-triazole that accumulates rapidly and extensively in the skin with a low toxicity profile. It is considered safe for children. Oral fluconazole is readily available and has a reasonable cost. In this work, we estimated the efficacy of fluconazole in treating localized cutaneous leishmaniasis caused by L. guyanensis in Manaus. The study had to be stopped due to the disappointing results, with only a 5% cure rate. Among the failures, 55% developed a marked worsening of their ulcers and 55% of those patients had lymphatic dissemination of the infection. In conclusion, fluconazole at the dose of 450mg per day is not efficacious in treating localized cutaneous leishmaniasis caused by L. guyanensis.
| Cutaneous leishmaniasis (CL) is one of the 17 neglected diseases, with 350 million people at risk in 98 endemic countries [1]. Latin America is a highly endemic region with a documented 30% increase in the number of reported cases during a period of ten years (2001–2011) [2]. Brazil together with eight other countries reports 90% of all registered cases of CL [1]. The incidence of CL in the Brazilian Amazon region surpasses the Brazilian national average (15.3 cases/100.000 inhabitants), reaching in some regions the incidence of 30 cases/100,000 inhabitants [3].
American tegumentary leishmaniasis (ATL), especially in the Amazon region, has a peculiar epidemiological profile characterized by higher intra- and inter-specific variations and heterogeneity of transmission cycles, reservoir hosts and sand fly vectors, with sympatric circulation of various Leishmania species. These characteristics altogether lead to diverse clinical presentations and distinct clinical responses to treatment [4].
Leishmania guyanensis is the most common cause of human leishmaniasis in the Guianan Ecoregion Complex (GEC) a region that covers Guiana, Suriname, French Guiana, the southern portion of Venezuela and the northern portion of the Amazon Basin [5, 6]. Cutaneous leishmaniasis is the most common clinical presentation of the infection caused by Leishmania guyanensis [7]. Lymphatic involvement, manifested as adenomegaly or lymphangitis, is the second most common presentation of the disease affecting 60% of the cases [8]. Disseminated cutaneous leishmaniasis and mucosal disease are less common presentations of leishmaniasis caused by this Leishmania species [9, 10].
The current treatment regimens for CL caused by Leishmania guyanensis follow a weak strength of recommendation based on a low-quality evidence [11]. Intramuscular administration of 3 mg/kg pentamidine isethionate every other day for up to four injections is considered the treatment of choice [12]. Meglumine antimoniate (Glucantime) 20mg Sv/Kg/day for 20 days is the recommended treatment in Brazil with a cure ratebetween 53% and 70% [13, 14]. Miltefosine has been used in some studies with a cure rate between 54% and 72% [15, 16], although that drug is not available in South America. Amphotericin B is the drug choice in severe cases.
The present treatment options are parenteral drugs with low cure rates and low therapeutic indexes leading to a high risk for undesired effects. This scenario leads us to research a new therapeutic option.
One drug that has been studied in different clinical presentations of Leishmania is fluconazole. This agent is a hydrophilic bis-triazole that is easily absorbed via the oral route and interferes with the production of leishmanial ergosterol, an essential component of the membrane structure [17]. Fluconazole, has a low toxicity profile, it accumulates rapidly and extensively in the skin, and it is readily available, with a reasonable cost. Fluconazole is also considered safe for children [18, 19].
This adaptive phase II trial evaluated the efficacy of fluconazole in the treatment of cutaneous leishmaniasis caused by L. guyanensis.
From December 2014 through February 2016, twenty-eight subjects with parasitological confirmed diagnosis of LCL were recruited at the Tropical Medicine Foundation Dr. Heitor Vieira Dourado, in the state of Amazonas, North Brazil, an endemic area of L. guyanensis infection. The flow diagram of participants through the different study phases is described in Fig 1.
Localized cutaneous leishmaniasis was defined as the presence of up to five ulcerous lesions without lymphatic or mucousal disease, with amastigotes visualized in direct examination of Giemsa-stained smears of a dermal scrapping taken from the ulcerated border of at least one lesion.
Inclusion criteria were as follows: (1) diagnosis of LCL based on case definition, (2) illness duration <3 months, (3) male sex with age of at least 18 years, (4) 1 to 5 ulcerated lesions, and (5) no previous treatment for leishmaniasis.
Exclusion criteria were as follows: (1) Leishmania species could not be identified; (2) infection caused by species other than L. guyanensis; (3) any uncontrolled active infectious or severe disease, and (4) an allergy to fluconazole.
Complete blood cell count, tests for the levels of aspartate and alanine aminotransferase, amylase, lipase, urea, creatinine and glucose, and an electrocardiogram were performed in all participants before therapy and at 30, 60, 90 and 180 days after treatment. All patients were subjected to a rapid human immunodeficiency virus test and serology for hepatitis B and C. All the ulcers were measured and photographed. A biopsy of each ulcer was performed and the material was used for a parasite culture and histopathology. Leishmania species were identified as described by Marfurt et al [20].
Scheduled patient visits were made 30, 60, 90 and 180 days after beginning the treatment. If a patient did not return to follow-up at the specified time, visits were conducted in the patient’s home on the same day or within 7 days of the missed appointment. Patients’ ulcers were measured with a flexible ruler at the initial visit and at each follow-up visit. Standardized digital photographs of the patients’ lesions were obtained at the same time points.
Patients were monitored for adverse events (AEs) and treatment adherence. Patients returned the blister packs of fluconazole to verify compliance. Clinical and laboratory AEs were graded according to the Common Terminology Criteria for Adverse Events of the National Cancer Institute [21].
Outcome measures followed the protocols published by Olliaro et al [22]. The primary endpoint was a definitive cure six months after the end of treatment. A definitive cure was defined as the complete epithelialization of all lesions without raised borders, infiltrations, inflammation or crusts. The secondary endpoint included an initial cure defined as complete epithelialization of ulcers two months after the end of treatment. If an initial cure was not attained it was considered a therapeutic failure. Any interruption of the treatment was also considered a therapeutic failure.
All patients included in the therapeutic failure group received a rescue therapy of meglumine antimoniate, 20mg Sbv/Kg/day for 20 days.
Fluconazole 150 mg capsules conditioned in blisters packs containing 10 capsules were self-administered orally for 30 days. The first dosing schedule involved 450 mg of fluconazole administered once daily, and the second dosing schedule involved 900 mg of fluconazole administered as two daily doses of 450 mg. Both schedules lasted 30 days.
An adaptive phase II trial was adopted based on the successful use of fluconazole in treating L. (Viannia) braziliensis in a study published by Sousa et. al [23].
The adaptive trial followed FDA recommendations [24].
The sample size of 30 patients in the first step of the study was calculated considering an estimated cure rate of 60%, a precision of +/- 20% and an alpha error of 5%. The patient was instructed to take three capsules (450mg total dose) orally once daily in the morning for 30 days. If 18 patients had reached the primary endpoint with this regimen, then the second step of the study would have begun.
The second step was designed based on the assumption of an improvement in magnitude of response of at least 10% after doubling the dose of fluconazole. Considering a cure rate of 70%, a precision of +/- 10% and an alpha error of 5%, a sample size of 73 patients was calculated. Patients in the second step would have received three 150 mg fluconazol in the morning and three 150 mg fluconazole in the afternoon (900 mg total dose) for 30 days.
All statistical analyses were performed with SPSS 21.0 software for Windows.
This trial was conducted according to the Declaration of Helsinki. Before they were enrolled in the study, written informed consent was obtained from all patients. The study was approved by the Ethics Committee of the Tropical Medicine Foundation Dr Heitor Vieira Dourado, Brazil - registration number 26118613.4.0000.0005. This clinical trial was registered in ReBEC (Brazilian Registry of Clinical Trials) with the identifier RBR-8w292w and is available from http://www.ensaiosclinicos.gov.br/rg/view/2668/ UTN number–1158–2421
Twenty-eight adult male patients fulfilled the inclusion criteria and were accepted into the study. Their median age was 38.3 years old (range 18–56). Of those, eight (28.6%) were excluded. In three cases the causative species was not Leishmania guyanensis, and in the other five, the parasite species could not be identified. Twenty adult male patients (71.4%) remained in the study for further analysis.
Those 20 patients presented 40 lesions, all with less than 3 months of duration. Fifty-five percent (n = 11) presented with one lesion, 15% (n = 3) with two lesions, 15% (n = 3) with three lesions, 5% (n = 1) with four lesions and 10% (n = 2) with five lesions. Concerning the distribution of the lesions, they were predominantly on exposed areas of the body, with 37.5% (n = 15) located on the lower limbs, 37.5% (n = 15) located on the upper limbs and 15% (n = 6) located on the face. Ten percent of the lesions (n = 4) were located on the trunk.
The median diameter of all 40 ulcers on the first day of treatment was 1.70cm (range 0.2 cm– 4 cm). During the trial, five participants with 17 total ulcers were excluded before the end of 30 days of treatment. At the end of treatment (day 30), 23 lesions on 15 patients remained active ulcers with a median diameter of 2.80 cm (range 0.4 cm– 5 cm). Twelve lesions on six patients remained active ulcers at day 60 (30 days after the end of treatment), with a median diameter of 2.37 cm (range 0.3 cm– 4.5 cm). The clinical evolution of the ulcers is represented in Fig 2.
A definitive cure was documented in 5% (n = 1) of the cases, as shown in S1 Fig. The remaining 19 patients were considered treatment failures. Five patients asked to change medication due to ulcer enlargement. In eleven cases (55%) there was worsening of the ulcers with marked inflammatory signs (Fig 3). In six patients (30%) lymphatic spread of the disease was also noted (Fig 3). In those 19 patients, fluconazole treatment was stopped and rescue therapy was instituted with complete resolution of the ulcers.
The drug was well tolerated with mild self-limited systemic adverse events in five patients (27.78%) as shown in Fig 4.
The first report of the activity of an azole against a species of the genus Leishmania came from an in vitro test of CIBA 32,644-Ba in 1965 and another of 2-amino-5-(1-methyl-5-nitro-2-imidazolyl)1-3-thiadiazole in1968 [25, 26]. At the end of the 1960s and in the 1970s, in Brazil, the clinical efficacy of 1-(5-nitro-2-tiazolil) 2-imidazolidinone and niridazole in cutaneous and mucosal leishmaniasis caused by L. (Viannia) braziliensis was reported, in case series, with some clinical response, at the cost of serious neurologic adverse events [27, 28].
The interest in using azoles in the treatment of leishmaniasis was revived after the report published by Berman indicating the activity of ketoconazole against leishmanial species in macrophage culture [29].
The efficacy of the azoles in treating ATL is mainly based on case series or small trials with heterogeneous results. The cure rate of ketoconazole 400mg bid was as follows: one out of six patients with ATL caused by L. guyanensis [30], three out of three patients with ATL caused by L. braziliensis [31] and 16 out of 22 patients with ATL by L. panamensis [32]. In two studies, itraconazole cured six cases out of ten [33] in one study and three out of 13 patients in the other [34].
The first clinical use of fluconazole in leishmaniasis was against kala-azar with 0% definite cure. Some patients had early apparent cures with later relapsing [35]. In 2002, Alrajhi et al. published a randomized, placebo-controlled trial and concluded that a six-week course of 200 mg fluconazole daily was safe and useful to treat CL caused by L. major [36]. Afterwards, this drug became an alternative for the treatment of Old World cutaneous leishmaniasis. A trial published by Emad et al. evidenced that 400mg of fluconazole daily was more efficacious in infections caused by L. major when compared to 200mg daily, with six-week cure rates of 81% versus 48.3% respectively [37].
In Brazil, a case series was conducted where 28 patients with confirmed leishmaniasis caused by L. (V.) braziliensis, who refused or could not use antimonials, received oral fluconazole for 20 days. Eight patients received 5mg/Kg/day with a cure rate of 75%, 14 patients received 6.5mg/Kg/day with a cure rate of 92.8% and six patients received 8mg/Kg/day with a cure rate of 100% [23]. The authors concluded that there was a higher efficacy at higher doses of fluconazole.
Treatment of leishmaniasis caused by L. guyanensis follows a weak strength of recommendation based on a low-quality of evidence [11]. Eight published trials analyzed the cure rate in cases of ATL caused by this species, with results varying from 53.6% to 91.7% (S1 Table).
Considering the lack of an optimal treatment for ATL caused by L. guyanensis, fluconazole seemed to be a promising drug alternative for treating this disease. This assumption was based on the mechanism of action of the drug in the protozoan ergosterol metabolism and on the clinical efficacy evidenced in clinical trials of Old World leishmaniasis and the clinical response against L. braziliensis [23].
In an adaptive clinical trial design it is possible to evaluate the desired outcomes exposing fewer patients without losing the quality of the evidence [38]. The sequential groups with scaled doses allow timely suspension of the research after initial failures with a smaller dose, minimizing unnecessary drug exposure.
During the execution of the study, the first clinical trial evaluating fluconazole in ATL caused by L. braziliensis was published. The intention-to-treat analysis two months after treatment showed cure rates of 22.2% (6 out of 27) in the fluconazole group and 53.8% (14 out of 26) in the Glucantime group. The per protocol results were the same at six months after the end of treatment. The AEs were similar in both groups [39].
In this study, the cure observed with 450 mg of fluconazole daily against L. guyanensis was 5% (1 out of 20). Five patients asked to stop taking fluconazole and to receive the rescue therapy. One patient who had clinical failure refused to receive any parenteral medication, and after 120 days he presented no signs of the disease. It is not clear if this outcome was a spontaneous evolution to a cure or a delayed response to the fluconazole treatment.
The remaining 11 patients presented a peculiar clinical outcome. The ulcers, after a period of ten days, began to show remarkable inflammatory signs associated with intense pain, increased size and, in some cases the lymphatic spread of the disease (Fig 3). The early enlargement of the ulcers during treatment with meglumine antimoniate was previously reported, but the intensity of inflammatory signs and the concomitant lymphatic involvement constitutes a novel finding never reported in patients with leishmaniasis exposed to fluconazole [40].
The mechanism of this inflammatory phenomenon deserves more investigation, considering that this high failure rate may be justified by an unrecognized immunological mediated effect associated to fluconazole.
All 16 patients that received Glucantime as rescue therapy were cured after 30 days. One patient moved from Manaus and the researchers lost contact.
Fluconazole was well tolerated systemically, with five cases of mild AEs.
In conclusion, fluconazole, at the dose of 450mg per day, is not efficacious against leishmaniasis in adult men infected by L. guyanensis. The clinical worsening during fluconazol exposure deserves more attention and should be evaluated in future studies involving fluconazole or other azole therapies.
|
10.1371/journal.pgen.1003355 | Ubiquitous Polygenicity of Human Complex Traits: Genome-Wide Analysis of 49 Traits in Koreans | Recent studies in population of European ancestry have shown that 30%∼50% of heritability for human complex traits such as height and body mass index, and common diseases such as schizophrenia and rheumatoid arthritis, can be captured by common SNPs and that genetic variation attributed to chromosomes are in proportion to their length. Using genome-wide estimation and partitioning approaches, we analysed 49 human quantitative traits, many of which are relevant to human diseases, in 7,170 unrelated Korean individuals genotyped on 326,262 SNPs. For 43 of the 49 traits, we estimated a nominally significant (P<0.05) proportion of variance explained by all SNPs on the Affymetrix 5.0 genotyping array (). On average across 47 of the 49 traits for which the estimate of is non-zero, common SNPs explain approximately one-third (range of 7.8% to 76.8%) of narrow sense heritability.
The estimate of is highly correlated with the proportion of SNPs with association P<0.031 (r2 = 0.92). Longer genomic segments tend to explain more phenotypic variation, with a correlation of 0.78 between the estimate of variance explained by individual chromosomes and their physical length, and 1% of the genome explains approximately 1% of the genetic variance. Despite the fact that there are a few SNPs with large effects for some traits, these results suggest that polygenicity is ubiquitous for most human complex traits and that a substantial proportion of the “missing heritability” is captured by common SNPs.
| The “missing heritability” problem has been intensely debated for the last few years. Possible explanations include the existence of many genetic variants each with a small effect, rare variants with large effects, and heritability being over-estimated. Previous studies using whole-genome estimation have demonstrated that for human complex traits such as height, body mass index, and intelligence, a large portion of the heritability can be captured by all the common SNPs on the current genotyping arrays. These studies, however, were all concentrated only on a few traits. In this study, we analysed 49 quantitative traits in a sample of ∼7,000 unrelated Korean individuals. We found that, on average over all the traits, common SNPs on the Affymetrix 5.0 genotyping array explain approximately a third of the heritability, that genetic variants are widely distributed across the whole genome with longer chromosomes explaining more phenotypic variation, and that approximately any 1% of the genome explains 1% of the heritability. Despite examples where a few variants explain a substantial amount of variation, all these results are consistent with polygenicity being ubiquitous for most complex traits.
| The five years wave of genome-wide association studies (GWAS) has uncovered thousands of single nucleotide polymorphisms (SNPs) to be associated with hundreds of human complex traits including common diseases [1], [2]. Yet, for most complex traits, the gap between the proportion of phenotypic variance accounted for by the top SNPs that reached genome-wide significance level in GWAS and the heritability estimated from pedigree analyses remains unexplained [3]. This was called the “missing heritability” problem [4], explanations to which have been debated in the field [3]. Taking height and BMI for example, well-powered studies with a discovery sample of over 100,000 individuals have identified 180 and 32 loci to be associated with height [5] and BMI [6], which explain ∼10% and ∼1.5% of variance for height and BMI, respectively, while the heritability was estimated to be ∼80% for height [7] and 40∼60% for BMI [8], [9]. On the other hand, however, recent studies using whole-genome estimation approaches have demonstrated that a large proportion of heritability for height [10], [11], body mass index (BMI) [11], schizophrenia [12] and rheumatoid arthritis (RA) [13] can be captured by all the common SNPs on the current genotyping arrays, which implies that there are a large number of variants each with an effect too small to pass the stringent genome-wide significance level. It could be argued that the evidence from these whole-genome estimation analyses are for the traits that are known to be highly polygenic and therefore are not representative for most human complex traits. Therefore, it remains unclear whether polygenic inheritance is a general phenomenon for most human complex traits or a unique feature for a particular group of traits such as height and BMI. There has been evidence from a review of a number of GWAS that more variants have been identified with increased sample size [2], consistent with a pattern of polygenic inheritance for most common diseases and complex traits. In this study, using the whole-genome estimation and partitioning approaches [10], [11], [14], we directly estimated the proportion of phenotypic variance explained by the common SNPs all together on a genotyping array for a range of quantitative traits in a large homogenous sample of Koreans. We demonstrated by a number of different analyses that polygenic inheritance is likely to be ubiquitous for most human complex traits.
We used the data from the Korea Association Resource (KARE) project [15]. The KARE cohort consists of 10,038 individuals recruited from two different sites in South Korea, genotyped at 500,568 SNPs on Affymetrix Human SNP array 5.0. There were 7,170 unrelated individuals and 326,262 autosomal SNPs after quality controls (Materials & Methods). We show by principal component analysis that all the individuals are of eastern Asian ancestry (Figure S1). All the individuals were measured for 49 quantitative traits, which are related to obesity, blood pressure, hyperglycemia, diabetes, liver functions, lung functions, and kidney functions (Table S1). The phenotypic correlations between pairwise traits are visualized in Figure S2, with traits within the same classification groups being more correlated than between groups.
We then estimated the proportion of variance explained by fitting all the SNPs in a mixed linear model for each of the 49 traits (Materials & Methods). In general, there was a substantial amount of variance explained by all SNPs on the Affymetrix 5.0 genotyping array () for most traits with a mean of 12.8% (a range from 0 to 31.6%) across all the 49 traits (Table 1). For 47 of the 49 traits, the estimate of was non-zero, 43 of which reached the nominal significance level (likelihood ratio test P<0.05) and 26 of which reached experimental-wise significance level after Bonferroni correction for multiple traits (likelihood ratio test P<0.001) [14]. We compared the estimates of with the narrow-sense heritability (h2) estimated from pedigree analyses in the literature (Table S2), and observed a significant trend (P = 0.017) that traits with a higher estimate of h2 were more likely to have a larger estimate of (Figure S3) and that all the common SNPs explain approximately 33.3% (a range from 7.8% to 76.8%) of the narrow-sense heritability, despite that the estimates of h2 were from various different studies, usually with large standard errors and mostly in samples of European ancestry. In contrast, when we performed a genome-wide association (GWA) analysis in the same sample, we identified genome-wide significant (P<5×10−8) SNPs for 25 of the 49 traits. On average across the 25 traits, the top associated SNPs from GWA analyses explained only 1.5% (range of 0.5% to 3.8%) of phenotypic variance (Table S2), nearly 10-fold smaller than the estimate of , suggesting there are many SNPs remaining undetected because of the lack of statistical power. In addition, we estimated the variance explained by all the SNPs imputed to HapMap2 CHB and JPT panels (Materials & Methods and Table S2). The estimate of averaged across all the traits using imputed data (13.8%) was slightly higher than that using genotyped data (12.8%).
We calculated the proportion of SNPs with p-values that passed a threshold p-value in a GWA analysis (θP) for each trait. We calculated θP for a range of threshold p-values and plotted them against the expected values under the null hypothesis of no association (i.e. the threshold p-values) (Figure S4). This plot is an analogue to the QQ plot. The averaged θP over all the traits started deviating from the expected value when the threshold p-value became small (Figure S4A) and such deviation varied across traits (Figure S4B). The question is whether a trait that shows a larger value of θP will also tend to have a larger estimate of . We then correlated θP with the estimates of across all the traits for a threshold p-value and calculated such correlations for a range of threshold p-values, from 0.001 to 0.201 by 0.05. We found a maximum of squared correlation of 0.923 at the threshold p-value of 0.031 (Figure 1), meaning that traits that have more proportion of SNPs passed a significance level in GWAS also have more proportion of phenotypic variance explained by all SNPs. It should be noted that the threshold p-value at which the maximum correlation between the estimate of and θP was found depends on sample size. This analysis is an alternative way to demonstrate the equivalence between GWAS and the whole-genome estimation analysis as implemented in GCTA. Although the whole-genome estimation approach estimates the variance explained by all SNPs regardless of individual SNP-trait associations, the estimate of is actually mainly attributed to SNPs that show stronger evidence for association with the trait, e.g. ∼92% of the estimate of could be determined by SNPs with association p-values<0.031 given the sample size of ∼7,000 in this study. These results also suggest that there are many common variants associated with the traits at nominally significant level (P<0.05) but their effect sizes are too small to be genome-wide significant (P<5×10−8).
Using the same method as above but allowing to fit multiple genetic components simultaneously in the model (Materials & Methods), we then partitioned into the contributions of individual chromosomes for all the 49 traits (Table S3) except HOMA and INS0 for which the estimates of were zero (Table 1), and plotted the estimate of variance explained by each chromosome () against chromosome length (LC) for each trait. We did not observe a linear correlation between and LC for any particular traits (Figure S5) as strong as that shown in the previous studies for height [11] and schizophrenia [12]. The squared correlation between and LC was from 0.00 to 0.48 with a mean of 0.15 and a standard deviation of 0.12. This result is not unexpected because the sample size of this study is smaller than that of the previous analysis so that in our analysis were estimated with larger sampling errors. We then averaged the estimates of over all the traits to reduce the sampling error variance and found that the averaged estimate of was strongly correlated with LC with a correlation of 0.78 (Figure 2A). We show by hierarchical cluster analysis that the correlation between averaged and LC was not driven by a few traits (Figure 3) and by randomly sampling the same number of SNPs from each chromosome that it was also not due to longer chromosomes having more SNPs (Figure S6). We also demonstrate that the estimates of on longer chromosomes were more variable than those on shorter chromosomes (Figure S7). We further took the weighted average of the estimates of across traits by , which is defined as the proportion of genetic variance attributed to each chromosome, and plotted it against the proportion of the genome represented by each chromosome (LC/L, with L being the total length of the genome) (Figure 2B). The regression slope of the proportion of the genetic variance attributed to each chromosome on the proportion of the genome represented by each chromosome was 0.875 with a standard error (SE) of 0.150 which was not significantly different from 1 (P = 0.413), and the intercept was 0.008 (SE = 0.007) which was not significantly different from zero (P = 0.289), suggesting that on average 1% of the genome approximately explains 1% of the genetic variance. Despite that there are SNPs with large effects for some traits (Figure S8), all these results are consistent with that many genetic variants each with a small effect widely spread across the whole genome.
In addition, we partitioned into the contributions of genic () and intergenic () regions of the whole genome (Materials & Methods) and averaged the estimates of and across all the traits. The result shows that SNPs in genic regions explain disproportionally more variation than those in intergenic regions (Table S4). We further estimated the variance explained by the genic () and intergenic () regions of each chromosome and again averaged the estimates of and across all traits. The numbers of genic and intergenic SNPs on each chromosome are presented in Table S5. We show that the variance explained by the genic (intergenic) regions on each chromosome is also proportional to the total length of the genic (intergenic) regions (Figure 4).
Previous studies using the whole-genome estimation approach [10], [14] have shown that common SNPs explain a large proportion of heritability for traits and diseases such as height [10], [11], BMI [11], cognition ability [16], [17], rheumatoid arthritis [13] and schizophrenia [12]. The reason why GWAS have not yet identified all the common SNPs that explain this amount of variation is mainly because there are many of them each with an effect too small to pass the stringent genome-wide significance level. However, each of these studies focused only on one or a few diseases or traits. We estimated and partitioned the genetic variance that tagged by all common SNPs for 49 traits in an eastern Asian population and showed by a number of analyses that polygenic inheritance is ubiquitous for most human complex traits.
The estimates of for 6 traits, however, were not different from zero at the nominal significance level (0.05) and the estimates for two insulin related traits INS0 (fasting blood insulin level) and HOMA (homoeostasis model assessment for insulin resistance) were constrained at zero in the analysis because the estimates were converged at small negative values during the estimation process. It does not necessarily mean that common SNPs do not explain any genetic variance for INS0 and HOMA. It could mean that for the two traits are small and their estimates approached zero just because of random sampling. For example, if the true parameter of for a trait is 0.05, given a SE of 0.04 (similar magnitude as those presented in Table 1), the probability of getting a zero estimate of is approximately 0.11, meaning that it is not surprising to observe a few zero estimates from an analysis of 49 estimates if the true parameters of for these traits have a spectrum from moderate to small values.
The estimate of for height was 31.6% (SE = 4.6%), which was smaller than the estimate from a study in Australians ( = 44.9%, SE = 8.3%) [10] but not statistically significant (P = 0.161), and was significantly (P = 0.015) smaller than the estimate from another study in European Americans ( = 44.8%, SE = 2.9%) [11]. There could be two possible reasons: 1) there is a difference in heritability for height between Koreans and Europeans and 2) the tagging of Affymetrix 5.0 array is not as good as the later version Affymetrix 6.0 and the Illumina HumanCNV370 arrays used in the previous studies in Europeans. The estimate for BMI ( = 14.7%, SE = 4.1%) was also slightly smaller than that in European Americans ( = 16.5%, SE = 2.9%) [11] but the difference was not significant (P = 0.741). We estimated the narrow-sense heritability for 11 traits by from a family study in Koreans (Text S1 and Table S6). The estimate of heritability either for height (h2 = 0.744, SE = 0.048) or for BMI (h2 = 0.478, SE = 0.057) in Koreans was comparable to that estimated in Europeans. We then estimated the variance explained by all SNPs on Affymetrix 5.0 array in the sample of 11,586 unrelated European Americans as used in [11] (Text S1). The estimate of variance explained by all SNPs on Affymetrix 5.0 array in European Americans was 0.394 (SE = 0.027) for height, which was not significantly different from that estimated in this study (P = 0.118). Therefore, the difference between the estimate of in this study and in previous studies is partly due to the use of different types of SNP genotyping arrays and partly due to sampling error.
It is demonstrated by the genome partitioning analysis that there was a strong linear relationship between the estimates of variance explained by individual chromosomes and chromosome length (Figure 2). The correlation between variance explained and DNA length was stronger in the intergenic regions than that in the genic regions if we define the genic region as ±0 Kb or ±20 Kb of UTRs, while it was stronger in the genic regions than that in the intergenic regions if we define the genic region as ±50 Kb of UTRs (Figure 4). We show by a number of analyses that the result was driven neither by the difference between the number of SNPs in genic regions and in intergenic regions nor by the difference in MAF distribution between genic and intergeinc SNPs (Text S2). If trait-associated genetic variants are enriched in functional elements such as introns and UTRs and diluted in exons, the relationship between the variance explain and DNA length will be attenuated in the genic region. However, this could also be just due to sampling. The sampling variance of a regression R2 is approximately 4ρ2(1−ρ2)/N where E(R2) = ρ2 and N is number of observations (number of chromosomes in this case). Given ρ2 = 0.5 and N = 22, the SE of the regression R2 is ∼0.2. Therefore, the difference between the correlation (between the variance explained and DNA length) in genic regions and that in intergenic regions is unlikely to be significant. In addition, in the partitioning analysis of intergenic regions, chromosome 2 seems to be an outlier (Figure 4). For example, for the definition of genic region of ±50 Kb, the variance explained by the intergenic regions on chromosome 2 averaged across 47 traits was 0.68% (SE = ∼0.16%), which was 0.25% larger than the expected value from the fitted line. Given the SE of ∼0.16%, the difference was, however, not greater than what we would expect by chance (P = 0.118).
Moreover, we attempted to investigate the enrichment of genetic variants in genes involved in biological pathways. For any particular trait, there are a number of biological pathways that are important to the trait development. We chose the well-known insulin signal transduction pathway as an example to demonstrate the use of GCTA to partition the genetic variance based on functional annotations. We took SNPs that are ±20 kb away from 103 genes that are involved in insulin signaling pathway. There were 955 SNPs which covered ∼0.45% of the genome. We then performed the genome partitioning analysis to decompose into two components, i.e. the contribution of the genes involved in insulin pathway and that of the rest of the genome for 11 lipids and diabetes related traits. As shown in Table S7, we did not find any evidence that genes involved in insulin pathway explained disproportionally more proportion of variance. This is not surprising because these gene regions cover ∼0.45% of the genome and the SE of the estimate was ∼0.3% so that even if there is an enrichment of genetic variants in these gene regions, it is unable to be detected due to the lack of power. Larger sample size is required for such kind of analysis in the future.
In conclusion, we showed by whole genome estimation and partitioning analyses that, most human complex traits, if not all, appear to be highly polygenic, i.e. there are a large number of genetic variants segregating in the population with a small effect widely distributed across the whole genome. All the common SNPs on the Affymetrix 5.0 array explain approximately a third of heritability on average over all the 49 traits analysed in this study. The remaining unexplained two thirds of heritability could be due to causal variants including the common and rare ones that are not well tagged by SNPs on the array or possibly due to the heritability was over-estimated in the family/twin studies. The conclusion drawn from previous studies that heritability is not missing but due to many variants with small effects is not specific for human height in European populations but likely to be in common for most human complex traits and populations. Taken all together, it implies that although whole genome sequencing data will provide much denser genomic coverage than the current genotyping array and will therefore identify more associated variants and explain more genetic variance, large sample size is still essential.
This study used the data from the Korea Association Resource (KARE) project, which has been described elsewhere [15]. In brief, there were 10,038 individuals recruited from two community-based cohorts, 5,018 from Ansung and 5,020 from Ansan, in Gyeonggi Province, South Korea. The individuals were aged from 40 to 69 years old and born in 1931 to 1963. All the individuals were measured for a range of quantitative traits through epidemiological surveys, physical examinations and laboratory tests, including traits related to obesity, blood condition, pulse, bone mineral density, lipids, diabetes index, liver functions, lung functions and kidney functions. A description of the 49 traits used in this study is summarized in Table S1. We adjusted the phenotypes of each trait for age by simple regression and then standardized the residuals to z-scores, in each of the two cohorts (Ansung and Ansan) and in each gender group separately.
The genomic DNAs were isolated from peripheral blood drawn from the participants and were genotyped with 500,568 SNPs on the Affymetrix 5.0 genotyping array [15]. We excluded the SNPs with missingness rate >5%, minor allele frequency (MAF)<0.01, and Hardy-Weinberg equilibrium (HWE) test P value<10−6 using PLINK [18], and retained 326,262 autosomal SNPs for further analysis. The KARE GWAS data had been imputed to HapMap2 CHB and JPT panels [19]. After removing SNPs with MAF<0.01 and SNP missing rate >0.05, there were 2,153,258 genotyped/imputed SNPs [15].
We estimated the genetic relationship matrix (GRM) between all pairs of individuals from all the genotyped SNPs and excluded one of each pair of individuals with estimated relationship >0.025 retaining 7,170 unrelated individuals. For each trait, we then estimated the variance that can be captured by all SNPs using the restricted maximum likelihood (REML) approach in mixed linear model , where y is a vector of phenotypes, b is a vector of fixed effects with its incidence matrix X, is a vector of aggregate effects of all SNPs, and with AG being the SNP-derived GRM and being the additive genetic variance. The proportion of variance explained by all SNPs is defined as with being the phenotypic variance. Details of the model and parameter estimation have been described elsewhere [10], [14]. In addition, using the same method as above but allowing to fit multiple genetic components simultaneously in the model, we partitioned into the contributions of genic () and intergenic () regions of the whole genome [11] and averaged the estimates of and across all the traits. The genic regions were defined as ±0 kb, ±20 kb and ±50 kb of the 3′ and 5′ UTRs. A total of 135,491, 175,637 and 205,901 SNPs were located within the boundaries of 12,310, 15,140 and 15,274 protein-coding genes for the three definitions (±0 kb, ±20 kb and ±50 kb), respectively, which covered 36.1%, 49.2% and 58.9% of the genome.
|
10.1371/journal.pntd.0003179 | Cutaneous Manifestations of Spotted Fever Rickettsial Infections in the Central Province of Sri Lanka: A Descriptive Study | Characteristic skin lesions play a key role in clinical diagnosis of spotted fever group rickettsioses and this study describes these cutaneous manifestations along with basic histological features.
Study was conducted at Medical Unit, Teaching Hospital, Peradeniya, from November 2009 to October 2011, where a prospective data base of all rickettsial infections is maintained. Confirmation of diagnosis was made when IgM and IgG immunofluorescent antibody titre of 1/32 and >1/256 respectively. Of the 210 clinical cases, 134 had cutoff antibody titers for Rickettsia conorii antigen for confirmation. All these 134 patients had fever and skin rash, and of them 132(98%) had discrete maculopapular rash while eight (6%) had fern leaf type skin necrosis. Eight patients (6%) had healed tick bite marks. Average size of a skin lesion was 5 mm and rash involved 52% of body surface, distributed mainly in limbs and back of the chest. Generally the facial and leg skin was slightly oedematous particularly in old aged patients. Sixteen patients (12%) had pain and swelling of ankle joints where swelling extended to feet and leg. Biopsies from skin rash of six patients showed evidence of cutaneous vasculitis and of them, 247 bp region of the 17-kDa spotted fever group specific protein antigen was amplified using PCR.
A discrete maculopapular rash and occasional variations such as fern leaf shape necrosis and arthritis are found in spotted fever group. Histology found vasculitis as the pathology of these lesions.
| Rickettsial organisms infect humans causing a wider array of clinical features and have re-emerged in Sri Lanka where three known disease entities; spotted fever group, murine typhus and scrub typhus do exist. These diseases cause clinical illnesses varying from mild febrile illness to severe multiple organ involvement even leading to fatal outcomes when there is a delay in diagnosis. Occasionally, clinical features could be nonspecific or atypical. Nevertheless, detection of skin lesions mostly facilitates the clinical diagnosis. Hence, clinicians need to be familiar with common as well as uncommon variations of skin manifestations. Being a treatable infection, early diagnosis is important and is heavily based on clinical features in settings where laboratory diagnostics are unavailable; at the same time delaying of treatment could lead to high morbidity and mortality. We identified some important variations of skin lesions associated with spotted fever group rickettsial infections which include fern leaf type skin necrosis mainly involving superficial skin with blackish discoloration which dries up with time and peels off. In addition, mild cutaneous oedema is seen over the face and ankles especially in older patients. Acute arthritis involving ankle joints were common manifestations which together with typical skin lesions facilitate the clinical diagnosis.
| Rickettsiae are a group of alpha-proteobacteria found as an obligatory intracellular parasite of eukaryotic cells [1]. Rickettsia cause human infections giving rise to a wider array of clinical features. Rickettsial infections have re-emerged in Sri Lanka where three known disease entities namely spotted fever group (SFG), murine typhus and scrub typhus are being reported from different parts of the island [2]–[4]. Disease spectrum varies mainly depending on the rickettsial species that causes the disease, for example the Rocky Mountain spotted fever (RMSF) caused by Rickettsia rickettsii is known to be the most severe form of tick borne rickettsioses around the globe [1]. Clinical illness may vary from mild to severe with multiple organ involvement, sometimes leading to fatal outcomes [5], [6]. Generally, clinical features of the infection could be nonspecific or atypical. Nevertheless, the presence of cutanoeus lesions facilitates the clinical diagnosis of the infection. These include eschars, skin eruptions and rash with patchy necrosis [2], [4], [7]. Further, it is important to be familiar with the common cutaneous manifestations as well as uncommon variations of skin lesions. Being a treatable infection, early diagnosis is heavily based on clinical features in settings where laboratory diagnostics are not available and at the same time delaying of treatment could lead to high morbidity and mortality [8], [9]. Of the clinical features, cutaneous lesions play a major role that supports the diagnosis. However, these cutaneous lesions tend to have varying patterns influenced by many factors. Thus, clinicians need to get used to these variations to make a presumptive diagnosis of rickettsial infection. Moreover, identifying pathological changes of skin lesions are important as supportive tools in verifying the clinical diagnosis and also to understand the nature of the pathology caused by rickettsiae. The basic pathological changes have been described previously in other regions of the globe [10]. The aims of this study were to describe the morphology of cutaneous manifestations and their basic histological features of spotted fever rickettsial infections in Sri Lanka.
Patient recruitment and sample collection for the study were done in the Medical Unit, Teaching Hospital, Peradeniya from November 2009 to October 2011. This study was conducted according to the Declaration of Helsinki with approval from the Ethics review committee, Faculty of Medicine, University of Peradeniya, Sri Lanka. Informed written consent was obtained from all the adult patients and from guardians on behalf of the minors enrolled in the study.
All patients admitted to the unit during the study period, with clinical features suggestive of rickettsial infection, were included in the study. Clinical case definition was based on the presence of fever for more than five days, associated skin rash and rapid defervescence with an anti-rickettsial antibiotic treatment [2], [4]. Patients were interviewed and examined, and followed up while in the hospital and after discharge. All clinical details were recorded on individual formatted data sheets, after obtaining the informed consent. Clinical details included the history, general examination and systemic examination findings. The details of the skin rash such as its distribution, size, and shape, presence of eschar and colour variations were recorded in a printed figure of the human body in the data sheet. The human body surface was divided into eight zones as, head and neck, anterior trunk, posterior trunk, arms and forearms, palms, thigh and gluteal regions, legs and soles. A score adopted from a locally devised method was used for each patient to indicate the extent of the skin rash [2]. A single score was given to each zone and the total extent of skin involvement was calculated as a percentage of skin finally.
In 2012, we saw two patients with extensive or unusual cutaneous manifestations who qualified for the diagnosis of rickettsial infections and included in this study as additional cases.
Confirmation of diagnosis was made by positive serology using immunofluorescent antibody assays (IFA) and polymerase chain reaction (PCR). IFA test was carried out to detect antibodies against three groups of rickettsial antigens. Frozen antigen pellets of Rickettsia conorii (Strain Malish) of spotted fever group, Orientia tsutsugamushi (Strain Karp) of scrub typhus and Rickettsia typhi (Strain Wilmington) of murine typhus were used for slide preparation and the antigens were obtained from WHO Reference center for Rickettsial & Bartonella Associated diseases, CDC, Atlanta, USA. The final diagnosis of rickettsial infection was defined on the basis of their clinical criteria and the presence of specific IFA IgM and IgG seropositivity. Baseline cutoff titre value for IFA testing was 1/32, for both IgG and IgM. Samples with cutoff titre of IgG were further tested to obtain individual end point titres. Final serodiagnosis was based on the detection of IgM seropositivity and the IgG tire >1/256. The IgG titre above 1/256 has been used in published data in Sri Lanka where attempts have been made to validate this level as the standard cutoff value. Previous literature on serological values were used for the final disease confirmation despite some constraints [2], [11]–[13].
Cutaneous biopsy samples were obtained from six patients from areas with definitive maculopapular or vasculitic rash. A covered part of the body with the skin lesion was selected and the biopsy specimen was obtained under local anesthesia using a punch biopsy needle. Formalin fixed samples were processed and microscopic slides were prepared. Sections were stained with hematoxylin and eosin and were examined under microscope at magnifications of ×100, ×400, ×1000.
DNA was extracted from skin biopsy samples using QIAGEN spin column (Qiagen Sciences, Maryland 20874, USA) kit. Nested polymerase chain reaction (nPCR) assay was performed on extracted DNA to amplify the 17-kDa antigen gene. The primers used for primary and nested PCRs were R17-122, forward (5′- CAG AGT GCT ATG AAC AAA CAA GG-3′); R17-500, reverse (5′- CTT GCC ATT GCC CAT CAG GTT G -3′) and TZ 15, forward; (5′- TTC TCA ATT CGG TAA GGG C -3′) TZ 16, reverse (5′ - ATA TTG ACC AGT GCT ATT TC - 3′) respectively. DNA products were electrophoresed at 100 V. The eletrophoresed gel was observed under UV light (Gel documentation system, Vilber Lourmat, 77202 Marne la Vallee, France).
Individual data points were stored in a computerized data base (Excel,Microsoft) and the basic descriptive analyses were done by means measures of central tendency. The data were analyzed by Minitab, version 14.0, Minitab Inc., USA.
A total of 210 patients qualified for the clinical diagnosis of rickettsial infections during the two year study period and their serum samples were tested for IgG and IgM antibodies for SFG, murine typhus group and scrub typhus groups of rickettsial infections. The average time gap between the onset of the illness and acute phase antibody testing was seven days. Ninety percent of the total (n = 188) showed positive titre above 1/32 of IgM and/or IgG to SFG. None of the patients had a positive IgG or IgM seroreativity for Typhus and Scrub Typhus groups alone. One patient showed a positive IgG seroreativity for both scrub typhus and SFG. Of the SFG, 164 (78%) had positive IgM seroreactivity and 188 (90%) had positive IgG seroreactivity. Of the IgG positive group, 144 patients had IgG titres >1/256. Moreover, both the IgM and IgG seropositivity was seen among 150 (71%) with 38 (18%) being positive only for IgG and 14 (7%) being positive only for IgM. Negative seroreactivity was observed in 22 (10%) patients of whom none of the serological tests were positive in acute sera. Of the total, 134 had high titers of IgM and IgG (IgG titre >1/256 and IgM titre >1/32) in the acute sera tested with IFA for Rickettsia conorii antigen (Table 1). This included 69 (52%) males and 65 (48%) females. Mean age of the group was 44 years (range, 12–84). Skin colour of the patients varied from fair to dark complexion with majority being dark skinned people.
All 134 patients with high antibody titers for SFG had fever and skin rash. Average duration of fever on admission was six days (range, 2–21days) and that of skin rash was two days (range, 1–7 days). Of this group, 119 (89%) had developed skin rash following the onset of fever and the average time gap between the appearance of these two symptoms was three days. The rest, 15 (11%) patients had developed fever and skin rash together.
Of the cutaneous lesions, eight (6%) patients had healed tick bite marks apart from the skin rash (Figure 1). The skin rash was maculopapular in 132 (98.5%), macular in one patient and papular in another one. Eight patients (6%) of the group had fern leaf type skin necrosis (Figure 2) and the rest had erythematous rash (Figure 3), (Table 2). Mean age of the patients who had necrotic skin rash was 64 years. Appearance of the erythematous lesions varied depending on the innate skin colour of the subjects. Lesions were distinctively erythematous in fair skinned patients whereas those were more of dusky red in dark skinned patients. Furthermore, in some patients the rash was visible only if the skin was visualized from an angle in the daylight. Average size of a skin lesion was 5 mm and it ranged from 2–10 mm. The rash was always discrete with normal looking skin in between and the shape of the lesions varied. Commonly they were either ovular or round in shape. Necrotic rash had a fern leaf like pattern. These lesions took time to heal with black discoloration leaving pinkish base once peeled off. In general, the facial and leg skin of the patient seemed slightly oedematous particularly in old aged patients. Average extent of the skin involvement with the rash was 52% (range, 12.5–100%). Out of the eight body regions considered, majority of patients had the skin rash in their arms and forearms (n = 108, 81%) and legs (n = 90, 67%) while other areas involved were palms (74, 55%), soles (75, 56%), anterior trunk (61, 46%), thighs and gluteal region (53, 40%), head and neck (45, 34%) and posterior trunk (44, 33%), (Table 3). Apart from the skin rash, old aged patients had cutaneous oedema around the ankles (Figure 4) and puffiness of face with the skin becoming dusky discolouration. Sixteen (12%) patients had pain and swelling of ankle joints where mild oedema extended to feet and mid leg suggestive of acute arthritis.
In 2012, a 65-year-old man presented with fever of 7 days and erythematous macular rash which progressed to patches of skin necrosis and gangrene of fern leaf shape mainly involving limbs. He had scrotal swelling and gangrene of scrotal skin (Figure 5). During the same period, a 64-year-old man presented with the similar history and developed dark brown skin lesions and vasculitic rash in toes and plantar surface of the feet (Figure 6). Both patients had positive titre of IFA for Rickettsia conorii antigens. They made recovery with doxycycline.
All six skin biopsy samples examined for histopathological characteristics showed evidence of cutaneous vasculitis. Main histopathological features were foci of basal cell vacuolation with exocytosis and lymphocytic infiltrates in perivascular spaces. Further, there were ectatic upper dermal blood vessels, focal swelling of endothelium, fibrinoid necrosis of vessel walls, extravasated red cells and presence of fibrin thrombi (Figure 7). The 247 bp region of the 17-kDa spotted fever group specific protein antigen was successfully amplified using PCR further confirming the presence of spotted fever among those patients.
This study describes both external morphology and histopathological features of cutaneous manifestations of spotted fever rickettsial infections in Sri Lanka where the causative rickettsial pathogens are yet unknown. All cases in the study group qualified for the diagnosis of spotted fever group rickettsioses, as the seroreactivity was noted only against R. conorii antigen, except in one patient who was positive both SFG and scrub typhus. Serological evidence with high antibody titres in acute serum samples alone has been considered confirmative based on previous studies [2], [3], [13]. However, rising titre in convalescent sera would have been the best. Due to practical difficulties, majority of patients failed to visit the clinic to give convalescent blood sample in two weeks after being discharged from the wards. Convalescent serum samples were tested only in six patients and all six showed a fourfold rise in antibody titre for SFG. The average duration between the acute and convalescent serum sample collection was two weeks. Moreover, we were aware of the concerns about serology such as IgM positivity that indicates a recent infection and is not necessarily the acute cause of fever. Further it is often found not very early during the course of infection and could give false positive results in other infections such as Epstein Barr viral infection. Also IgG positivity needs IgM or any other specific disease manifestations for confirmation. We used clinical case definition and high titre serology to overcome these issues.
None of the patients in this study group had classical eschars and eight patients had only possible tick bite marks. Classic eschars with thick and dry necrotic tissue have been described in some SFG rickettsioses such as Mediterranian spotted fever (MSF), African tick bite fever and rickettsial pox [14], [15]. Predominance of maculopapular erythematous rash in SFG has been reported earlier and the fern leaf type skin necrosis has not often been described other than in few studies done in the same region in Sri Lanka [2], [4]. However, there are reports of gangrene of digits, ear lobes, scrotum, nose or limbs occurring in rickettsial infections such as RMSF [16], similar to two cases presented in 2012. Even though, the skin rash had a wide distribution in different regions of the body, majority had dominant involvement of arms, forearms and legs. Although the involvement of palms and soles are characteristic in rickettsial infections, only about 56% of this group had the rash involving palms and soles. Moreover, the study describes some important general characteristics basically seen in old aged patients. These include cutaneous oedema around the ankles and puffiness of face with the skin becoming dusky discolouration. Importantly, such focused descriptions are not found in earlier studies. Even though, majority of patients did not come for follow up, our experience suggests that cutaneous rashes take 2–3 weeks to fade off after defervescence. Furthermore, rash negative forms of rickettsial cases have also been described from Sri Lanka amounting to a smaller percentage [18].
Clinical features of SFG of Sri Lanka described in the current study do differ from the presenting features of SFG in some regions of the globe. RMSF is a well known SFG in America, caused by R.rickettsii, used to present with high fever, headache, myalgia and skin rash as the predominant clinical manifestations. However eschars are rare in this disease and skin necrosis has been reported [17], [19]. Generally the maculopapular skin rash is mainly distributed in the limbs. Commonest SFG rickettsioses reported in Southern Europe and Northern Africa is MSF caused by R.conorii. In contrast to SFG in Sri Lanka and RMSF, eschars are common in MSF where the distribution of maculopapular or petichial skin rash in palms and soles is remarkable. In addition to these, R. japonica causes Japanese spotted fever mainly in East Asia. Major clinical manifestations of Japanese spotted fever are headache, fever, skin rash and eschars [1]. There are about 18 different SFG agents identified from different regions of the globe with variations in clinical picture and mode of transmission. This vast diversity can mainly be attributed to geographical factors, socio-economic factors and vector-human relationships [1], [14]. The cutaneous histopathology of SFG rickettsioses is caused by endothelial damage by the rickettsial organisms. As per the previous findings the basic histopathological changes that can be seen in skin eruptions and patchy necrotic lesions include lymphohistiocytic capillaritis and venulitis, perivascular oedema, erythrocyte extravasation, interstitial infiltrate, leukocytoclastic vasculitis and presence of nuclear dust. There can also be epidermal changes including basal layer vacuolar degeneration with mild dermoepidermal interface lymphocytic exocytosis, focal fibrin thrombi, capillary wall necrosis and perivascular inflammation [10]. Most of these were evident in this study and besides there were small vessel ectasis and focal swelling of endothelium which are also supportive of the vasculitis. This is the first time a study described histopathological features of the cutaneous lesions of rickettsioses in Sri Lanka. Furthermore, use of immunohistochemical stains may have identified the causative rickettsial organisms in the skin lesions.
Pattern of cutaneous manifestations have a pivotal role in making a clinical diagnosis of SFG. The typical features of the skin rash include discrete maculopapular lesions with dusky erythemtous hue, distributed mainly in the limbs, back of the chest, anterior abdomen and soles. Variations are common, such as fern leaf pattern of skin necrosis mainly involving superficial skin with blackish discoloration which with time dries up and peels off. Mild cutaneous oedema is common over the ankles and face specially in older patients. Histological features like disruption of small vessels by inflammatory cells, deposition of fibrin thrombi within the lumen and leukocytic infiltrates in perivascular spaces suggest vasculitis caused by infecting organism.
|
10.1371/journal.pcbi.1002411 | Network Evolution: Rewiring and Signatures of Conservation in Signaling | The analysis of network evolution has been hampered by limited availability of protein interaction data for different organisms. In this study, we investigate evolutionary mechanisms in Src Homology 3 (SH3) domain and kinase interaction networks using high-resolution specificity profiles. We constructed and examined networks for 23 fungal species ranging from Saccharomyces cerevisiae to Schizosaccharomyces pombe. We quantify rates of different rewiring mechanisms and show that interaction change through binding site evolution is faster than through gene gain or loss. We found that SH3 interactions evolve swiftly, at rates similar to those found in phosphoregulation evolution. Importantly, we show that interaction changes are sufficiently rapid to exhibit saturation phenomena at the observed timescales. Finally, focusing on the SH3 interaction network, we observe extensive clustering of binding sites on target proteins by SH3 domains and a strong correlation between the number of domains that bind a target protein (target in-degree) and interaction conservation. The relationship between in-degree and interaction conservation is driven by two different effects, namely the number of clusters that correspond to interaction interfaces and the number of domains that bind to each cluster leads to sequence specific conservation, which in turn results in interaction conservation. In summary, we uncover several network evolution mechanisms likely to generalize across peptide recognition modules.
| Protein interaction networks control virtually all cellular processes. The rules governing their evolution have remained elusive, as comprehensive protein interaction data is available for only a small number of very distant species, making evolutionary network studies difficult. Here we attempt to overcome this limitation by computationally constructing protein interaction networks for 23 relatively tightly spaced yeast species. We focus on networks consisting of kinase and peptide binding domain interactions, which play central roles in signaling pathways. These networks enable us to investigate evolutionary network mechanisms. We are able, for the first time, to accurately quantify the contribution of different rewiring mechanisms. Interaction change appears to be mainly accomplished through binding site evolution rather than through gene gain or loss. This is in contrast to other evolutionary processes, where gene duplication or deletion is a major driving factor. Moreover, our analysis reveals that interaction changes are very fast – fast enough that the number of changes saturates, i.e., the actual rate of change has been strongly underestimated in previous studies. Our analysis also reveals different mechanisms by which certain interactions are conserved throughout evolution. Our results likely transfer to other species and networks, and will benefit future evolutionary studies of signaling pathways.
| Peptide recognition modules (PRMs) and kinase domains bind short linear peptide motifs on their protein binding partners and are integral members of many signaling pathways [1]–[3]. PRM members include the SH3 (Src homology 3), SH2 (Src homology 2), and PDZ (PSD-95/Discs-large/ZO-1) domains [1]. In this study, we focus on the SH3 domain, a small (∼60 amino acids) domain, implicated in crucial regulatory processes such as signal transduction, cytoskeleton organization, and cell polarization [4], [5]. SH3 domains typically bind short proline rich peptides containing a PxxP binding motif [5]. Initial structural analysis revealed two main binding classes, although variations to these canonical SH3 binding motifs have also been discovered.
Experimental identification of the short peptide binding motifs recognized by PRMs has been performed using a number of methods, such as synthetic peptide arrays (SPOT), oriented peptide array libraries (OPAL), protein domain microarrays, and phage display [1], [6]–[10]. Binding specificity maps have been generated for Saccharomyces cerevisiae SH3 and kinase domains using phage display and combinatorial peptide library screening approaches, respectively [6], [11]. Domain binding specificities from these experimental methods are captured in position weight matrices (PWMs) enabling comprehensive and high confidence predictions of physical interactions involving SH3 and kinase domains. The high accuracy of these PWM predictions has been demonstrated in their ability to recapitulate interactions derived from orthogonal experimental methods such as yeast two-hybrid [6].
Interaction network studies have uncovered network properties such as scale-free and hierarchical topologies [12], resulting in the development of models describing protein interaction network evolution [12]–[15]. Network rewiring rates for protein interaction networks have also been established for proteins in S. cerevisiae that have paralogs [16], model eukaryotic protein interaction networks [17], and yeast regulatory networks [18]–[20]. A global comparative analysis on network rewiring from existing experimental datasets has suggested that regulatory networks are among the fastest evolving biological networks [21]. However, these comparative studies are hampered by two problems: The analyzed networks are often incomplete and the species examined are highly diverged. An obvious problem is that interactions in species similar to the model organisms (such as yeast or worm) are usually inferred by means of orthology mapping, which prohibits any kind of evolutionary analysis based on them [22], [23]. Unlike these mapping methods, predicting interactions via PWMs enables generation of interaction networks in a species-independent manner. This permits a more accurate means to identify both conserved and diverged interactions, as interactions are determined by a domain's binding specificity and not by orthologous protein pairs, thus enabling the evaluation of different evolutionary mechanisms that give rise to the observed rates of network rewiring.
In this study, we use the aforementioned high resolution S. cerevisiae SH3 and kinase specificity maps [6], [11] to computationally predict high confidence SH3 and kinase interaction networks for 23 species belonging to the Ascomycota phylum of the Fungal kingdom, representing over 300 million years of evolution [24]. We quantify network evolution rates for different network rewiring mechanisms and compare them to other network evolutionary rates [18]. Furthermore, we show that the rate of network rewiring reaches saturation due to the rapid rate of interaction change. Moreover, we uncover interaction conservation patterns related to multiple SH3 domains binding the same proline rich region on a protein binding partner. Finally, we show motif specific sequence conservation translates to the conservation of interactions.
Using position weight matrix (PWM) profiles derived from S. cerevisiae phage display experiments and combinatorial peptide library screens [6], [11], we constructed SH3 and kinase interaction networks in 23 different fungal species spanning over 300 million years of evolution (Figure 1, Materials and Methods). Note that our methodology does not rely on sequence homology to predict interactions in different species, enabling the identification of species-specific interactions through binding profiles. Sequence homology is used only in the comparison of interaction networks between different species. The SH3 predicted network, using 30 PWMs from S. cerevisiae resulted in ∼800 interactions and ∼400 unique proteins for each of the 23 yeast species. Likewise, the kinase network, using 63 PWMs, resulted in ∼1800 interactions and ∼450 unique proteins. Parameters to create the networks were selected using the area under the receiver operator curve (AUROC) and the Matthews correlation coefficient metrics (Materials and Methods). In Figure S1 we provide estimates over a range of true positive and true negative ratios.
To compare the constructed networks and infer interaction conservation among the different yeast species, orthology assignments provided by Wapinski and co-workers were used to establish orthology relationships for all proteins in the networks (Materials and Methods) [25]. We further ensured that orthologs to the S. cerevisiae SH3 and kinase proteins contained the particular SH3 and kinase domains, respectively. Here, we assume domain binding specificities found in S. cerevisiae to be similar for orthologous proteins, even for distant species. While this is a critical assumption, five different observations suggest it is reasonable for the proteins analyzed. First, we examined paralogous domains in S. cerevisiae since they tend to be under weaker purifying selection than non-duplicated genes and are likely to diverge at faster rates than proteins in different species [26]. We find paralogous SH3 proteins to have very high PWM similarity, especially considering those above 80% amino acid sequence identity (Figure S2). Second, we show orthologs and paralogs share the same amino acids at similar, presumably binding determining, positions in a multiple sequence alignment (Figure S3). Third, SH3 domain crystal structures reveal contact amino acids to a bound ligand are highly conserved in orthologs (Figure S4). Fourth, orthologs to S. cerevisiae SH3 and kinase domains exhibit a high degree of amino acid identity (Figure S5A and S6). Fifth and finally, for many following analysis, we use two species sets: one where we use all 23 species and a restricted set, where we use orthologs with an amino acid sequence identity greater than 80% (for which binding specificity is almost guaranteed to be conserved). In all cases, we observe the same results.
To assess the similarity of the fungal networks with each other, we created phylogenetic trees from the orthology mapped interaction networks, based on the number of conserved interactions between species (Materials and Methods). Importantly, we found that the phylogenetic trees derived from the predicted SH3 and kinase interactions () are remarkably similar to the canonical protein sequence-based phylogeny (Figure 2A), suggesting that the interaction networks share similar evolutionary properties as genome sequences [27]. While we observe similar phylogenetic trees for the fungal species, analysis spanning the 3 domains of life revealed topological differences between the metabolic pathway and sequence based phylogenetic trees, representing many more years of evolution [28]–[30]. Despite the phylogenetic similarities for the fungal species, only 5 SH3 interactions are conserved across all 23 yeast species, translating to ∼1% of all S. cerevisiae SH3 interactions. For kinases, not a single interaction is conserved across all kinase interaction networks given the defined thresholds and orthology mappings. The limited number of globally conserved interactions indicates that phylogenetic similarities are due to conservation of the network structure and topology rather than individual interactions between orthologs.
The observed lack of conservation of specific interactions may be attributed to two main evolutionary mechanisms. First, an interaction can be lost or gained but both the binding and target proteins are evolutionary retained, which we refer to as “interaction rewiring” (i.e., orthologs exist for the associated genes of the binding and target proteins). Second, the binding or target protein itself can be gained or lost (i.e. through a gene duplication), which we refer to as a “protein change”. To examine interaction differences observed in these networks, we quantified rates of interaction rewiring and protein change in both the SH3 and kinase networks and compared them to previously reported rates for other types of interaction networks.
To attain a rate of interaction rewiring, we need to ascertain the number of interaction changes between two species as well as their divergence time (see Materials and Methods). We find that the interaction rewiring rate decreases sharply as the divergence distance increases for the SH3 and kinase interaction networks (Figure 2B, Figure S8A) in addition to being slower than in random networks (p-value<0.001, Figure S9). This suggests that to obtain an accurate rewiring rate one has to use closely related species, as the estimated rewiring rates are dependent on divergence distance and hence the selected reference species. For example, the rates of interaction rewiring in the SH3 network between closely related species 1) S. cerevisiae and S. paradoxus and between distant species 2) S. cerevisiae and S. pombe are 3.86×10−4 and 4.01×10−5 interaction rewirings per protein pair per million years respectively. Indeed, using the nearest evolutionary species as a reference, S. octosporus, for S. pombe results in an over 4 times increase to 1.85×10−4 interaction rewiring per protein per million years (Table S1). Intuitively, this phenomenon may be explained by a saturation of interaction changes at longer evolutionary distances. A similar result is observed when selecting SH3 domains above a range of amino acid identity thresholds (Table S2, Figure S10), indicating that our assumption of orthologs retaining similar specificities is reasonable, as noted above.
To investigate this saturation effect, we examined the absolute number of interaction changes with respect to divergence time. Importantly, we find the number of SH3 and kinase interaction rewiring events is sufficiently rapid to reach saturation in about 200 million years of divergence since the last common ancestor (Figure 2C, Figure S8B). This saturation effect is analogous to the saturation of neutral sequence substitutions in the comparison of highly divergent sequences resulting in a biased dN/dS ratio, due to the inability to observe multiple substitutions at the same nucleotide positions thus resulting in a deflated dS value. The observed decrease in interaction rewiring rate is dominated by divergence distance and possibly reflects the inability to observe multiple rewirings of the same interaction, as the loss followed by a gain of the same interaction appears as a conserved interaction (Table S1). Previous studies that quantified evolutionary rates for network changes [18], [19] used relatively distant species for comparison and were hence likely hampered by this issue. Thus we expect that they have underestimated the true evolutionary rate.
Similar to interaction rewiring, the rate of protein change is also dependent on the selected reference species. Using the nearest species to calculate rates in the SH3 network, the average rate of protein change is 1.95×10−5 protein changes per protein pair per million years while the average rate of interaction rewiring is 1.61×10−4 interaction rewiring per protein pair per million years (Table S1 and S3). An almost 10-fold difference between rates in interaction rewiring and protein change suggests that interaction rewiring is the primary mechanism for determining the overall rate of interaction change in the SH3 interaction network. A similar result is observed in the kinase network (Table S2 and S4). SH3 and kinase signaling interactions are known to evolve more rapidly than metabolic and other protein interactions [17], [21], due to the short protein binding surface area forming the interaction interface. Additionally, the time scale to observe the saturation phenomena in these networks is significantly smaller than the time scales used in previous comparison studies [17], [21]–[23]. Here we find the dominant mechanism for interaction change, at the observed time scale, is by interaction rewiring for these signaling networks. This is reminiscent of the study by Zhong et al. who found the mechanisms of interaction rewiring and protein change corresponded to distinct mutation events. They further found single genes associated with multiple diseases could be explained by interaction rewiring (ie. network perturbation) [31]. Thus the rapid rewiring rate exhibited in the SH3 and kinase networks conceivably enables the discovery of new functionality while maintaining the same gene repertoire.
A previous study examined the rate of interaction change (i.e. the combination of interaction rewiring and protein change rates) associated with phosphosite and transcription factor regulatory networks [18]. We therefore compared rates for these networks to rates for our SH3 and kinase interaction networks. To provide an unbiased comparison, by compensating for the aforementioned saturation phenomenon, we computed interaction change rates of our kinase network using S. cerevisae as the sole reference strain instead of the closest related species and observed similar rates between the two networks. Specifically, interactions in the C. albicans and S. pombe phosphosite network were reported to change at a rate of 1.09×10−5 and 1.24×10−5 interaction changes per protein pair per million years, respectively after adjusting to our orthology mappings [18] (Table S5). Importantly, we found that interaction changes in the C. albicans and S. pombe kinase networks occur at a comparable rate, 3.68×10−5 and 3.89×10−5 interaction changes per protein pair per million years respectively (Table S2 and S4), when S. cerevisiae was used as the reference species. Considering that our computational approach is likely to contain some false positives and false negatives, the calculated rates may be overestimated. Thus, the close agreement between our findings and the Beltrao et al phosphosite network study supports the validity of our computational approach.
We next quantified rates in the C. albicans and S. pombe SH3 networks and found that they change at a rate of 5.90×10−5 and 5.47×10−5 interaction changes per protein pair per million years, respectively (Table S1 and S3). Interestingly, interaction changes in the SH3 networks occur at similar rates compared to interaction changes in the kinase and phosphosite networks.
Rates for the transcription factor-DNA (TF-DNA) interactions have also been deduced for S. mikatae and S. bayanus regulatory networks using S. cerevisiae as a reference species. We found that the rates of interaction change associated with these regulatory networks (1.02×10−3 and 5.91×10−4 interaction changes per transcription factor-gene pair per million years for S. mikatae and S. bayanus, respectively) (Table S6) [18] are similar to the SH3 interaction network rates of interaction change for the same species (4.82×10−4 and 4.24×10−4 interaction changes per protein pair per million years for S. mikatae and S. bayanus respectively) (Table 1). Given enhancer regions diverge at a fast rate [32], [33], transcription factor interactions are expected to change rapidly.
Here we find that different peptide recognition domain networks evolve at different rates, all of which are faster by an order of magnitude than the rate of change of protein-protein interaction networks [21]. This is the case even when considering the estimated error in the rate of network rewiring within the SH3 interaction network between the evolutionary closest species of S.cerevisiae and S. paradoxus (estimated at 1.7×10−4 interaction changes per protein pair per million years). Given the rate at which the signaling and regulatory networks rewire, it is tempting to speculate that the ability to rapidly reorganize their structure is a mechanism for swift adaptation to selective constraints while minimizing disruption to a core network responsible for basic cellular functionality.
The rates above highlight the plastic nature of SH3 and kinase interaction changes which in addition to the lack of a significant correlation between a SH3 PWM's entropy and the rate of interaction rewiring (ρ = −0.067, p-value = 0.724, Figure S11) suggest detecting conserved interaction signals to be difficult. Interestingly, we readily observed global trends of network conservation. Using S. cerevisiae as the reference species, we found a significant correlation between the number of domains that bind a target protein and the degree to with interactions are conserved, for both the SH3 (ρ = 0.466, p-value<2.2×10−16) and kinase (ρ = 0.337, p-value<2.2×10−16) interaction networks (Figure 3A and 3B). These correlations explain 16% and 10% of the variance, respectively, where interaction conservation is the fraction of species retaining an interaction found in S. cerevisiae. This suggests targeted proteins may retain interactions by maintaining many interaction partners (ie. the target protein has a high in-degree).
To identify mechanisms giving rise to the above correlation, we use the position specific binding information provided by the PWMs to determine the exact region bound by a domain on a target protein (Materials and Methods). We found that high in-degree target proteins have peptide regions bound by multiple SH3 domains, forming binding site clusters primarily in proline rich regions (Figure cluster 4A). Binding site cluster formation may be attributed to two modes: 1) binding by the same SH3 specificity class and 2) binding by multiple SH3 specificity classes that share a common PXXP core, where X is any amino acid. As an example, Srv2p contains a multiclass cluster composed of class I, II, and III SH3 domain binders (Figure 4A). While we observe the formation of clusters, two proteins cannot simultaneously occupy the same binding site, thus multiple binding domains forming a cluster competitively bind for the target binding site.
Having identified cluster formations at SH3 target binding sites, we explored the relationship between the size of the binding site clusters and interaction conservation. A significant correlation was found between cluster size and interaction conservation (ρ = 0.192, p-value = 4.67×10−6) (Figure 4B), though not as significant as the correlation found at the global level of protein interactions between the number of interacting SH3 domains and interaction conservation. Interestingly, cluster sizes greater than 7 fail to exhibit the same degree of interaction conservation as proteins whose in-degree are of the same magnitude. Since proteins with many interacting SH3 domains may contain multiple clusters, this suggests the number of binding clusters may play a role in determining an interaction's conservation degree. Pursuing this observation, we find a significant correlation between the number of binding site clusters and interaction conservation (ρ = 0.461, p-value = 5.85×10−12) (Figure 4C). This is in agreement with previous studies suggesting that the amount of a protein participating in interactions is more conserved [34], [35]. The existence of disjoint clusters is analogous to the existence of several different interaction interfaces participating in a complex formation, which is most likely driving the correlation with interaction conservation. Similar observations are found within the kinase interaction network where correlations between interaction conservation and both binding site cluster size (ρ = 0.210, p-value = 1.40×10−14) and the number of clusters (ρ = 0.296, p-value = 6.48×10−10) are found (Figure S12).
Cluster formation at binding sites suggests that the sequence could be evolutionary constrained to preserve recognition by multiple binding domains. To investigate the relationship between sequence conservation and interaction conservation we measured the binding site divergence using the AL2CO algorithm (Materials and Methods) [36] and observe a significant correlation between sequence conservation and binding cluster size (ρ = 0.241, p-value = 2.80×10−5), indicating the existence of selective pressure to maintain sequence conservation due to multiple interacting partners. However, the relative contribution of different target binding site amino acids to a domain's binding specificity varies wildly. Consider the binding specificity of a class I SH3 domain with the consensus sequence RXXPXXP, where X is any amino acid. The first, middle, and last positions of the binding site are constrained to specific anchor amino acids, whereas the other positions are free to evolve to any other amino acid. Thus from the perspective of the class I domain, a binding site is conserved when the 3 anchor amino acids are present, while the other amino acids are free to evolve.
To measure the conservation of a target peptide sequence relative to a SH3 domain's binding specificity we developed a new metric. Using S. cerevisiae as the reference species, for an observed target binding site, we first calculate its PWM score and the PWM score for all orthologous proteins. The difference between a species' PWM score to that of the S. cerevisiae ortholog indicates the relative amount the target site has evolved. To provide a summary metric, PWM scores for each ortholog are weighted by their divergence distance from the S. cerevisiae ortholog (Materials and Methods). While many previous studies sought to identify conserved protein interactions [22], [23], here we show that this metric enables identification of both conserved and evolving interactions. Conserved interactions have values near zero as is the case for Srv2p which contains a highly conserved binding cluster of size 6. Evolutionary changing interactions bind evolving target binding sites identified by negative values, larger magnitudes indicate recently acquired target binding sites. One example of an evolved target binding site is found on Ubp7p in a cluster size of 5. The metric captures the evolving binding site, while only weak sequence conservation is observed. Not surprisingly, a significant correlation between the metric and interaction conservation (ρ = 0.425, p-value<2.2×10−16) is observed, highlighting the metric's ability to capture sequence conservation relevant to interaction conservation. In line with this, we find a significant correlation between the metric and binding site cluster size (ρ = 0.271, p-value = 2.41×10−6). This result demonstrates that binding specificity from multiple domains indeed places evolutionary constraint on the target binding site sequence.
In conclusion, we use the specificity profiles of a common signaling domain to generate a model network for the fungal clade spanning a wide range of evolutionary distances. In this manner we overcome the difficulty of comparing interaction networks of very divergent model organisms derived from a limited amount of experimental interaction data. We draw a number of conclusions that are biologically important: First, we find that the major driver of evolution of signaling pathways is interaction change and not gene duplication or loss. Second, we find network interaction changes are so rapid that they swiftly saturate, a phenomenon future studies will need to consider. Finally, we find signatures of network conservation and propose associated mechanisms. We expect our results can be generalized to many other signaling domains, such as SH2, WW, and PDZ domains, since they are affected by the same fundamental evolutionary processes.
Genomic and proteomic data for 23 fungal species from the Ascomycota phylum were obtained from a variety of sources [37]–[43]. The same genomic and proteomic datasets were selected as those used to generate the orthology group assignments by Wapinski et al. [25], but extend to 23 fungal species which is available on their website as version 1.1 (retrieved August 2009).
30 SH3 position weight matrices (PWMs) for 25 of 27 S. cerevisiae SH3 domains were obtained from Tonikian et al. [6]. Kinase PWMs were obtained from Mok et al. for 61 of 122 S. cerevisiae kinase proteins, for a total of 63 kinase PWM classes [11]. SH3 PWMs were constructed by aligning peptide sequences derived from phage display to create amino acid frequencies for each ligand position. Kinase PWMs are based on intensity signal ratios for each amino acid at every ligand position from peptide library screens.
Identifying putative target binding sites for each SH3 and kinase domain was determined using the Motif Analysis Pipeline (MOTIPS) [44], a method similar to ScanSite [45]. Binding target sites were identified for the 23 yeast species using the S. cerevisiae SH3 and kinase domain PWM specificity classes (see above). Here we use the term target binding site to refer to a peptide target prediction corresponding to a PWM class. Given a PWM, the binding target site score is defined bywhere c is the PWM class, is the optimal binding score for the PWM class, and is defined below.where Nc is the length of the target binding site for class c, p is the position of an amino acid within the target binding site, and aa is an amino acid's entropic adjusted “count” [46] at a given position in the PWM. Each predicted binding target site is further associated with parameters of disorder (lack of tertiary structure) and surface solvent accessibility respectively computed by disopred2 and sable [47], [48] to provide additional features to enhance target binding site predictions.
Parameter selection was performed using the area under the receiver operator curve (AUROC). Other binary classification metrics such as Matthews correlation coefficient (MCC) resulted in the same parameters being selected. To measure model performance, both true positive and true negative sets are required. For SH3 interactions, true positive interactions were retrieved from Tonikian et al. [6] and true negative interactions were created using randomly selected interacting partners for each SH3 domains. The true negative set was constrained to exclude protein pairs annotated with overlapping cellular compartments, additionally true positive interactions were removed from the negative set. The parameter selected for the SH3 interaction network were the top scoring 30 interactions for each class, with accessibility and disorder scores respectively greater than 3 and 0 to achieve AUROC and MCC values of 0.86 and 0.79 respectively. For the kinase interaction network we used data from Breitkreutz et al [49]. Unfortunately, this dataset identifies only interaction partners and does not include interaction directionality, which is crucial for our purposes. To determine interaction directionality, S. cerevisiae phosphosite data from literature [18], [50]–[54] was superimposed on the Breitkreutz et al kinase network. Kinase PWMs from Mok et al. were subsequently used to identify kinase domains that bind specific phosphosites [11], [49]. To ensure the interaction is significant and likely a true positive, phosphoproteins were scanned by every kinase PWM to create a best score background distribution. For each kinase PWM best score distribution we assume it follows a Student's t-distribution and use a threshold of p-value<1×10−16 to define the true positive set. The kinase network true negative set was constructed in the same manner as the SH3 negative. Parameters selected for the kinase interaction network were top 360 interactions for each PWM class, accessibility score greater than 4.0 and a disorder score greater than 0.9 to achieve AUROC and MCC values of 1 and 0.16 respectively. In all comparisons, a 1 to 5 ratio between true positives and true negatives were used. Various metrics describing the confusion matrix were found to be similar across a range of true positive and true negative ratios (Figure S1).
Comparisons between the 23 different species interaction networks were made using orthology mappings provided by the SYNERGY algorithm [55]. Two network comparison types were performed, global and local, respectively based on the absence or use of exact amino acid positional target binding information. Global comparisons involve determining the interaction conservation between two proteins. Here an interaction is conserved if any of the target proteins has an ortholog found to also be a binding partner for a given protein domain (one-to-one and one-to-many orthology relationships). Local comparisons are used in identifying conserved interactions when estimating the network rewiring rate. To simplify network comparisons for evolutionary rate calculations, we constrained orthology mappings to one-to-one and many-to-one relationships with respect to the reference species, thus ensuring at most a single protein per species exists in a multiple sequence alignment for each reference species' protein. Orthology mappings are established using the shortest distance between genes, where the distances are derived from the orthogroup's reconstructed gene tree, as orthogroups encompass both orthologs and paralogs. The reconstructed gene trees for each orthogroup were retrieved from the January 2009 data revision by Wapinski I. et al. containing the 23 fungal species used in this study [25].
Blastp [56] and the SH3 Pfam HMMs [57] were used to determine SH3 domain existence in the orthologs to S. cerevisiae proteins containing a SH3 domain. If both blastp and the SH3 Pfam HMMs failed to identify a SH3 domain, the protein was removed from the interaction network. S. cerevisiae domain regions were retrieved from Tonikain et al. [6]. Parameters used for blastp were ‘-evalue 0.1’ and for the Pfam HMMs, ‘-e_seq 0.1 -e_dom 0.1’. In a similar manner, kinase proteins from Mok et al. were selected [11] and the Pfam Pkinase HMM was run with ‘-e_seq 0.01 -e_dom 0.01’. The S. cerevisiae kinase domain boundaries were extracted from the Pfam output, and blastp was run with ‘-evalue 0.1’. If both blastp and the Pkinase Pfam HMM failed to identify a kinase domain in an orthologous S. cerevisiae protein, the protein was removed from the kinase interaction network.
For each gene target involved in a SH3 or kinase interaction, the protein sequences of the gene target and its associated orthologs were aligned using MAFFT v6.717b [58] with the ‘–auto’ parameter.
The R statistical program with the Analyses of Phylogenetics and Evolution (APE) package [59] was used to generate the phylogenetic trees derived from conserved interactions in the SH3 and kinase interaction networks for the 23 yeast species using a minimum evolution method. As input for such methods, a multiple sequence alignment (MSA) is required. A MSA based on conserved interactions can be created by considering an interaction between a protein pair as a single position within a MSA, whose value is 1 if the interaction is present and 0 otherwise.
The canonical phylogenetic tree was created by concatenating the protein MSA of 79 out of 153 genes families with synteny support [27], [60], where the 79 gene families correspond to orthogroups containing each of the 23 yeast species exactly once [25]. In other words, only gene families with no paralogs were selected. MAFFT was used to align the 79 gene family orthogroups. The alignments were concatenated and a phylogenetic tree was created using SEMPHY version 2.0b3 with the JTT matrix and parameters ‘–jtt –S –O’ [61]. The divergence distance from the last common ancestor between S. cerevisiae and S. pombe was set to 600 million years and the APE package was used to create the canonical phylogenetic tree.
To compare rates of interaction change with the computed SH3 and kinase interaction networks against prior literature rates, we used rates of interaction change provided by Beltrao and colleagues for the kinase and transcription factor networks. To provide a fair comparison, divergence times for K. lactis, C. albicans, and S. pombe with respect to S. cerevisiae were set to 300, 400, and 600 million years respectively, values use by Beltrao et al. [18]. Divergence times for the yeast sensu stricto group taken with respect to S. cerevisiae for S. paradoxus, S. mikatae, and S. bayanus were 10, 15, and 20 million years respectively [62].
Estimating rates of interaction change between interaction networks requires the divergence times between the networks to be known. To estimate divergence times, we created a canonical gene phylogenetic tree encompassing all 23 fungal species. The above section details the construction of the canonical phylogenetic tree.
The network rewiring mechanism of “interaction rewiring” and “protein change” require the identification of conserved interactions, as the absence of a conserved interaction is an interaction change. For interaction rewiring, an interaction is conserved if there exists a target binding site within a window centered on the reference species' target binding site plus 10 flanking amino acids. For protein change, interaction conservation is based solely on the existence of a protein ortholog.
Rates of interaction change with respect to a reference species were calculated in the same manner as Beltrao et. al. using the following equation:where intChanges is the number of gained and lost interactions, orthDomainProteins is the number of orthologous proteins between the two species in comparison containing a SH3 or kinase domain, orthProteins is the number of orthologous proteins, and divergenceTime is the divergence time in millions of years separating the two species [18]. Rates for both network rewiring mechanisms of interaction rewiring and protein change were calculated using the above equation. This equation can be viewed as the fraction of the interaction changes versus all possible interactions amongst proteins with orthologs and normalized by the divergence distance between the two compared species interaction networks.
Determining if rates of interaction rewiring significantly differ between the predicted network and random networks, 1000 sets of randomized interaction networks for each of the 23 species were created. The networks were randomized such that protein degree and the total number of nodes within the 23 interaction networks were maintained. For each set of randomized networks, the rate of interaction rewiring was calculated and compared against rates found in the original interaction networks.
Estimating the error in network rewiring was performed in the SH3 interaction network and between the two closest yeast species: S.cerevisiae and S. paradoxus. Using the predicted S.cerevisiae interaction network, the number of false positives and false negatives were calculated using the true positive and true negative datasets determined above. We assume the number of incorrect interactions is the same in the S. paradoxus SH3 interaction network, hence the maximum number of incorrect interactions is double the number of false interaction changes found in the S.cerevisiae network.
The set of target binding sites, linear peptide sequences, for each SH3 and kinase domain were provided by the MOTIPS pipeline. When multiple target binding sites overlap, a target binding site cluster is formed. Specifically, a binding site cluster is defined as a region on a protein target for which every peptide segment bound by a protein domain overlaps with every other bound peptide segment, and for every pair of overlapping bound segments, one segment overlaps another by more than 50% of its peptide length. A greedy approach is used to form the clusters.
Determining the binding site sequence conservation is measured by AL2CO [36] in conjunction with a weighted scoring scheme to account for gaps. AL2CO was run with the parameters ‘–f 1 –g 0.01’. Each amino acid position within the multiple sequence alignment was further weighted by 1 minus the ratio of gaps vs non-gaps at that amino acid position, thereby decreasing the conservation score of positions with many gaps.
Calculating the sequence conservation specific to a protein domain is computed in 3 steps: 1) use PWMs to calculate the highest scoring target binding site for all protein orthologs at the location of the reference species' target binding site in a MSA within a 10 amino acid flanking window, 2) obtain divergence distances for all protein orthologs with respect to S. cerevisiae, and 3) finally weight the difference in PWM scores between protein orthologs versus the S. cerevisiae protein PWM score with the reciprocal of the protein orthologs divergence distance from the S. cerevisiae protein. Computing the highest scoring target binding sites for all protein orthologs for a given protein domain is performed by making a MSA for each target protein, where each MSA consists of the target protein and its protein orthologs. Using S. cerevisiae as the reference species, a window is set to the boundaries of the target binding site plus 10 flanking amino acids on either side. For each protein ortholog, the highest scoring target binding site is attained from the sequence within the window. A score threshold defined as twice the worst PWM score in the S.cerevisiae network is applied to the PWM score difference between protein orthologs and the S. cerevisiae protein to capture sequences likely to have diverged towards the S. cerevisiae binding target site. Divergence distances for protein orthologs of a S. cerevisiae target protein are obtained from the reconstructed gene tree by Wapinski I. et al. of the target protein's orthogroup [25]. If protein paralogs exist within the orthogroup, the protein with the shortest distance to the S. cerevisiae target protein is retrieved.
|
10.1371/journal.pbio.1001028 | Cryptic Patterning of Avian Skin Confers a Developmental Facility for Loss of Neck Feathering | Vertebrate skin is characterized by its patterned array of appendages, whether feathers, hairs, or scales. In avian skin the distribution of feathers occurs on two distinct spatial levels. Grouping of feathers within discrete tracts, with bare skin lying between the tracts, is termed the macropattern, while the smaller scale periodic spacing between individual feathers is referred to as the micropattern. The degree of integration between the patterning mechanisms that operate on these two scales during development and the mechanisms underlying the remarkable evolvability of skin macropatterns are unknown. A striking example of macropattern variation is the convergent loss of neck feathering in multiple species, a trait associated with heat tolerance in both wild and domestic birds. In chicken, a mutation called Naked neck is characterized by a reduction of body feathering and completely bare neck. Here we perform genetic fine mapping of the causative region and identify a large insertion associated with the Naked neck trait. A strong candidate gene in the critical interval, BMP12/GDF7, displays markedly elevated expression in Naked neck embryonic skin due to a cis-regulatory effect of the causative mutation. BMP family members inhibit embryonic feather formation by acting in a reaction-diffusion mechanism, and we find that selective production of retinoic acid by neck skin potentiates BMP signaling, making neck skin more sensitive than body skin to suppression of feather development. This selective production of retinoic acid by neck skin constitutes a cryptic pattern as its effects on feathering are not revealed until gross BMP levels are altered. This developmental modularity of neck and body skin allows simple quantitative changes in BMP levels to produce a sparsely feathered or bare neck while maintaining robust feather patterning on the body.
| The distribution of hairs or feathers across the body is not homogeneous, and many animals have characteristic regions of their skin with either profuse or reduced coverage. These features, such as manes, crests, or bald patches, are seen in diverse species, suggesting that they can be selectively advantageous and also that the mechanisms by which the skin develops somehow enables such features to appear repeatedly in the course of evolution. In this study we explore the basis of loss of neck feathering, a feature associated with heat tolerance that has arisen independently several times during bird evolution. We find that in chickens a bare neck is caused by increased production of BMPs, factors previously implicated in defining the size of the gaps between neighboring feathers. Selective production of retinoic acid by embryonic neck skin enhances BMP signaling, thereby bringing this skin region close to the threshold of BMP action required to completely suppress feather development. This usually innocuous distinction between neck and body skin enables mutations that increase BMP action to render the neck completely bare while permitting normal feathering on the body. Thus an underlying map within the skin provides a one-step route to a bare neck.
| The vertebrate skin carries a highly ordered arrangement of pigments and morphological structures such as hairs and feathers. These patterns in the skin occur on two distinct spatial scales. Repetitive patterns of follicles or of pigment spots and stripes are laid out in a periodic manner, with each element in the micropattern positioned at a characteristic distance from its neighbors. On a larger anatomical scale, different parts of the body display periodic pattern variations in terms of the density and size of the repeated structures, and in regions of bare skin no periodic micropattern is present at all. These regional differences in micropattern across the skin constitute the macropattern.
Feathers are distributed in the avian skin on both of these spatial scales. The feather tracts, separated by bare skin, are macropattern elements, while the regular spacing between individual feathers defines the micropattern [1]–[3]. Both levels of organization arise in the embryo, beginning with the stereotypical positioning of the 14 feather tracts. In chicken this macropatterning is initiated at embryonic day 7 (E7) by dermal signals that induce stripes of cells that are competent to undergo feather development. These stripes can be detected using molecular markers to reveal the location of each incipient tract. The stripes broaden and propagate bilaterally across the skin, with micropatterning occurring just behind the propagating wavefront, resulting in the laying out of rows of feather primordia, called placodes [1],[4]. The placodes contain tightly packed cells that undergo rapid proliferation to produce a tubular outgrowth and subsequently undergo branching and differentiation to yield a mature feather fiber [5]–[7].
The sequential addition of new rows of feather placodes to tract margins terminates before the tracts meet, resulting in bare or downy spaces, called apterylae, between them. These bare patches persist through life and their area is associated with thermoregulatory capacity, particularly when present on the neck [8]–[10]. The extent and shape of feather tracts and apterylae are highly variable among bird species [11],[12], indicating an evolutionary malleability in the developmental processes that generate the skin's macropattern.
Distinct developmental mechanisms underlie the formation of macro- and microscale patterns. Classic embryological experiments have clearly demonstrated that of the two tissue layers that compose the skin, the dermis acts as the repository of positional information during macropatterning and that this information is conveyed to a positionally naïve epidermis [13]–[16]. In contrast to the rigid anatomical coordinates that define macropattern regions, experimental evidence and theoretical predictions [17]–[23] suggest that periodic micropatterning of the skin is achieved by the action of a reaction-diffusion mechanism whereby a field of cells is apportioned to placode or non-placode fates by the action of opposing Activatory and Inhibitory signals with specific regulatory connections and spatial ranges of action [24]–[26]. Such systems produce self-organizing patterns with relative pattern positions, in contrast to the absolute anatomical locations defined by the macropattern. The density of the pattern elements produced by these Activator-Inhibitor interactions depends on the relative potency of the Activator and the Inhibitor and their spatial ranges of action. In studies of micropatterning of chicken and mouse skin, experimental evidence points to members of the BMP family as being Inhibitory factors [17],[18],[21],[27], while WNT/β-catenin [20],[28]–[32] and FGF [33]–[36] pathways act as Activators.
Standard reaction-diffusion systems yield a single characteristic follicle density as the pattern output [25],[37]. However, the skin's micropattern is not uniform across the entire body, raising the question of how different densities of hair and feather follicles, or patches of entirely bare skin, are laid out to achieve the diverse skin patterns so characteristic of the external anatomy of the vertebrates. Here we address this question by analyzing the genetic and developmental basis of the Naked neck chicken, an example of macropattern variation in a single species.
In order to dissect the mechanisms that modulate feather patterning on neck skin, we had previously mapped the Naked neck (Na) mutation to a 13 cM interval in the distal region of chicken chromosome 3q, 5.7 cM from the closest microsatellite marker [38]. Here we use the original mapping family to fine-map the causative gene by searching for recombination breakpoints with SNP (Single Nucleotide Polymorphism) markers in order to narrow the interval. By analyzing the candidate genes in the reduced interval, we found that the Naked neck mutation causes elevated expression of the BMP12 gene in developing skin, which is associated with a large insertion approximately 260 kb from BMP12. We then establish that the feather patterning of the neck skin is influenced by the existence of a cryptic molecular macropattern that has modest phenotypic effects until revealed by alteration of BMP levels. This illustrates how the periodicity-generating interactions of a reaction-diffusion network are integrated with the positional information encoded at different anatomical sites to produce the skin's diverse macropattern.
Domestic Naked neck fowl lack feathers on the neck and have narrow feather tracts on the body (Figure 1A). As in wild species, the Naked neck trait in chicken is associated with enhanced thermotolerance and with increased agricultural production in hot climates [39],[40]. This trait is caused by a single incompletely dominant locus, which abolishes neck feathering and reduces body feathering by approximately 20% in heterozygotes and by 40% in homozygotes [41]. Patterning of feathers, rather than their morphogenesis or maintenance, is affected by the Naked neck mutation as mutant embryos lack feather placodes on the neck and display reduced tract expansion on the body (Figure 1B and 1C). Naked neck embryos and adults exhibit a discrete boundary between feathered and unfeathered regions, though in wild type birds there is no overt boundary demarcating neck from body skin (Figure 1B and 1C) and both regions are considered to carry a continuous spinal tract that runs from head to tail [11],[12].
To gain molecular insight into the basis of macropattern variations, we started by refining the location of the causative mutation. As we had already mapped the Na locus to a 13 cM interval of chicken chromosome 3 [38], we developed 11 new markers from this region to refine the location in the original mapping family. Recombination events in two individuals led to refinement of the candidate gene to a region of 770 kb, containing five annotated genes (Figure S1). We sequenced all predicted exons of these genes (HS1BP3, XM_419977; BMP12, XR_026709; CB043, NM_001031093; APOB, NM_001044633; and TDC6, XM_419980) from Na/Na genomic DNA and did not identify any mutations predicted to affect the coding sequences or splice junctions of any of these genes. This suggested that the Na mutation influences transcriptional regulation, resulting in altered expression of one or more genes in the region. We found that only one of the five candidate genes, BMP12 (also known as GDF7), is normally expressed in developing skin and embryonic feather placodes (Figure S2), and that this gene exhibits strongly increased expression in Naked neck mutant skin at the onset of feather patterning (Figure 1D). None of the other genes within the Na critical region has altered expression levels in Naked neck mutant skin (Figure S3). In situ hybridization revealed that the elevated expression is widespread throughout the skin of mutant embryos (Figure 1E and 1F). By sequencing across an indel polymorphism in the 3′UTR of BMP12, we found that in Na/+ heterozygous embryos the expression of the mutant allele is greater than that of the wild type in the skin, but not in internal organs (Figure 1G), demonstrating the action of a cis-regulatory mutation with a tissue-specific effect.
To further refine the location of the genetic modification causing the Naked neck trait, we genotyped multiple wild type and Naked neck individuals from geographically dispersed flocks for markers across the 770 kb critical region. This identified an approximately 200 kb region that was identical by descent in all available Na/Na individuals (Table S1). While tiling this region by overlapping PCRs we found that we could not amplify across one specific region (chromosome 3: nucleotides 105089664–105089844) in Naked neck individuals, suggesting the presence of a genomic rearrangement at this location. We used inverse PCR to define the sequences flanking this rearrangement, finding on both sides the insertion of chromosome 1 sequences that map 73 kb apart from one another in the reference genome (Figure 1H, Figure S4). These inserted sequences map to an intergenic region flanked by the WNT11 (NM_204784) and UVRAG (NM_001030839) genes on chromosome 1. We confirmed the presence of a large insertion at this location by PCR using chromosome 1 and chromosome 3 primers (Figure S5) and further confirmed that this insertion was both present in all Naked neck genomes available and absent from >500 wild type chromosomes from diverse breeds (Table S2). As this large insertion is unique to Naked neck genomes it appears that this mutation is responsible for the increased BMP12 expression in skin of Naked neck embryos through a long-range (>260 kb downstream) cis-regulatory effect.
As several BMP family members act during early feather development [17],[18],[42],[43] we determined whether the increased BMP12 expression in Naked neck embryos leads to an appreciably increased overall BMP signal response. SOSTDC1 (NM_204373) is a target of BMP signaling in developing mouse skin [21] and we confirmed that this gene is a BMP target in chicken skin also (Figure 2A). We then used SOSTDC1 as a marker to visualize the distribution of BMP responses in the developing neck skin. SOSTDC1 expression is detected at the periphery of nascent feather placodes in wild type skin (Figure 2B), consistent with these zones experiencing BMP-mediated lateral inhibition of feather identity during periodic patterning. At E7.5 the anterior region of the spinal tract, including the neck, displays one row of feather primordia on each side of the midline, and over the next 24 h the entire dorsal region of the neck becomes populated with feather placodes (Figure 2C). In contrast, Naked neck embryos display a broad swathe of SOSTDC1 expression across the neck (Figure 2D and 2E), consistent with the failure of feather placode formation in this region being a result of inhibition by elevated BMP12 levels. Confirming that excessive BMP signaling causes the Naked neck phenotype, we found that pharmacological suppression of BMP signal transduction rescues feather development on the neck of cultured Na/Na skin (Figure 2F).
Initially, we considered that the basis for the complete loss of neck feathers coupled with retention of body feathers in Naked neck mutants was likely a result of the disproportionate elevation of BMP12 expression in Na/Na neck skin compared to body skin (Figure 1D). However, we found that treating explant cultures of wild type skin with soluble BMP12 protein did not cause a homogeneous disruption of feather patterning, but instead reproduced the Naked neck phenotype (Figure 3A and Figure S6). Application of recombinant BMP4 yielded similar results, demonstrating that this skin regional effect on feather placode suppression is not a unique property of BMP12 but is general to these BMP ligands. Although the strongly elevated BMP12 expression on neck compared to body skin in chickens carrying the Na mutation is likely to influence the precise nature of the feather macropattern in this mutant, the greater sensitivity to BMP signals of the neck relative to the body in wild type embryos is sufficient to enable loss of neck feathering in response to quantitative changes in total BMP levels.
This finding demonstrates that regional macropatterning of avian skin, in particular the distinction between the neck and body, involves the same signaling molecules as employed for the conceptually distinct periodic micropatterning of individual feathers. To explore the relationship between periodic and anatomical patterning, we treated skin with different doses of BMP12 and assessed the effects on feather density and on body tract width. We found that the density of feather placodes on the neck is normally lower than that of the body and that neck skin placode density falls sharply when exposed to exogenous BMP12. In contrast, on the body the periodic pattern is relatively robust to increasing BMP12 levels (Figure 3A and 3B), though this treatment causes a dose-dependent reduction in the number of placode rows, and hence overall tract size (Figure 3C and 3D). To visualize regional differences in BMP-sensitivity we assessed SOSTDC1 expression in response to applied BMP12 and found elevated BMP responses on the neck, with a sharp gradient of sensitivity from neck to body (Figure 3E). Thus BMPs elicit greater transcriptional responses on the neck, in addition to being more effective inhibitors of feather development in this region.
The periodic micropatterning of feather placodes relies on the interaction of factors that activate or inhibit placode formation [17],[18],[42],[44], operating in a reaction-diffusion mechanism. BMPs have been proposed to represent inhibitory factors during feather placode patterning [17],[18], with the WNT/β-catenin and FGF pathways serving as key activators. Reaction-diffusion mechanisms rely on the action of an Activator, which stimulates production of more Activator in a positive feedback loop and which also promotes the synthesis of its own Inhibitor. Attainment of a high Activator concentration by a cell alters its fate, in this case to that of feather placode. When the Inhibitor possesses a greater range of action than the Activator and when the relative signaling potencies of Activator and Inhibitor are appropriately balanced, these interactions will produce a periodic pattern from near homogeneous initial conditions [24]–[26]. In such systems Inhibitor production is a result of both widespread, constitutive synthesis starting prior to patterning, denoted here by CI, as well as the Activator-induced Inhibitor upregulation that occurs during the patterning process (Figure 3F). We performed computational simulations to determine whether the operation of a reaction-diffusion system on a field with differing Inhibitor sensitivities could explain the different neck versus body patterning behaviors observed upon BMP12 treatment of embryonic skin. We applied differential Inhibitor sensitivity to our patterning field according to the profile of SOSTDC1 expression in BMP12 stimulated skin. Thus the simulations now explored periodic patterning on a field with an Inhibitor sensitive region, representing the neck, and a less sensitive region, representing the body, with a steep gradient of Inhibitor sensitivity between these regions. Varying CI in the patterning simulations, which mimics the application of recombinant BMP12 to cultured skin, altered the simulated placode patterns in the manner observed in experimental treatments. Thus, high CI values caused ablation of Activator foci in the sensitive “neck” domain, while pattern density on the simulated body was little affected (Figure 3G and 3H). As observed in Naked neck fowl and in BMP12 treated skin cultures, the simulations also yielded a sharp boundary between the neck and body, the location of which was stable with varying Inhibitor levels (Figure 3H). Further simulations testing a range of Inhibitor sensitivity gradient slopes revealed that the observed sharp, but not step-change, gradient of Inhibitor sensitivity best fits our experimental observations of chicken skin pattern behavior (Figure S7).
In contrast to the effects of augmenting Inhibitor production, our simulations predicted that graded suppression of BMP function would produce stronger pattern alterations on the body than on the neck. A numerical sensitivity analysis of the model demonstrated that moderate suppression of the Inhibitor causes a transition from a spotted pattern to a striped one (Tables S3 and S4; Figure S8) [45],[46] and our model predicted that such spot to stripe transitions would occur readily on the body, with stripe production on the more sensitive neck requiring further suppression of the Inhibitor's action (Figure 4A and 4B). We tested this prediction by inhibiting the Smad1/5/8 and p38MAPK transducers of the bifurcated BMP signaling pathway [47] in cultured skin. We found, as predicted by simulation, that neck and body patterns did indeed respond differently to BMP signal suppression, with stripes being more prevalent on body than neck skin at low doses, while upon further suppression of BMP responses the pattern on body and neck converged to yield ubiquitous β-catenin expression within the tracts (Figure 4C and 4D). Intuitively, this phenomenon can be understood as the suppression of Inhibitor/BMP activity leading to over-accumulation and saturation of the opposing Activator levels, and hence to expansion of Activator foci. The symmetric expansion of Activator foci becomes restricted with the narrowing of the inhibited zones separating them and adjacent foci are thus forced to expand laterally, creating elongated placodes. As more lateral expansion of foci occurs, the prevailing pattern becomes one of activated stripes, rather than spots. The higher sensitivity of neck skin prevents Activator over-accumulation at moderate levels of Inhibitor/BMP suppression, requiring further suppression of BMP signaling to achieve Activator saturation and stripe production. These findings show that a reaction-diffusion system operating on a field with different Inhibitor sensitivities explains both the modest difference in placode density between neck and body in unmanipulated embryonic skin, as well as the greater pattern divergences between neck and body caused by experimental titration of BMP signaling.
To elucidate the molecular basis of the different sensitivities of chicken neck and body to BMPs, we compared the gene expression profiles of these two skin regions by array hybridization at E7.0 (Table S5). This approach identified expression of components of the retinoic acid (RA) signaling pathway as a very prominent difference between neck and body skin. The RA synthesizing enzymes RALDH2 (NM_204995) and RALDH3 (NM_204669) [48],[49] and the RA target genes DHRS3 (XM_417636) and CYP26A1 (NM_001001129) [50],[51], displayed significantly elevated expression in neck compared to body skin. RA signaling is important for determining skin appendage identity and orientation during morphogenesis [52],[53] but has not previously been implicated in influencing the periodic patterning of skin appendages. Whole mount in situ hybridization confirmed that RALDH2 expression is more pronounced on neck than body, with strong expression also observed in developing neural tissue along the midline (Figure 5A and 5B). RALDH3 expression was predominantly on the neck, with some extension onto the body peripheral to the presumptive feather tract (Figure 5C and 5D), and prominent expression on the hindlimb at the margin of the femoral tract was also observed (Figure S9). Visualization of the sites of RA signal responses by detection of DHRS3 showed that while neck and body skin are both sensitive to the action of exogenous RA (Figure 5E), endogenous RA synthesis elicits responses specifically on the neck and in a diminishing gradient onto the anterior region of the body at the lateral margins of the feather tract (Figure 5F and 5G). The definition of the neck as a site of selective RA signaling is not unique to chicken as we observed very similar RA pathway gene expression profiles in duck, turkey, quail, and guinea fowl embryos during their feather patterning (Figure S10). Quantification of RA pathway gene expression revealed the transient nature of the neck/body disparity, with the neck displaying higher transcript levels only during feather patterning (E7 and E8) and little difference between neck and body prior to and following completion of this process (Figure 5H). To determine which skin layer produces RA and which layer responds to this signal, we quantified gene expression in isolated dermis and epidermis (Figure 5I). RALDH2 and RALDH3 expression were detected only in the dermis, while DHRS3 expression was predominantly epidermal. This shows that RA is produced in the dermis and acts as a signal to the overlying epidermis, a finding consistent with classical skin recombinations which demonstrated that macropattern information is encoded within the dermis [13].
Based on the finding that RA signaling occurs at higher levels in neck than body skin at the onset of feather patterning, we considered that this factor might be responsible for the heightened sensitivity of neck skin to BMP-mediated inhibition of feather development. We tested this first by determining the effect of RA on placode patterning, and then by asking whether the differences in patterning behavior observed between neck and body skin could be minimized by reducing the difference in RA signal intensity between these two regions. We found that RA acts as an inhibitor of feather placode formation, with increasing doses of RA leading to a reduction in placode density and ultimately to complete suppression of placode formation (Figure 6A and 6B). In contrast to BMP administration, RA signaling effectively suppresses placode formation on both neck and body. RA inhibition of placode formation requires active BMP signaling (Figure 6A), suggesting that the primary action of RA might be to sensitize the skin to BMP signals. To test this idea directly, we co-treated skin with modest doses of both RA and BMP12 and observed a synergistic effect of these two signals, with low doses of RA potentiating the action of BMPs to allow complete suppression of placode formation on the body (Figure 6C). Thus the ability of the body skin to resist BMP signals, which enables feather development in the presence of moderate levels of BMP, depends on the absence of RA signaling in this region. To confirm that RA signaling is responsible for sensitizing the neck to BMP action, we cotreated skin cultures with Citral, an inhibitor of the RALDH enzymes, together with BMP12 and found that this suppression of endogenous RA production allowed feather patterning on the neck (Figure 6C). These results show that RA sensitization of skin to BMP signals accounts for the different pattern behaviors of neck and body skin, allowing quantitative changes in gross BMP levels to selectively reduce or abolish neck feathering.
Hairs and feathers are laid out in different patterns on different parts of the body according to their roles in thermoregulation, defense, and display. We have explored the developmental basis for variation of neck feathering in birds, finding that the Naked neck trait in domestic fowl is caused by suppression of embryonic feather development through increased BMP12/GDF7 expression. This adds to the catalog of agricultural production traits associated with altered GDF (Growth and Differentiation Factor) function, which includes increased muscle growth for meat production (GDF8/myostatin) [54],[55] and fecundity (GDF9B/BMP15) [56] in livestock.
The increased BMP12 expression that we observe in Na skin is completely associated with the insertion of chromosome 1 sequence downstream of this gene. This inserted sequence lies between WNT11 and UVRAG and contains conserved non-coding elements, but no sequence predicted to be transcribed. While determination of the precise mechanism of action of the mutation requires further investigation, the well-characterized expression of WNT11 in developing chicken skin [57] suggests that BMP12 expression may be upregulated in Na mutants due to the acquisition of WNT11 enhancers lying within the insertion. This notion is supported by our finding that upregulation of BMP12 in Naked neck embryos is particularly strong on the neck compared to the body (Figure 1), and WNT11 expression also appears to be significantly stronger on the neck than the body (Table S5). Alternatively, the insertion could act to abolish the function of a repressive regulatory element on chromosome 3, a similar mechanism having been shown to be the cause of increased IGF2 expression contributing to enhanced muscle growth in pigs [58],[59]. The large distance between the insertion and the BMP12 coding sequence that it influences is consistent with an emerging picture of the strikingly long-range action of cis-regulatory elements that tend to be responsible for control expression of BMP family genes [60].
Based on their expression patterns and ability to suppress feather development, BMP family members have been proposed to be Inhibitors in a reaction-diffusion system that dictates the micropattern spacing between individual feather follicles [17],[18],[42], though no genetic evidence in favor of such an activity in vivo has previously been reported. Using graded stimulation and suppression of BMP signaling coupled with analysis of pattern transitions, we provide further evidence in support of the BMP family playing the key Inhibitory roles during periodic feather patterning. More importantly, we find that different regions of the skin display differing sensitivities to BMPs during feather patterning, revealing a molecular link between micro- and macroscale patterning. Appropriately balanced activities of Activatory and Inhibitory signals are key to the operation of reaction-diffusion systems; if either function is too potent, then no periodic pattern can be produced. Thus, above a given threshold of BMP signaling, the micropattern Activatory functions (probably mediated by WNTs and FGFs [28],[29],[33]–[36], though the precise regulatory connections between BMPs and these genes remain to be defined) are overwhelmed and cannot stabilize the positive feedback loop required to generate placodes. In this way a region of skin can be rendered refractory to periodic patterning by the amount of BMP signaling it experiences.
That neck skin has a greater sensitivity to BMP signals than body skin demonstrates that the apparently continuous spinal feather tract is in fact composed of two partly independent developmental modules. This modularity is enabled by the level of RA signaling, which is high on the neck and low on the body due to differential expression of RALDH genes. RA plays a key role in defining the placode pattern on neck skin by potentiating BMP signaling, thereby reducing feather density in a manner that depends on gross BMP levels. It is important to note that RA does not itself act as a component of the periodicity generator as we observe no evidence that RA synthesis within feather placodes acts to laterally inhibit placode identity in surrounding skin (Figure 5). Rather, RA acts as an external input that modulates the output of the periodic patterning mechanism (Figure 7). Previous theoretical studies have indicated that spatially distributed inputs can significantly modulate the form and variety of patterning [61]–[63] and the results here suggest that this type of external modulation is likely to be a recurring theme in reaction-diffusion patterning, as the imposition of such inputs allows a single set of Activator-Inhibitor interactions to produce distinct pattern outputs on different regions of a field, yielding a macropattern.
As no new feathers are inserted between existing ones as the skin expands to maturity, the adult feather pattern is a product of both the cell signaling processes focused on here together with the diluting effects of subsequent skin growth. Though placode density on the embryonic neck is significantly lower than that of the body, in adults the neck and body feather densities are the same (Figure S11). Thus the impact of RA in reducing neck placode density during patterning is compensated for by subsequent unequal growth of neck and body skin, with the body pattern being stretched to a greater degree than the neck, ultimately resulting in a homogeneous feather distribution across these two regions in the adult.
We also find that the lateral parts of the body skin are more sensitive to BMP-mediated suppression of feather development than the medial skin. However, we see no evidence that RA is involved in this phenomenon, as RA pathway genes are not expressed by lateral body skin and suppression of RA production using Citral does not impair BMP-driven reduction in body tract width (Figure 5 and Figure 6C). It is likely that this apparent medial-lateral BMP sensitivity gradient simply reflects the later formation of placodes on lateral than on medial skin. This results in lateral skin experiencing a greater duration of BMP stimulation prior to placode formation than medial skin. In addition, ventral skin is also likely to exhibit a higher BMP sensitivity than dorsal skin, since a marked reduction of feather cover is observed on the belly region of homozygous Na/Na individuals, while Na/+ heterozygotes have a more normally feathered ventrum. Thus different BMP sensitivities, perhaps based on a range of different molecular mechanisms, may play a widespread role in defining the macropattern across the entire body.
These findings have implications for the developmental mechanisms underlying the evolutionary diversity of skin patterns. During the course of avian evolution neck feathering has been lost independently in several lineages, notably in large species of the tropics, such as members of the Accipitridae (Old World vultures), Cathartidae (New World vultures), Ciconiidae (genus Leptoptilos, including the Marabou stork), and the large ratites (ostrich, emu, cassowary, and rhea). The fossil record does not support a bare neck as an ancestral feature of the feather pattern [64], raising the question of how this character could have evolved so frequently. The modularity of neck skin that we report illustrates that the positional information distinguishing neck from body is generally present in avian embryonic skin, requiring only changes to gross signal levels to produce a sparsely feathered or bare neck. In general, developmental modularity of this kind enhances evolvability by dissociating the effects of genetic change on distinct anatomical regions [65],[66]. The presence of cryptic skin patterns removes the need for evolutionary generation of positional information de novo, enabling the translation of spatially homogeneous changes in signal levels into spatially heterogeneous (i.e. patterned) morphological change. Such cryptic patterns may be widespread in vertebrate skin, imposing a substantial bias on the types of morphological changes likely to occur from mutation and so be exposed to natural, sexual, and human selection.
The population used for mapping of the Na mutation, with 70 informative progeny, has been described [38]. Genotyping was performed with 11 additional microsatellite markers designed from the available chicken sequence assembly (Table S6). White Leghorn embryos were used as wild type controls for in situ hybridizations, quantitative RT-PCR, and skin explant cultures. Naked neck samples were obtained from England, Scotland, France, and Mexico. Additional DNA samples were obtained from the Transylvanian Naked Neck population provided by the Godollo Institute in Hungary to the AvianDiv collection. Wild type samples of various breeds were obtained from The Wernlas Collection, Shropshire, United Kingdom, and from the INRA collection of experimental lines. DNA was isolated from embryos or blood using proteinase K digestion, phenol/chloroform or high salt extraction, and ethanol precipitation. Na/+ heterozygous embryos used to determine imbalanced allele expression were a cross between Na/Na and Silver Appenzeller. Oligonucleotides used for amplification across the indel polymorphism within the BMP12 3′UTR were: Forward: 5′-CGTGGTGTACAAACAGTACG-3′; Reverse: 5′-AAGCCCGGCCTTTTTATAGC-3′. PCR products were purified (QIAGEN) and directly sequenced.
Embryos or skin cultures were fixed overnight in 4% paraformaldehyde in PBS at 4°C. Samples were dehydrated into methanol, bleached using H2O2, rehydrated, treated with 5 µg/ml proteinase K, post-fixed, and hybridized. Samples were washed to remove unbound probe and hybridization detected using an alkaline phosphatase conjugated sheep anti-digoxigenin (Roche) and a BCIP/NBT color reaction.
Total RNA was isolated using TRI reagent and reverse transcribed using random primers and AMV reverse transcriptase (Roche) in a 20 µl total volume. Reactions were diluted 10-fold and 5 µl used as template for each qPCR. Double dye (5′FAM, 3′TAMRA) probes and primers were supplied by Eurogentec and Applied Biosystems. Probe sequences used were: GAPD: 5′-FAM-CATCGATCTGAACTACATGGTTTA-TAMRA-3′; BMP12: 5′-FAM-TCGGCACCGTCACCGGCTTC-TAMRA-3′; SOSTDC1: 5′-FAM-ACTTGAACGCGATTGTTAC-TAMRA-3′; DHRS3: 5′-FAM-AGGCGAGGAGCCAGGAAGATCATCC-TAMRA-3′; RALDH2/ALDH1A2: 5′-FAM-CAGATGCTGATTTGGATTATGCTGT-TAMRA-3′; and RALDH3/ALDH1A3: 5′-FAM-TGAGGAAGGAGACAAGCCTGATGTG-TAMRA-3′.
Twenty-microliter reactions were performed in triplicate, with at least four biological replicates used to determine each data point. Relative levels of GAPD, SOSTDC1, RALDH2, RALDH3, and DHRS3 were determined from a dilution standard curve, while a plasmid standard curve was used to determine BMP12 levels.
Dorsal skin from the entire crown-caudal length of E7.0 White Leghorn embryos was dissected and placed onto an MF-Millipore filter on a metal grid and submerged in DMEM containing 2% FBS in a centre well dish (Falcon) at 37°C, 5% CO2. Recombinant human BMP4 and mouse BMP12 (R&D Systems) were used. Dorsomorphin, Citral, and all-trans retinoic acid were supplied by Sigma-Aldrich and SB203580 by Merck. Feather placode densities, shapes, and areas were measured on β-catenin hybridized skin samples using ImagePro PLUS (Mediacybernetics). Circular placodes were defined as β-catenin positive foci with a circularity ratio (perimeter2/4πarea) of ≤1.2. Placode densities were determined only within tracts and did not include non-feathered areas. Mathematical modeling methods are described in Text S1.
All predicted exons in the ENSEMBL database lying between chromosome 3: 104754409–105526289 were amplified by PCR from genomic DNA of individuals in the French Na/Na experimental population and directly sequenced using the primers used for PCR amplification (oligonucleotide sequences available on request). Functional variants were defined as non-synonymous, frameshift, or nonsense SNPs within a predicted open reading frame, or as nucleotide substitutions within 10 bases of an intron/exon junction, based on comparison to the reference genome. Putative functional variants that were not in the dbSNP database were then sequenced from wild type individuals. No functional variants that were unique to Na/Na were identified.
The microarray study used the Agilent Chicken expression arrays (design 015068: Agilent Technologies, Berks, UK) in a two dye reference experiment. Neck skin total RNA was labeled with Cy5 and body skin total RNA was labeled with Cy3 using the Ambion MessageAMP kit with aminoallyl labeled UTP (Applied Biosystems, UK) and the Cy3 and Cy 5 Dyes (GE Healthcare Life Sciences, Bucks, UK) according to manufacturers' protocols. Four independent E7.0 White leghorn body/neck RNA pairs were used for independent, unpooled hybridizations, which were carried out using the Agilent hybridization chambers and equipment. The slides were washed according to Agilent Technologies protocols and scanned in an Axon 4200AL scanner (Molecular Devices, UK) at 10 micron resolution. The scanned images were processed using the Feature Extraction software from Agilent Technologies.
|
10.1371/journal.pcbi.1003083 | A Mechanical Design Principle for Tissue Structure and Function in the Airway Tree | With every breath, the dynamically changing mechanical pressures must work in unison with the cells and soft tissue structures of the lung to permit air to efficiently traverse the airway tree and undergo gas exchange in the alveoli. The influence of mechanics on cell and tissue function is becoming apparent, raising the question: how does the airway tree co-exist within its mechanical environment to maintain normal cell function throughout its branching structure of diminishing dimensions? We introduce a new mechanical design principle for the conducting airway tree in which mechanotransduction at the level of cells is driven to orchestrate airway wall structural changes that can best maintain a preferred mechanical microenvironment. To support this principle, we report in vitro radius-transmural pressure relations for a range of airway radii obtained from healthy bovine lungs and model the data using a strain energy function together with a thick-walled cylinder description. From this framework, we estimate circumferential stresses and incremental Young's moduli throughout the airway tree. Our results indicate that the conducting airways consistently operate within a preferred mechanical homeostatic state, termed mechanical homeostasis, that is characterized by a narrow range of circumferential stresses and Young's moduli. This mechanical homeostatic state is maintained for all airways throughout the tree via airway wall dimensional and mechanical relationships. As a consequence, cells within the airway walls throughout the airway tree experience similar oscillatory strains during breathing that are much smaller than previously thought. Finally, we discuss the potential implications of how the maintenance of mechanical homeostasis, while facilitating healthy tissue-level alterations necessary for maturation, may lead to airway wall structural changes capable of chronic asthma.
| With every breath, mechanical pressures change in the lung and permit air to efficiently traverse the airway tree and undergo gas exchange. These pressure variations also influence cell and tissue function, raising the question: how does the airway tree co-exist within its mechanical environment to maintain normal cell function throughout its branching structure of diminishing dimensions? We introduce a new mechanical design principle for the conducting airway tree in which mechanotransduction, the process that converts mechanical forces on cells to biochemical signals, is driven to orchestrate tissue-level structural changes that can best restore a preferred mechanical microenvironment; a concept termed mechanical homeostasis. We report in vitro mechanical properties for a range of airway sizes and present a mathematical model that describes the data. Our results indicate that airways indeed consistently operate within a preferred mechanical homeostatic state. We further describe how this mechanical homeostasis while facilitating healthy tissue-level alterations necessary for maturation can inadvertently lead to airway wall structural changes capable of chronic asthma.
| The act of breathing creates a mechanical environment that pervades all structures in the lungs down to the molecular level [1]. As a pressure difference across the lungs draws air through the airway tree for gas exchange, all airways dilate and transmit mechanical stresses and strains to their cellular constituents – including smooth muscle, epithelial, and fibroblast cells. All of these airway cell types reside within a complex bifurcating airway tree, and through mechanotransduction, they actively sense and respond to their mechanical environment [1]–[3]. Indeed, growth and remodeling of the extracellular matrix is consistently observed in response to chronically altered mechanical stress and occurs through a concerted response of multiple airway cell types [4]–[6].
This study asks the following question: What are the fundamental principles guiding the distribution of airway wall properties – in particular, airway wall thicknesses and airway wall material properties - throughout the conducting airway tree residing in a dynamic mechanical environment? Previous studies have examined airway tree design principles but have ignored both the dynamic mechanical forces to which the airway tree is perpetually subjected and the tissue-level biomechanical properties of the airway wall. Instead, based on the concept of a self-similar rigid tree [7] with optimal space-filling properties [8], it was presumed that the fractal design of airway luminal radii through the airway tree is optimal to transport fresh air to the periphery of the lung for efficient gas exchange. A tree design based purely on optimizing gas transport through rigid pipes ignores the fact that breathing dynamics can produce mechanically-driven alterations in the cells and tissue of the airway walls. Moreover, it is conceivable that these alterations would eventually modify the material properties and hence the caliber of the pipes themselves perhaps in a fashion destroying the underlying physical optimization related to gas transport [9].
Here, we introduce a new mechanically-based design principle for the airway walls of the conducting airways, in which mechanotransduction at the level of cells occurs in response to an altered mechanical microenvironment and is driven to orchestrate tissue-level structural changes of the airway wall to restore and maintain a preferred mechanical microenvironment. This theory is termed mechanical homeostasis and provides a plausible physiological role for mechanotransduction [10] in mechanically-driven tissue and organ systems. The concept has emerged as a prevalent theory in the vascular system [11]. However, mechanical homeostasis has not yet been conceptualized nor tested in the airway system.
Using experimental and modeling approaches, we show that the distribution of tissue-level biomechanical properties of the airway walls within the normal conducting airway tree is consistent with the existence of mechanical homeostasis. We also provide the likely desired homeostatic conditions for a healthy airway tree undergoing tidal breathing and occasional deep inspirations. Lastly, we address the implications of a mechanically-driven design principle with regard to airway disease. We conjecture that changes in the mechanical environment alone would facilitate healthy tissue-level alterations necessary for maturation on the one hand, but could also lead to airway wall structural changes capable of chronic asthma in a “misguided” attempt to sustain such mechanical homeostasis.
We first examine if and how the principles of mechanical homeostasis occur in structurally intact bovine airways in vitro. We measure the quasi-static relationships between airway luminal radius (Rin) and transmural pressure (PTM) (Fig. 1a, red circles). By adopting a computational model of vascular mechanics using a strain-energy formulation for thick-walled cylindrical tubes [12], [13], we estimate the three-dimensional stresses and strains within intact airway wall tissue (Methods). We implement this analysis to determine the optimal model fits to our data for positive PTM (Fig. 1a, solid lines), and calculate the relationships between the circumferential stress at the inner wall (σθ) and PTM (Fig. 1b, solid lines) and the corresponding incremental circumferential elastic modulus (Yinc) and PTM (Fig. 1c, solid lines). This analysis identifies two salient mechanical features present within our airways. First, within the typical operating PTM in vivo (0.5 to 1 kPa), every airway experiences a relatively narrow fixed range of σθ between about 8 and 15 kPa, which is less than 6% of the maximum stress. We approximate this as a fixed circumferential stress of 12 kPa at the mean operating PTM of 0.75 kPa in Fig. 1b. Second, for all 11 airways measured, there is a nearly identical relationship between Yinc and PTM with values between 180 and 310 kPa at PTM = 0.75 kPa, as defined by the dashed curves in Fig. 1c. Importantly, we find no correlations between Yinc at 0.75 kPa PTM and absolute airway size (Fig. 1c inset) suggesting that the airway wall stiffness is approximately the same for many generations of the airway tree when evaluated around the physiological operating PTM of 0.75 kPa.
We next utilize a computational modeling approach to examine how the uniform σθ and Yinc would be maintained throughout an airway tree structure exposed to the transmural pressures of breathing. The substantial decreases in luminal radius from the trachea (generation 0) to the periphery (generation 26) presumably allow for optimal gas transport via fractal branching but would also result in vastly different circumferential stresses for the resident cells within the airways, as evident in simplistic terms by the geometric relationship of LaPlace's law (σθ = PTM * Rin/H where H = wall thickness). To maintain a constant σθ and Yinc at the operating PTM throughout the airway tree, the airway wall areas and material properties would need to play a compensatory role within the thick-walled cylindrical airway. We re-analyze high resolution computed tomography (HRCT) data measured in humans by several investigators [14], [15] (Methods) to discover a strong linear relation between airway wall area and airway luminal radius throughout the airway tree (Fig. 2a). Remarkably, this relationship is maintained throughout all stages of lung growth from birth to adulthood. That is, an airway at the periphery of an adult lung, which has the same luminal radius as a central airway in a child, also has the same wall area. Taken together with our results of radius-independent micromechanical environment in Fig. 1, these findings are consistent with the existence of mechanical homeostasis within the airway tree.
Using the geometric relationship in Fig. 2a, we next predict the airway wall material parameters required to maintain the data-derived conditions of mechanical homeostasis (from Fig. 1) throughout the airway tree. By necessity, the model-determined material parameters describing the nonlinear elasticity of the airway wall (Eq. 1) increase from the periphery to the trachea (Fig. 2b). From the generation-dependent geometric (Fig. 2a) and material (Fig. 2b) relationships, we then compute σθ and the circumferential stretch ratio (λθ) (Fig. 2c, solid lines) for the entire airway tree and consequently, we obtain Yinc (Fig. 2c, slopes of solid lines). The peripheral airways have smaller Yinc than the central airways at any given λθ, which is consistent with observed decreases in collagen and cartilage content down the airway tree [16]. As a direct consequence, the progressively increasing stretch ratios along the nonlinear σθ- λθ curves allow the peripheral airways to experience the same σθ and Yinc as the central airways when exposed to typical operating PTM (Fig. 2c, cyan circles).
To validate our model prediction, we utilize the predicted σθ- λθ curves to calculate the relationships between λθ and PTM for the airway tree (Fig. 3a). These relationships are directly measurable in human lungs in situ for large airways [17] (greater than 2 mm luminal radius) and in vitro for peripheral human airways [18], [19]. From the trachea to the periphery, our analysis predicts that the airways are progressively more compliant. The specific airway compliance (Δλθ/ΔPTM) from 0 to 3 kPa PTM modestly increases from the trachea (4.4 Pa−1) to the periphery (5.1 Pa−1) (Fig. 3b, solid black line). Our predictions of specific airway compliance are consistent with the human data in the literature [17]–[19] (Fig. 3b, black circles), our bovine data (Fig. 3b, blue crosses), and data in dogs [20] and rabbits [21] (Fig. 3b, green squares and cyan stars). Importantly, our predictions are also consistently below the estimate of parenchymal hole expansion, which is generally accepted as the limit of airway expansion (Fig. 3b, dashed line). We also performed a sensitivity analysis that shows that these predictions remain essentially the same when the calculations are repeated for an optimal PTM of 0.5 kPa (solid magenta) instead of 0.75 kPa (solid black). Additionally, the root mean square error (RMSE) between model and data are similar for both PTM of 0.5 kPa and 0.75 kPa independent of airway radius (Fig. 3c). Thus, the agreement between our model predictions and the data in Fig. 3b is consistent with the notion that the known generation-dependence in specific airway compliance exists as a means to maintain a constant intrinsic mechanical microenvironment in response to PTM throughout the airway tree.
With only modest increases in specific airway compliance down the tree, our results also suggest that cells in the airway wall are well-equipped to respond uniformly to strain-dependent phenomena, such as cellular fluidization from deep inspirations [3]. Since it has not been possible to directly measure small airway mechanical properties in vivo, computational models of airway wall mechanics are vital to estimate the mechanical strains of breathing that are used in mechanobiological cell culture and tissue strip experiments. One prevailing model developed by Lambert et al. [22] was based on the extrapolation of limited empirical fits of radius-PTM data [17]. This model has been used in isolated ASM strips to apply in vivo like loads that would mimic a small airway's structure (1.1 mm luminal radius). In this experimental model, deep inspirations to TLC result in large sustained reductions in ASM constriction [23]. These results are in stark contrast to PTM oscillations applied directly to constricted bovine [24] and human [25] airways, which have little to no impact on airway caliber. Interestingly, we find that while the Lambert model has similar predictions to our model for large airways, the Lambert model also predicts airways having unrealistically large compliance as radius decreases (Fig. 3b, solid gray line), which eventually become much larger than the available data and the parenchymal hole expansion limit. In contrast, our model suggests that the mechanical strains experienced by ASM cells during breathing are much smaller than previously thought as our simulations demonstrate (Fig. 4). Furthermore, these strains depend only mildly on airway size increasing from 3.3% and 6.5% at the trachea to 4.9% and 9% in small airways during tidal breathing and deep inspiration, respectively. As a consequence, their sustained functional impact on airway responsiveness would be greatly attenuated when tested under physiologically appropriate mechanical conditions [24], [25].
Our data and modeling analyses suggest that the organ-level structure of the airway system plays a crucial role in sustaining not just organ level function, namely gas transport, but also the function of its individual components down to the cellular level in every airway wall throughout the airway tree during breathing. Specifically, the airway tree's structural design maintains circumferential stresses and circumferential elastic moduli in a relatively narrow range. The circumferential stress and the incremental modulus determine the effective stiffness of the extracellular matrix which is known to significantly influence cellular behavior [2]. Thus, the regulation of the mechanical microenvironment throughout the airway tree (Fig. 1), during growth (Fig. 2a), and in response to tidal breathing (Fig. 4) to maintain a homeostatic state ensures normal cellular and tissue-level function within a dynamic PTM environment in a way that maintains compatibility with the physical structure of the airway tree required for optimal gas transport. It is interesting to note that while homeostasis is formulated in terms of stress and modulus, it also results in a nearly constant circumferential strain independent of airway radius (Fig. 4) implying that similar cells in the walls but with different sensitivity to strain did not have to evolve for efficient airway function. Thus, the consistency of tissue-level structure with a mechanical design principle further implies an active and essential role of airway cells as controllers of mechanical homeostasis for the airway wall system. While airway luminal radius and length have evolved to efficiently deliver gas through a tree structure [8], the airway wall tissue structure has evolved to provide the proper mechanical milieu for its constituent cells.
Before examining the implications of a mechanical design principle, we first discuss the limitations of our intact airway experiments and computational modeling approaches. We used bovine intact airways in vitro to obtain pressure-radius relationships. Compared to tissue strips and cell culture, this preparation maintains the 3-dimensional architecture of the airway and applies a physiological transmural pressure as a means of stretching the cells and the extracellular structures in the airway wall. However, neural airway tone that is known to constantly modulate airway diameter in vivo [26] is removed in this preparation. When using these data in our computational approach, we neglected the pressure-diameter relation for negative PTM, which may be important during flow limitation; we found that to describe the negative PTM data, we would need a different form of the strain energy function. While our analysis of the data advances airway wall modeling by invoking a thick-walled cylinder approach, the model uses several simplifying assumptions. We assume that the airway walls are elastic, which is a limitation of the strain energy density formalism itself. We also assume that the airway walls are homogeneous and maintain a cylindrical shape at all PTM. During breathing, the lung is exposed to cyclic stretch and the irregular nature of the stretch pattern has important consequences on cellular function [27]. If cellular growth or remodeling of the airway are of interest, these irregular stretch patterns may influence these processes. The thick-walled hollow cylinder is assumed to be homogeneous and isotropic, which are the usual assumptions in airway wall modeling. At the length scale of single cells, however, this is certainly not true. We also assume that the typical operating pressure is 0.75 kPa for all conducting airways and for all stages of life (birth to adulthood). It is probable that the operating pressure may change throughout the course of development, and it may also differ slightly throughout the airway tree due to gravitational differences in pleural pressure, resistive losses in airway pressure down the airway tree, surface tension, and local differences in parenchymal tethering forces. A sensitivity analysis, however, revealed that almost identical results are obtained for operating pressures of 0.75 kPa and 0.5 kPa.
Another limiting assumption in both our data analysis and computational approach is plane strain, in which the properties and deformation of the airway wall are assumed to not change in the axial direction. In our in-vitro experiments, the two ends of the airways are fixed at a pre-stretch ratio that is consistent with in-vivo lengthening that occurs during tidal breathing [28]. In this setup, the plane strain condition is indeed valid for a section of the airway in the middle away from the boundaries which we verified using the ultrasound; data only from this region are used in the analysis. However, in the in-vitro experiments and in our computational approach, airways cannot lengthen dynamically with radial expansion, as presumably occur in-vivo. Relaxing the plane strain assumption to allow for airway lengthening would require complex finite element modeling and knowing the in vivo boundary conditions for the airway which are beyond the scope of this study.
We next discuss the implications of mechanical homeostasis on the behavior of airways over various time scales. On short time scales, ASM cells can actively control local stresses in the wall by contraction and relaxation. This has indeed been observed in dogs where airway lumen varied substantially from day-to-day over a period of a year [29]. On longer time scales, many other cell types participate in controlling stresses via remodeling of the airway wall. In fact, the consistency of geometric dimensional relationships from birth to maturity in the airway tree (see Fig. 2a) suggests that proper tissue growth is not a pure biochemical process but it also requires significant mechanically-driven feedback. As growth factors cause an airway's luminal radius to increase in size [30], the circumferential stress and elastic modulus would also initially increase at a given operating PTM. However, cellular compensatory mechanisms would detect this deviation from mechanical homeostasis and trigger the airway cells to build more wall tissue to restore the homeostatic state, characterized by an optimal circumferential stress and incremental Young's modulus. Therefore, a generation 0 airway at birth would grow to a much larger luminal radius in maturity while maintaining a similar preferred mechanical environment. As a consequence of this airway wall growth process, our analysis predicts a generation-dependent decrease in airway wall compliance as the airways mature (Fig. 5a, black circles), which is in agreement with data in the literature [21] (Fig. 5a, squares and triangles). This proposed process is consistent with an emerging hypothesis that the unique geometric and material properties of mature organs derive from mechanical stimuli and feedback throughout development [31] and may add another mechanical stimulus for growth in the respiratory system [32].
In an analogous fashion, our results also have implications for airway diseases. While the airway wall structure is designed to maintain normal cell function under physiological operating conditions throughout life, it remains ill-prepared for persistent ASM activation from environmental sources. ASM contraction immediately changes airway wall circumferential stresses by introducing an active mechanical stress that reduces airway caliber and increases wall thickness. Thus, airway cells would be immediately driven from their physiologic mechanical homeostatic state. Over time, repetitive ASM activation would stimulate airway wall remodeling processes consistent with growth in an attempt to restore mechanical homeostasis. Chronic constriction would result in an airway wall that is thicker and stiffer (Fig. 5b), as is consistent with data on asthmatic airways in the literature [33]. However, as ASM activation is sporadic in nature and occurs on a much faster timescale than structural remodeling, the presence of airway remodeling indicates a tissue-level system struggling to restore a steady-state mechanical homeostasis that may never be fully realized. Applied throughout the airway tree, this deviation from mechanical homeostasis would have disastrous impacts to function at the level of the cell, tissue, organ, and organism [6], [9] and would not be readily reversible with bronchodilators [33]. Mechanical homeostasis may thus emerge as a governing principle that divides health and disease within the respiratory system, and may also unify respiratory diseases with a host of others whose progression is intimately coupled to mechanical signaling [34].
Experiments were carried out using bovine lungs obtained from a local slaughterhouse immediately after death (Research 87, Bolyston, MA). Protocol approval was not required.
Our system for intact airway experiments has been described previously [35]. Briefly, bovine lungs were obtained from a local slaughterhouse. A bronchus of the right lung (generations 10–17) was dissected and the side branches were ligated. The airway was cannulated at each end and mounted horizontally in a tissue bath containing gassed (95% O2-5% CO2) and heated (37°C) Krebs solution. The airway was stretched longitudinally (110–120% of its resting length) and held fixed at its extended length for the entire experiment. Tissue viability was then confirmed with both electric field stimulation and acetylcholine (ACh; 10 5 M).
A computer-driven pressure-controlled syringe pump delivered PTM changes to the intact airway. The proximal cannula was mounted in series to a hydrostatic pressure column filled with Krebs solution, which also filled the airway lumen. The difference in fluid height between the horizontally mounted airway and the pressure column determined the intraluminal pressure (thus PTM) experienced by the airway.
A portable ultrasound system (Terason 2000), consisting of a high-frequency linear array transducer (10L5) and an external beamformer module, was used to visualize the intact airway. The hardware was connected directly to a personal computer running software that both controlled the imaging settings (focal depth, focal length, and gain) and acquired and stored the images in real time. The ultrasound transducer was mounted above the intact airway and partially submerged in the tissue bath. Using a three-directional micromanipulator, the transducer was positioned over the airway's longitudinal axis at its diameter. The airway was imaged at 30 fps with fixed ultrasound imaging settings (focal depth: 30 mm, focal length: 13 mm, gain: 0.2).
We controlled Pin delivered to the structurally intact bovine airways of fixed length while Rin and Rout were directly measured with ultrasound imaging. Quasi-static, passive R-PTM curves were measured via slow ramps in Pin (−1.5 to 3 kPa, 0.1 kPa/second) with Pout set to 0 kPa. The positive expiratory limb (3 to 0 kPa) was used for our analysis to probe the full physiological range of breathing (0.5 kPa PTM at functional residual capacity to 3 kPa at total lung capacity). Eleven bovine airways were obtained from the right lower lobes of eight different animals.
|
10.1371/journal.pcbi.1001013 | Self-Organized Criticality in Developing Neuronal Networks | Recently evidence has accumulated that many neural networks exhibit self-organized criticality. In this state, activity is similar across temporal scales and this is beneficial with respect to information flow. If subcritical, activity can die out, if supercritical epileptiform patterns may occur. Little is known about how developing networks will reach and stabilize criticality. Here we monitor the development between 13 and 95 days in vitro (DIV) of cortical cell cultures (n = 20) and find four different phases, related to their morphological maturation: An initial low-activity state (≈19 DIV) is followed by a supercritical (≈20 DIV) and then a subcritical one (≈36 DIV) until the network finally reaches stable criticality (≈58 DIV). Using network modeling and mathematical analysis we describe the dynamics of the emergent connectivity in such developing systems. Based on physiological observations, the synaptic development in the model is determined by the drive of the neurons to adjust their connectivity for reaching on average firing rate homeostasis. We predict a specific time course for the maturation of inhibition, with strong onset and delayed pruning, and that total synaptic connectivity should be strongly linked to the relative levels of excitation and inhibition. These results demonstrate that the interplay between activity and connectivity guides developing networks into criticality suggesting that this may be a generic and stable state of many networks in vivo and in vitro.
| Learning depends crucially on the synaptic distribution in a neural network. Therefore, investigating the development from which a certain distribution emerges is crucial for our understanding of network function. Morphological development is controlled by many different parameters, most importantly: neuronal activity, synapse formation, and the balance between excitation and inhibition, but it is largely unknown how these parameters interact on different time scales and how they influence the developing network structure. In our work, we consider the well-known concept of self-organized criticality. We have measured how real cell cultures change their activity patterns during the first 60 days of development traversing through different stages of criticality. With a dynamic model we can reproduce the observed developmental states and predict specific time-courses for the network parameters. For example, the model predicts a delayed, overshooting onset of inhibition with a longer time to reach maturation as compared to excitation. Furthermore, we suggest that the balance of dendrites and axons in the mature state is quite sensitive to the initial conditions of development. These and several more predictions are accessible by future experimental work and can help us to better understand neuronal networks and their parameters during development and also in the mature state.
| During the last years increasing evidence has accumulated that networks in the brain can exhibit “self-organized criticality” [1]–[3]. Self-organized criticality is one of the key concepts to describe the emergence of complexity in nature and has been found in many systems – ranging from the development of earthquakes [4] to nuclear chain reactions [5]. All these systems exhibit spatial and temporal distributions of cascades of events called avalanches which can be well described by power laws [6]–[8]. This indicates that the system is in a critical state [6], [9] and that similar dynamic behavior exists across many different scales. Several neural network models have predicted that neural activity might also been organized this way [10]–[14] and recently this had been confirmed experimentally [1], [15]–[17]. A recent study by Levina and colleagues [18] addresses the question how self-organized criticality can emerge in such networks in a robust way by using dynamical synapses, which alter their synaptic connection strength on a fast time scale. This contribution, which is able to analytically predict the network behavior, is a theoretical milestone in our understanding of criticality in neural systems. In general, however, theoretical and experimental investigations have so far usually focused on mature networks [1], [16] sometimes including adaptive processes [18]–[20]. Little is known how developing networks can reach a final state of self-organized criticality [10], [17], [21]. In the current paper, we are therefore experimentally investigating the different stages of developing cortical cell cultures [22] to assess under which conditions these networks develop into a critical state. Specifically we are asking the following questions: 1) do the investigated cell cultures undergo a significant transition in their activity states and how is this related to self-organized criticality and 2) can specific predictions be made with respect to network activity and connectivity which would explain the observed behavior. To address the second aspect we are designing a model to simulate network development, which is based on activity-dependent axonal and dendritic growth leading to homeostasis in neuronal activity [23]–[28].
In order to assess how self-organized criticality develops in cell cultures, we have monitored a total of 20 cultures and recorded their activity patterns between 13 and 95 days in vitro (DIV). In general, cultures start with about 500,000 dissociated cortical neurons, which develop over time into an interconnected network. To assess the different network states the activity at 59 electrodes was measured and analyzed at different DIV (see Methods). Figure 1 A shows 15 minutes of recorded activity for one typical culture at 42 DIV. At this temporal resolution individual bursts are visible as vertical dot-lines indicating activity at almost all electrodes, separated by rather long pauses which allow for robust separation of these bursts required for avalanche analysis. At a fine temporal resolution (Figure 1 B) one sees that the burst activity expresses certain patterns. Note, pauses have been graphically shortened in panels (B) and (C). Panel (C) shows the activity pattern that arises in our model, which at a first glance looks similar to that in the culture. Details about the model and an analysis which support similarity of model and real data, will be provided later. First we would like to describe the developmental stages in the cultures with respect to their avalanche distributions. In this work, avalanches are defined by the number of spikes between two windows without activity (see Methods).
At early stages during development, usually before 13 DIV, connectivity is small and activity in the network very low. So, it is very difficult to obtain long enough recordings for plotting avalanche distributions. However, known from the literature [29], in this stage activity is best described by a Poisson like behavior. At about 13 DIV (see Figure 2), we receive the first distributions which develop towards criticality (Figure 3 A). Therefore, we call this state the initial state. The ideal power-law fit for each curve is shown by the dashed lines. If a distribution matches the power law line it can be called “critical” [6], [7]. A dominance of long avalanches is indicative of a supercritical state whereas a lack thereof is referred to as subcritical. This is measured by , which gives the quality of fit between ideal power law and actual distribution. For a system in a supercritical state is larger and for a subcritical state smaller than zero (see Methods). Values of are also shown in the different panels of Figure 3. For the cultures, we receive at (on average) 19 DIV values of in the interval from to . While this shows that the system develops towards criticality, we also observed that this behavior is very unstable. Quickly, within just (on average) one/two days, the distributions change shape and develop a substantial “bump” for larger avalanches. This indicates that at (on average) 22 DIV the network enters a supercritical regime (Figure 3 B). After (on average) 36 DIV network activity is curbed and it reaches a subcritical regime (Figure 3 C). This can be seen by the decrease of the distribution at larger avalanches. At (on average) 58 DIV the system becomes finally critical (Figure 3 d). Here we find that the deviation from a power law is nearly zero (for these examples ). In general we find that the differences between all states are significant for the measured values of (ANOVA test). Figure 2 provides the data of all 20 cultures (see Methods) divided into the different states. All completely measured cultures undergo the same transitions from initial (black) to supercritical (red) to subcritical (green) and finally to a critical state (blue). The overlap between the first two states results from the very quick transition between them together with small differences in the speed of development of the different cultures. Average values of for these four states are , , , and (see Table in Figure 2). Differences are significant using the multiple comparison procedure with Bonferroni correction based on the one-way ANOVA test. Only the difference between the initial and critical state is not significant as in the initial state the network develops towards criticality until strong morphological changes set in (see Phase I). However, the activity given by the number of action potentials per minute is for the supercritical state significantly higher than for the initial, subcritical and critical state, which has the lowest mean activity. These were the only differences that were observed.
In summary, these results show that there is a characteristic time course in the development of the avalanche distributions. The system starts with low activity and then enters a transitory initial state. Quickly it leaves this state and, passing supercritical and subcritical regimes, reaches the critical state.
The first phase (Phase I) of the network development is characterized by dendritic growth to establish first synaptic contacts and to rise neuronal activities. At the beginning of the model development the dendritic acceptance increases (Figure 4 C). By this outgrowth the system creates synapses and forms a network. The distribution of the avalanches, the mean membrane potential , and the mean calcium concentration also changes (mean values over all neurons are given as upper case letters, while lower case letters indicate individual values). Similar to real cell cultures, all neurons at this phase are excitatory. With the help of a mean field approach it is possible to calculate average membrane potential and average calcium concentration during this phase. Different from real networks, where the activity is too small to render reliable measurements for very early developmental stages, in the model we can also analyze these. For this, the term , which determines the increase of the membrane potential according to the activity of the connected neurons in Equation 14 (see Methods), is simplified to a product of the mean membrane potential and an monotonous function dependent on the mean synaptic density (see below) and we get for the activity change:(1)The differential equation of the calcium concentration (Equation 15 in Methods) can be written as:(2)With these equations, we can now consider three different degrees of synaptic densities in the first phase ; namely zero, small, and medium densities and for Phase II with a large density.
Phase II of the network development is characterized by an overshoot in network activity. The membrane potential and calcium concentration ( and ) reach their maximum. This causes a phase transition in axonal and dendritic development: At that point, the dendritic acceptance begins to shrink and the axonal supply increases (see 4 C,D, Phase II). Moreover, during such transitions (accompanied with the formation of very many synapses) the action of the transmitter GABA switches from excitatory to inhibitory due to a change in the intracellular chloride concentration [30]. As we do not model changes in ion concentrations, we just change 20% of all neurons and assign them a negative value of , thereby making them inhibitory ( is changed to in this second phase). To determine the influence of different degrees of inhibition, the ratio of to is chosen differently in different experiments (Figure 5 D–F).
We can calculate the membrane potential as before with Equations 1 and 2 now with the constraint for a network without inhibition. As the membrane potential has by definition an upper limit of 1, the limit for to infinity during the phase transition (Phase II) is:(7)The calcium concentration has no upper limit and will theoretical rise to infinity(8)As the system remains only for a finite time in this second stage, will, however, remain finite. The mean membrane potential on the other hand reaches in the simulations indeed a value of 1 while the calcium concentration approaches (Figure 4 D).
If the membrane potential is close to one, neurons theoretically fire at every time step. Due to the given refractory period of 4 time steps, however, only out of 100 neurons fire on average in one time step. Without inhibition too many neurons are active during this stage and distributions cannot be reasonably assessed because one will only measure one or two “endless” avalanches (Figure 5 D).
Introducing inhibition changes this behavior substantially. The mean membrane potential decreases from to (Table 2) and the avalanche distribution shows now a measurable supercritical behavior (Figure 5 E,F). For measuring this avalanche distribution both excitatory and inhibitory neurons are considered. The membrane potential for the inhibitory neurons is larger than that for the excitatory neurons . This is due to their lower density (20% inhibitory as compared to 80% excitatory neurons).
As in Phase I, the network will not reach a steady state in Phase II, either. However, by contrast to the first phase where activity and connectivity is slowly growing, in the second phase, connectivity and activity is quickly getting overly strong (Figure 4 B–D, Phase I and II). Therefore, the system remains supercritical for the whole second phase until pruning is reducing connectivity to the homeostatic value in Phase III. Note, that stronger inhibition dampens the membrane potential and the firing rate considerably but does not influence the supercritical behavior of the system; (Table 2) remains essentially the same across five orders of magnitude of increased inhibition (see also Figure 5 E,F).
So far we have described the three development phases for our network model showing how criticality depends on network state, where the final state suggests some kind of fixed point behavior. In the following we will assess to what degree this process is characteristic for the system. To this end, we calculate its nullclines analytically [24] and compare these results with the simulations in Figure 8. For simplicity here we treat only a purely excitatory network.
To be able to solve the problem analytically we assume that the change of the connectivity between neurons and their membrane potential is slow and derivatives can, thus, be set to zero. Furthermore, on longer time scales the differences between neurons are negligible and only the behavior of the means need to be considered. As a result one can calculate the nullcline of this system (see Supporting Material Text S2), which describes a hysteresis curve (Figure 8 A):(10) and are the mean values of and over all neurons. is a sigmoidal function as an approximation for the Heaviside function , which determines when an action potential is generated (see Supporting Material Text S2). In Figure 8 A we also plot the trajectories which belongs to this system and the other (trivial) nullcline , which describes the fact that the system develops into homeostasis. At fixed point development stops. In Figure 8 B we plot the actual development of and observed in the simulations. Ideally this curve should match one of the trajectories in panel A and one can see that this is essentially the case. The main deviation arises from the fact that, due to the required simplifications, the analytical solution in panel A shows during the phase transition (Phase II) infinite growth and this cannot be achieved in the simulation. This leads to a reduction in the rising slope of panel B and to the fact that the fixed point is shifted closer to the inflexion point of the isocline.
When considering axons and dendrites separately, fixed point splits into a zone of many points, which correspond to the same connectivity , and hence lie on a hyperbola in Figure 8 C (dashed line). These fixed points form an omega-limit set in phase space and are represented by the equilibrium point in the --space. The approximate path of a trajectory from panels A and B is shown in Figure 8 C by the solid white line. Above we had stated that rates and membrane potentials are in Phase III fully invariant against system parameters and initial conditions, while connectivity is influenced by the level of inhibition. To this we can now add that the actual balance between axonal supply and dendritic acceptance (location of the different fixed points) remains dependent on the initial conditions (as well as on the inhibition) and should, therefore, be the most sensitive developmental parameter, e.g. much susceptible to pharmacological interference.
Furthermore, as the rate essentially follows and and inversely , we can state that the isocline in Figure 8 A will, for larger inhibition, be shifted diagonally upwards away from the origin shifting the fixed point to a higher synaptic density.
The dynamic behavior shown in Figure 8 is similar to that observed in the studies of Van Ooyen and Van Pelt [24] and our results show that the three development phases (Phase I, II and III) of this system are generic and independent of the chosen simulation parameters and confirm the existence of a strong phase transition.
Figure 9 shows a comparison of the different criticality states between cell culture (top) and model data (bottom) summarizing some of the observations from above. Additionally, the exponent of the avalanche distribution and the time bins are given in Table 4 for each state in model and cell cultures. In the model, at the end of the transition from Poisson to power law (Figure 5 C), little connectivity in Phase I leads to an initial state similar to that observable in dissociated cell cultures (Figure 9 A). This is followed by strongly rising synaptic density in Phase II (B). Accompanying the overshoot in network activity and connectivity, the model network passes a transient phase of supercriticality (B, bottom) as do the cell cultures (B, top). Depending on the chosen strength of inhibition, we obtain in Phase III a subcritical state for the model (C, bottom, ) similar to that in cell culture data (C, top). Thereafter, still in Phase III, we have gradually reduced the inhibitory strength to , hence balancing synaptic strength for inhibition and excitation (while keeping the number of inhibitory neurons constant). This leads to a final critical state in the model (D, bottom) similar to that found in cell culture data (D, top).
Thus, this predicts that that developing inhibition is an important factor for the course of criticality in developing neuronal networks. Only if inhibition in the model is lowered in Phase III again, the network becomes critical. Therefore, it is likely that overall synaptic pruning in developing networks not only affects excitatory but also inhibitory synapses [31]. Moreover, neuronal networks seem to reach firing rate homeostasis earlier than the equilibrium for maturing inhibition (compare discussion in Figure 7).
Additional inter-spike interval (ISI) and cross-correlation (CC) analyzes have been performed. ISIs and CCs are very similar between cultures and model across all stages but they do not contain interesting features (like oscillations) and therefore we do not show these diagrams here to save some space.
The following predictions are derived from the model:
These predictions are quite specific as they do not depend on the parameter choices in the model, which is one strength of this approach. Most predictions, if not all, can be tested in a straight forward way in future experiments, albeit requiring substantial and sometimes difficult experimental work which can only be addressed in future work.
In the current study we have investigated how the activity patterns in developing cell cultures can be measured and modeled in terms of self-organized criticality. We have shown that the activity distributions in real cultures undergo a transition from a stage with little activity to a supercritical and then a subcritical state and finally to critical behavior. These transitions were significant for the cell cultures analyzed.
We used an extended version of the neurite outgrowth model by Van Ooyen and co-workers [10], [24], [25] with separate axonal and dendritic fields. The axonal and dendritic growth is driven by the goal to reach firing rate homeostasis as modeled in previous papers by Butz and co-workers [26], [27]. The model was able to reproduce the different developmental phases and several interesting predictions have been made.
In general the chosen abstractions in the model appear to match the data description level quite well, but the question arises to what degree this still corresponds to reality in developing networks. Most importantly, the network stages described above must be related to the morphological development of dissociated neurons and their growing connectivity in culture which determines the activity pattern at every point in time [37]–[39].
It is well known [40] that the development of connectivity in cultures follows several phases. An initial phase (Phase I) is characterized by neuritic growth, followed (Phase II) by a structural overshoot and pruning followed by a maturation phase (Phase III) which finally leads to stable mean connectivity. Slowly growing connectivity in Phase I [41] leads over to the fast building of many synapses and a strong increase in activity in Phase II [38], [42], while pruning leads to Phase III with reduced number of synapses and lower activity [37]. Thereafter, firing behavior remains unchanged for two months [38], [42]. One may conclude that when synaptic pruning ceases, connectivity becomes stable and neuronal activities turn into homeostasis. Stable connectivity means that the sum of existing synapses does not vary much in time. The topology of the network can, however, not be predicted by the model as for this purpose a more detailed model of axons and dendrites would be required. However, neuronal development towards homeostasis substantially accelerates by increasing neuronal activities due to disinhibition by picrotoxin, a GABAergic synapse blocker [43]. Considering other transmitter studies, neuronal activity via increased glutamate release is likely to promote axonal outgrowth [44], [45] and therefore leads to a faster synapse formation and to an earlier maturation of the cell culture. Importantly, the behavior of dissociated neurons forming networks spontaneously occurs in any cell culture regardless of the original source of the plated neurons like cortex or hippocampus [46].
While certain simplifying assumptions had to be made to arrive at the basic differential equations (Equations 14–17) of our model, these experimental results clearly support the general dynamics assumed for our model.
In our model, networks with about 20% inhibition where the only ones that reached a robust critical state. While this level of inhibition corresponds to that in real nets, the results is intriguing as homeostasis of the firing rate will also be reached with much different levels of inhibitory cells. As known from literature [47], [48], GABA changes during the development from an excitatory to an inhibitory transmitter. As this is a fast process, inhibition sets in rapidly in the overshoot Phase II [48], [49] and possibly with a too high level. As discussed above, the observed subcritical phase clearly suggests a pruning phase for the inhibition which lasts longer than the firing rate equilibration. An indication of the functional role of synaptic pruning of inhibitory synapses was recently obtained from the developing auditory system in gerbils [31].
Like others [24], [25], [50], [51], also our model assumes that the main determining force within a growing network is the attempt of the neurons to achieve on average activity homeostasis. Several existing studies indicate that neurons, which are too active, seek to reduce their firing [50], [52], whereas neurons that are too quiescent try to increase it [53], [54]. Activity reduction is achieved by a reduction of the inputs to the cell (for example dendritic withdrawal) and vice versa. At the same time, highly activated cells respond with axonal outgrowth [44], [45], [55], [56] as increased levels of intracellular calcium, as a second messenger, regulates growth cone motility and therefore affects neurite outgrowth [44], [56]–[60].
Self-organized criticality represents the situation that many systems of interconnected, nonlinear elements evolve over time into a critical state in which the probability distribution of avalanche sizes can be characterized by a power law. This process of evolution takes place without any external instructive signal. As analytically shown [7], an important feature of the power law is its scale invariance. This means that all neuronal avalanches regardless of their size (number of spikes) can be treated as physically equal [3]. Furthermore, avalanches remain stable in their spatial and temporal configuration for many hours, as already shown in cortical slices [15]. So, avalanches have optimal preconditions (equality and stability) to be a candidate for memory patterns. The stability of these effects is strongly supported by the way our model systems develop as will be discussed next.
The current study shows that networks in cell cultures undergo a certain transition during their morphological development. Thus, this paper is in the tradition of a sequence of investigations [17], [40], [43], [45] that try to link cell culture activity and development to possible in vivo stages. Indications exist indeed that different activity states in cultures could be matched to in vivo states [61], but one needs to clearly state that culture and in vivo development also show clear differences. In vivo development is much more structured which will lead to differences in (ongoing) activity. As discussed above, dendritic and axonal fine structure and their spatial distribution, however, does not seem to critically affect the observed state-transitions. Hence, this supports that, at the level of avalanches, little difference might indeed exist between culture and in vivo. A study by Stewart and Plenz [21] suggests that avalanche frequency is correlated to the integrated amplitude of local field potentials, which grows until 25 DIV in their study. This indicates that also their networks had developed from a low-activity state into states that follow a power-law distribution. They show that distributions have in general an exponent of −1.5, indicative of a branching parameter of 1 [1], and a closer look at their result suggests that transitory (sub- and supercritical) stages are also observable in this data set (see, e.g., Fig. 3D in Stewart and Plenz 2008 [21]). A related study by Pasquale et al. [17] confirms this observation. It, thus, seems that the critical state represents the final state of the development, which – in the model – is reached together with firing rate homeostasis. This leads to a high degree of stability, which would be desirable also from a functional viewpoint. This is supported by the observation that in Phase III in the model sudden changes of the network structure (e.g. by a sudden change of inhibition) will only lead transiently to a stronger disruption of criticality. Indeed, the system soon find its way back into homeostasis and criticality is only little affected.
Several previous studies [1], [17], [21], [32] focused on the exponent of the power law in the critical state. This is a characteristic parameter of the system and found to be around [1], [17], [21], [32]. We find that the exponent is in simulations and in cell cultures. Thus, the exponent matches previous results very well for the simulations. The difference in the experiments from the theoretical value of can occur from variations in the time bin, too harsh selection criteria, or a too small number of data points. Thus, deviations leading to the found value of fall into the tolerance range of these experiments. In addition, it is not clear if the theoretical value of gained from branching processes [62] can be applied to all self-organizing systems in the critical state (f.e. Bak et al., 1987 [6]). Hence, it is equally well possible that the activity of the cultures is critical but does not exactly follow a branching process.
In a previous study Beggs and Plenz [1] have shown, that the critical state is optimal for a neuronal system concerning information flow. If the system is subcritical information will die out. The opposite situation is an epileptic system with too many long avalanches (supercritical state). Thus, a neuronal network in the critical state has the maximal dynamical range to react to incoming (external) information arriving from complex interactions of the neural system with its environment. The experimental part of the current study shows that real networks will develop towards such a state and the model suggests that this state is rather stable and therefore computationally reliable. Follow-up investigations, hopefully triggered by this research, might shed a light on the structural and functional dynamics of self-organized criticality in real developing brains and possibly also contribute a better understanding of developmental pathologies.
In order to investigate the relationship between network development and self-organized criticality, we extended the previous neurite outgrowth model by Van Ooyen and Van Pelt [10], [24], [25] by separate axons and dendrites. The model is essentially a two-dimensional recurrent neuronal network with uni-directional synapses. Model neurons are described by four equations; for activity , internal calcium concentration as well as dendritic acceptance , and axonal supply . The last two parameters determine the connectivity which is a generalisation of synaptic weights and the number of synapses between neurons. In line with previous experimental [27], [45], [50] and modelling studies [23], [24], [71], [72], the processes which determine the dynamics of this system can be summarized very briefly as: The activity of each neuron affects its calcium concentration. This, in turn, specifies the change of the dendritic and axonal offers, hence, the connectivity which will then gradually influence activity and so on. In the following we define parameters and equations. These equations are solved by the Euler method with an interval length of one simulated time step.
|
10.1371/journal.ppat.1002429 | Sialidases Affect the Host Cell Adherence and Epsilon Toxin-Induced Cytotoxicity of Clostridium perfringens Type D Strain CN3718 | Clostridium perfringens type B or D isolates, which cause enterotoxemias or enteritis in livestock, produce epsilon toxin (ETX). ETX is exceptionally potent, earning it a listing as a CDC class B select toxin. Most C. perfringens strains also express up to three different sialidases, although the possible contributions of those enzymes to type B or D pathogenesis remain unclear. Type D isolate CN3718 was found to carry two genes (nanI and nanJ) encoding secreted sialidases and one gene (nanH) encoding a cytoplasmic sialidase. Construction in CN3718 of single nanI, nanJ and nanH null mutants, as well as a nanI/nanJ double null mutant and a triple sialidase null mutant, identified NanI as the major secreted sialidase of this strain. Pretreating MDCK cells with NanI sialidase, or with culture supernatants of BMC206 (an isogenic CN3718 etx null mutant that still produces sialidases) enhanced the subsequent binding and cytotoxic effects of purified ETX. Complementation of BMC207 (an etx/nanH/nanI/nanJ null mutant) showed this effect is mainly attributable to NanI production. Contact between BMC206 and certain mammalian cells (e.g., enterocyte-like Caco-2 cells) resulted in more rapid sialidase production and this effect involved increased transcription of BMC206 nanI gene. BMC206 was shown to adhere to some (e.g. Caco-2 cells), but not all mammalian cells, and this effect was dependent upon sialidase, particularly NanI, expression. Finally, the sialidase activity of NanI (but not NanJ or NanH) could be enhanced by trypsin. Collectively these in vitro findings suggest that, during type D disease originating in the intestines, trypsin may activate NanI, which (in turn) could contribute to intestinal colonization by C. perfringens type D isolates and also increase ETX action.
| Clostridium perfringens type D strains cause enteritis and enterotoxemias in livestock after colonizing the intestines and then producing toxins, notably epsilon toxin (ETX). Initially produced and secreted in an inactive form, ETX can be rapidly proteolytically-activated by trypsin and other intestinal proteases. While most C. perfringens strains produce three sialidases, no pathogenic role has yet been identified for these enzymes that remove terminal sialic acid residues from glycoproteins and glycolipids. Our current study found that trypsin increases the activity of the NanI sialidase made by type D strain CN3718. This effect enhanced the ability of NanI to modify the surface of MDCK cells, leading to increased ETX binding and cytotoxicity. We also found that modification of the host cell surface by NanI sialidase allows efficient attachment of CN3718 cells to Caco-2 cells. These results identify interactions between intestinal proteases, ETX, sialidases, and ETX-producing bacteria, whereby trypsin activates not only ETX but also NanI sialidase. If similar effects occur in the intestines, the activated NanI sialidase may modify the host cell surface to facilitate bacterial attachment and thereby worsen disease by facilitating intestinal colonization by type D strains to prolong toxin delivery and, in some species, increase ETX binding.
| Clostridium perfringens, a Gram-positive, spore-forming anaerobe, is an important pathogen of both humans (causing, for example, gas gangrene and type A human food poisoning) and livestock (causing severe enterotoxemias and enteritis) [1]. The virulence of this bacterium is largely attributable to its ability to express a plethora of potent toxins. However, while C. perfringens can produce >15 different toxins, individual strains express only portions of this toxin arsenal [1]-[3]. Therefore, based upon production of four typing toxins (α, β, ε, and ι), isolates of this organism are commonly classified into five toxinotypes (type A through E) [4].
By definition, C. perfringens type D isolates must produce alpha and epsilon toxins, while type B isolates must express alpha, beta and epsilon toxin [4]. Beyond those typing toxins, type D and type B isolates commonly produce additional toxins, e.g., perfringolysin O, enterotoxin, TpeL or beta2 toxin [5]-[7]. Type B and D isolates cause enterotoxemias in livestock that initiate with toxin production in the intestines, followed by absorption of those toxins into the circulation to affect other internal organs, such as the brain and kidneys. Type D isolates can also cause acute or chronic enteritis in goats [1], [8], [9].
Epsilon toxin (ETX) is considered important for the virulence of both type B and type D isolates [5], [6]. This CDC class B select toxin, which ranks as the third most-potent clostridial toxin after the botulinum toxins and tetanus toxin, belongs to the aerolysin family of pore-forming toxins [9], [10]. ETX is synthesized and secreted as an inactive prototoxin of 311 amino acids (32.7 kDa). In the animal intestines, the prototoxin can be proteolytically activated to the fully-active toxin (274 amino acids) by trypsin and chymotrypsin [11].
To date, only a few ETX-sensitive cultured cell lines have been identified, including MDCK cells, mpkCCDc14 cells, and human leiomyoblastoma (G402) cells [11]. The mechanism of ETX action on MDCK cells is still under active study but first involves the binding of this toxin to unidentified protein receptors on the MDCK cell membrane surface. The bound toxin then uses lipid rafts to form a heptameric prepore complex on the membrane surface [12], [13]. When this prepore complex inserts into the cell membrane, an active pore is created that causes, or strongly contributes to, MDCK cell death [9].
Genome sequencing has revealed that C. perfringens strains typically possess three sialidase-encoding genes, named nanH, nanI and nanJ, which are located on a conserved region of the chromosome [14]-[16]. The nanH gene product, which is not secreted, is the ∼43 kDa NanH sialidase. The nanI gene product is the secreted ∼77 kDa NanI sialidase, while the nanJ gene product is the secreted ∼129 kDa NanJ sialidase. The nanH ORF shares only 19% nucleotide sequence identity with nanI and nanJ, but the ORFs encoding the two larger exoenzymes are more closely related, sharing 57% nucleotide sequence identity [17].
Sialidase (neuraminidases) production by some pathogenic bacteria has been implicated in their virulence. For example, Vibrio cholerae sialidase enhances the activity of cholera toxin by modifying surface gangliosides to create additional toxin receptors and thereby increase toxin binding levels [18]. Sialidases can also apparently contribute to virulence in other ways besides enhancing toxin binding. For example, neuramindases are thought to assist Streptococcus pneumoniae pathogenesis by providing nutrients for growth, assisting in biofilm formation, and enhancing colonization by exposing adhesion sites for this bacterium in the airways [19].
Possible sialidase contributions to C. perfringens virulence have received only limited attention. A recent study [20] used mutants to evaluate the potential pathogencity contributions of NanI and NanJ sialidase when C. perfringens type A strain 13 causes clostridial myonecrosis (gas gangrene). That study [20] found sialidases can enhance alpha-toxin-mediated cytotoxic effects in vitro, but also reported that sialidase production is not necessary for strain 13 to cause gas gangrene in a mouse model. Possible virulence contributions of sialidases when other C. perfringens strains, such as type D strains, cause disease originating in the intestines has received even less study. It was reported that sialidases can enhance ETX cytotoxicity towards MDCK cells [21], but an apparently contradictory conclusion was reached by another study determining that treatment of synaptosomal membrane fractions with sialidases lowered ETX binding levels [22].
The current study has used both biochemical and isogenic mutant approaches to better evaluate the possible sialidase enhancement of in vitro ETX action. In addition, a possible role for sialidases in facilitating C. perfringens cell adhesion to host cells was examined for the first time. Results of these in vitro studies suggest that sialidases could contribute to type D pathogenesis via at least two distinct mechanisms.
To clarify whether C. perfringens sialidases can enhance ETX cytotoxicity in sensitive mammalian cells, MDCK cells were pretreated with purified C. perfringens NanI sialidase and, after washing, those cells were challenged with purified ETX. The results showed that sialidase pretreatment substantially increased ETX-induced cytotoxicity (Figure 1A). This enhancement was stronger using a 0.005 U/ml vs. a 0.001 U/ml dose of sialidase for the pretreatment. However, while enhancement of ETX cytotoxicity increased up to a 90 min sialidase pretreatment using the 0.001 U/ml NanI dose, the higher 0.005 U/ml NanI dose caused no further increase in cytotoxicity beyond a 60 min pretreatment (Figure 1A). Sialidase treatment alone, or treatment with prototoxin, did not increase cytotoxicity above background levels (data not shown).
Experiments were then performed to explore the mechanistic basis behind the sialidase-induced enhancement of ETX cytotoxicity for MDCK cells that was shown in Figure 1A. First, MDCK cells were pretreated with purified C. perfringens NanI sialidase and, after washing, those cells were incubated with Alexa Fluor 488-labeled epsilon prototoxin (AF488-pETX), which was used since it binds similarly as active ETX to MDCK cells, yet causes no cytotoxicity [23]. Results from this experiment (Figure 1B) showed that pretreating MDCK cells with either 0.001 U/ml or 0.005 U/ml of purified NanI sialidase substantially increased subsequent AF488-pETX binding levels. This enhancement of toxin binding increased with longer sialidase pretreatment time, although no further increase in AF488-pETX binding was noted after a 60 min sialidase pretreatment beyond the 0.005 U/ml sialidase dose (Figure 1B).
An experiment then assessed the ability of sialidase pretreatment to increase formation of the ETX oligomeric complex that is considered responsible for pore formation-induced ETX cytotoxicity [9]. SDS-PAGE analyses of MDCK cell lysates treated with 5 µg/ml of Alexa Fluor 488-labeled ETX (AF488-ETX) detected an increased formation of the ETX oligomeric complex after NanI pretreatment (Figure 1C and 1D). This effect increased up to a 60 min pretreatment with 0.005 U/ml of NanI. Lastly, similar NanI pretreatment also increased MDCK cell binding and complex formation using a 10 µg/ml dose of AF488-pETX or AF488-ETX, respectively (data not shown).
The Figure 1 results indicated that pretreating MDCK cells with purified C. perfringens NanI sialidase can enhance ETX cytotoxicity by increasing toxin binding and thus, complex formation. Therefore, we next sought to assess whether sialidases also contribute to ETX-induced cytotoxicity when these glycoside hydrolyases are expressed at natural levels and in the natural presence of other C. perfringens exoproducts, including other toxins. To initiate this work, we first surveyed sialidase activity in 6 h and overnight supernatants from cultures of various C. perfringens isolates (Figure 2). This survey detected considerable strain-to-strain variations in supernatant sialidase activity and also identified type D strain CN3718 as a moderately high sialidase producer. Since CN3718 is also transformable and produces ETX (see below), it was chosen for construction of sialidase mutants.
PCR analyses first determined that CN3718 carries all three identified C. perfringens sialidase genes (data not shown), including nanI and nanJ, which encode the secreted NanI and NanJ sialidases, and nanH, which encodes the NanH sialidase that lacks a signal peptide and thus localizes in the cytoplasm. Therefore, the current study used a Clostridium-modified Targetron insertional mutagenesis method [24] to construct isogenic single nanH, nanI, and nanJ null mutants, a nanI/nanJ double null mutant, and a nanH/nanI/nanJ triple null mutant in a CN3718 background.
The identity of the CN3718 nanJ, nanI and nanH single null mutants (named BMC201, BMC202 and BMC203, respectively) was first demonstrated by PCR using primers specific for internal nanJ ORF, nanI ORF or nanH ORF sequences. Using DNA from wild-type CN3718, these internal PCR primers specifically amplified the expected PCR products of 306 bp for nanJ, 467 bp for nanI, and 285 bp for nanH (data not shown). However, consistent with the insertion of an ∼900 bp intron into the target ORF, the same primers amplified PCR products of ∼1200 bp, ∼1400 bp, and ∼1200 bp for the nanJ, nanI and nanH null mutants, respectively (data not shown). Those PCR assays also confirmed the identity of a nanJ/I double null mutant (BMC204) and a nanJ/I/H triple null mutant (BMC205) constructed by Targetron technology (data not shown). The nanJ/I double null mutant PCR supported amplification of two larger bands, which matched the product sizes of a disrupted nanJ gene and nanI gene, along with a small band matching the ∼285 bp product size for the wild-type nanH gene (data not shown). Using BMC205 DNA, PCR amplified larger bands for all three sialidase ORFs (data not shown).
After curing the intron delivery plasmids from the BMC201, BMC202, BMC203, BMC204 or BMC205 sialidase null mutants, DNA from the wild-type strain and each null mutant was subjected to Southern blot analysis (Figure 3A) using an intron-specific probe [25]. No probe hybridization to wild-type DNA was detected, as expected. In contrast, the presence of a single intron insertion was visible on this Southern blot using DNA from each single sialidase null mutant. In addition, two intron insertions were detected using DNA from the BMC204 nanI/J double null mutant and three intron insertions were noted using DNA from the BMC205 nanI/J/H triple null mutant.
To initiate a phenotypic characterization, vegetative growth of the isogenic mutants was compared in Todd Hewitt (TH) medium. After similar inoculation into TH medium, the vegetative growth rate of the isogenic mutants was nearly identical to the CN3718 parent (data not shown). Western blots then compared sialidase expression between CN3718 and the isogenic single, double and triple sialidase null mutants. These analyses demonstrated that sonicated 8 h TH cultures of the wild-type strain contain the 129 kDa NanJ, the 77 kDa NanI and the 43 kDa NanH (Figure 3B). In contrast, the BMC201 null mutant only produced NanI and NanH, the BMC202 null mutant only produced NanJ and NanH, and the BMC203 null mutant produced only NanI and NanJ (Figure 3B). Similar Western blotting of sonicated cultures (Figure 3C) showed that the BMC204 nanI/J double null mutant only produced the 43 kDa NanH, while the BMC205 nanJ/I/H triple null mutant did not produce any sialidase protein.
To evaluate the effects of mutating each sialidase gene on the total sialidase activity of CN3718, sialidase activity assays were carried out using either 8 h TH culture supernatants (to measure exosialidase activity) or sonicated 8 h TH culture supernatants (to measure total sialidase activity) from the wild-type parent and each null mutant strain. Collectively, the results (Figure 3D) indicated that the exosialidase activity measured in supernatants from the wild-type strain is mainly attributable to NanI. The Figure 3D results also confirmed that NanH predominantly accumulates in the cytoplasm since 8 h supernatants of the BMC204 nanI/J double mutant exhibited almost no sialidase activity, while the sonicated culture supernatant of this double mutant possessed significant sialidase activity. Lastly, this analysis demonstrated that CN3718 produces only the three recognized sialidases since the supernatant and sonicated culture of the BMC205 nanI/J/H triple null mutant lacked sialidase activity.
Since a major goal of the current study was to evaluate a possible relationship between sialidases and ETX action, ETX expression was measured for wild-type CN3718 and the isogenic sialidase null mutants. Surprisingly, each single sialidase mutant, but especially the BMC202 nanI null mutant, showed decreased ETX expression compared to the wild-type parent (Figure 3E and 3F). ETX expression levels by the BMC204 nanI/J double null mutant and the BMC205 nanI/J/H triple mutant decreased further compared against ETX production levels by wild-type CN3718 (Figure 3E and 3F).
The considerable ETX expression differences between wild-type strain CN3718 and the sialidase mutants shown in Figure 3E and 3F precluded using those strains for comparative ETX cytotoxicity experiments aimed at evaluating sialidase contributions to ETX action under natural sialidase expression levels and in the presence of natural levels of other secreted C. perfringens products. To overcome this complication, etx null mutations were engineered into the CN3718 wild-type strain and the BMC205 triple sialidase null mutant strain to create BMC206 and BMC207, respectively. The availability of these two etx mutants would allow later comparative cytotoxicity experiments using equivalent amounts of ETX added to supernatants of these strains. These etx mutants were constructed by a Targetron insertional mutagenesis approach. PCR confirmed the presence of an intron insertion into the etx gene (Figure 4A). Using DNA from wild-type CN3718, internal etx PCR primers specifically amplified a PCR product of 117 bp. However, consistent with a 900 bp intron insertion into the etx ORF, the same primers amplified a PCR product of ∼1000 bp using DNA from the putative etx mutants of both BMC206 and BMC207. PCR approaches also confirmed that BMC207 maintains intron insertions in its nanI, nanJ and nanH ORFs (Figure 4B).
After curing the intron delivery plasmids from these mutants, DNA was isolated from the two putative etx null mutants and subjected to Southern blotting using an intron-specific probe (Figure 4C). This analysis detected no probe hybridization with wild-type CN3718 DNA, as expected. As no suitable enzyme was identified that could separate the four intron copies present in BMC207, EcoRI was used to distinguish the presence of the etx intron from the other three sialidase introns, whose presence in the BMC205 mutant had already been demonstrated (Figure 3A). By this approach, the presence of the etx intron was clearly shown in BMC206 and BMC207 by Southern blotting (Figure 4C).
To initiate phenotypic comparisons, the growth rates of the BMC206 and BMC207 etx null mutants were first determined to be similar to wild-type CN3718 in TH medium (data not shown). Western blotting then confirmed the lack of ETX expression by these etx null mutants (Figure 4D). For later experiments it was important to demonstrate that culture supernatants could be supplemented with purified ETX to achieve equivalent ETX levels. This was confirmed by adding 1, 5 or 10 µg/ml of ETX to supernatants of the two etx null mutants (Figure 4D).
Sialidase Western blotting and sialidase activity analyses (Figure 4E and 4F) were next performed using cultures or sonicated cultures (to release NanH) of wild-type CN3718, BMC206 and BMC207 to evaluate whether introducing an etx gene mutation had affected sialidase expression. Western blot results demonstrated comparable levels of NanJ, NanI and NanH production by wild-type CN3718 and BMC206 when grown for 8 h in TH. However, as expected, production of these three sialidases was absent from similar cultures of BMC207 (Figure 4E). Measurement of sialidase activity in sonicated cultures (Figure 4F) further confirmed those Western blotting results. The sialidase activity detected in TH sonicated lysates or TH culture supernatants of CN3718 vs. BMC206 was very similar. However, no sialidase activity was measured in similarly prepared BMC207 sonicated lysates or culture supernatants.
For preparing a nanI complementing strain of BMC207, we first constructed pJIR750Icomp, which consists of the CN3718 nanI ORF, 500 bp of upstream sequence and 300 bp of downstream sequence cloned into the C. perfringens/E. coli shuttle plasmid pJIR750. A nanJ complementing plasmid, named pJIR751Jcomp, was similarly prepared by PCR-amplifying a product corresponding to the nanJ ORF, 1000 bp of upstream sequence and 400 bp of downstream sequence, and then cloning that PCR product into the C. perfringens/E. coli shuttle plasmid pJIR751. Finally, by the same method, a nanH complementing plasmid named pJIR751Hcomp was prepared that contains 500 bp of upstream sequence, the nanH ORF and 300 bp of downstream sequence cloned into pJIR751. The three sialidase complementing plasmids were then individually transformed into BMC207 by electroporation. PCR confirmed the resultant transformants contained nanI, nanJ or nanH wild-type genes (Figure 5A). Using primers to internal sequences of each sialidase gene and BMC207 DNA, the nanI, nanJ or nanH PCR products amplified large bands indicating an intron insertion. In contrast, using DNA from the three complementing strains, the PCR products amplified were of smaller size, consistent with their containing a wild-type sialidase gene, which is preferentially amplified over the larger intron-disrupted sialidase gene also present in these complementing strains.
Western blot analyses then assessed sialidase expression by the complementing strains (Figure 5B). As expected, BMC206 produced the ∼43 kDa NanH, the ∼77 kDa NanI and the ∼130 kDa NanJ. In contrast, the BMC207 null mutant expressed no sialidase proteins. These Western blots also showed that the nanH complementing strain BMC2073 only produced NanH, the nanI complementing strain BMC2072 only made NanI, and the nanJ complementing strain BMC2071 only expressed NanJ (Figure 5B).
The sialidase activity of the complementing strains was also assessed (Figure 5C). When sialidase activity was measured in 2 h, 4 h, 6 h, 8 h, 10 h or overnight TH cultures, the nanI complementing strain BMC2072 started expressing NanI as early as 2 h in TH culture. In contrast, BMC206 and the nanJ complementing strain began producing sialidase at 4 h. Wild-type CN3718 also expressed sialidase starting at 4 h (data not shown). The NanI complementing strain BMC2072 produced more sialidase activity than other complementing strains. Since the BMC2072 nanI single complementing strain already expressed more sialidase activity than either CN3718 or BMC206, and also due to technical limitations, no double or triple sialidase complementing strains were prepared.
The availability of sialidase mutants and complementing strains allowed us to evaluate whether specific sialidases can enhance ETX binding to MDCK cells at natural sialidase expression levels and in the background of other secreted C. perfringens products. The same 5 µg/ml amount of AF488-pETX was added to concentrated supernatants from BMC206, the BMC207 triple sialidase null mutant strain, or the three BMC207 sialidase complementing strains (i.e. BMC2071, BMC2072 and BMC2073). Note that, i) a natural 1× concentration of C. perfringens culture supernatant was finally present in these MDCK cell cultures, ii) a 5 µg/ml ETX concentration corresponds to a typical natural supernatant concentration of this toxin for ETX-producing C. perfringens strains [5], [6] and iii) these binding experiments were performed at 37°C to obtain the enzymatic effects of sialidases (when present), but previous studies [13] have shown that prototoxin binding to MDCK cells is equivalent at 4°C or 37°C.
When these mixtures were added to MDCK cells, the fluorescent readings showed that the BMC207 supernatant sample supported less AF488-pETX binding to MDCK cells compared against the BMC206 supernatant (Figure 6). Supernatant from the BMC2072 nanI complementing strain exhibited even greater levels of AF488-pETX binding than did BMC206 supernatant after a similar supplementation with the labeled prototoxin. In contrast, a smaller increase in AF488-pETX binding was noted using supernatants of the nanJ or nanH complementing strains supplemented with labeled prototoxin.
Figure 1 results demonstrated that pretreating MDCK cells with purified C. perfringens NanI sialidase increases subsequent ETX cytotoxicity. Results shown in Figure 7A confirmed this conclusion and demonstrated that this effect is not due to direct sialidase-induced cytotoxicity.
The Figure 6 binding results suggested that, compared against supernatants from BMC207, increased MDCK cell cytotoxicity should also be observed using supernatants from BMC206 or the nanI complementing strain BMC2072 when those supernatants were supplemented with ETX. Verifying this suggestion, the ETX-supplemented culture supernatants of BMC206 strain caused about 30% more MDCK cell cytotoxicity than did similarly ETX-supplemented BMC207 supernatants (Figure 7B). Furthermore, complementation with the wild-type nanI gene substantially increased the cytotoxic effects of the sialidase-deficient BMC207 mutant, while complementation of the BMC207 mutant with the wild-type nanJ or nanH genes increased cytotoxicity to a lesser extent.
During natural disease, ETX is secreted into the animal intestines as an inactive prototoxin of ∼33 kDa, which can then be activated by intestinal proteases such as trypsin or chymotrypsin [11]. Interestingly, when CN3718 wild-type supernatants were trypsin-treated to activate prototoxin (data not shown), overall sialidase activity in these supernatants also significantly increased. This result was confirmed by treating purified C. perfringens NanI with trypsin. As shown in Figure 8A, the trypsin-treated NanI possessed significantly increased sialidase activity compared against the same amount of non-trypsin-treated NanI. When these samples were subjected to Western blotting, the results indicated that most of the trypsin-treated NanI ran as a slightly smaller protein than native NanI, supporting proteolytic processing (Figure 8B). In addition, chymotrypsin also activated NanI sialidase activity (data not shown).
A further experiment then compared ETX cytotoxicity in the presence of purified NanI sialidase that had, or had not, been trypsin-activated prior to mixing with purified active ETX. After blocking trypsin activity from the sialidase sample with trypsin inhibitor, an enhancement of ETX cytotoxicity was observed using the trypsin-activated sialidase compared against the same amount of sialidase that had not been trypsin-activated (Figure 9A). Trypsin-activated sialidase alone caused no increase in cytotoxicity above background.
Similar experiments using concentrated C. perfringens culture supernatants showed that trypsin activation of these supernatants enhanced MDCK cell cytotoxicity by about 25% for BMC206 (Figure 9B). Further supporting the involvement of sialidases in this trypsin-induced increase in the cytotoxic properties of BMC206, similar trypsin treatment of BMC207 supernatants did not increase ETX-induced cytotoxicity (Figure 9B).
The Figure 8A results indicated that NanI can be activated by trypsin. To assess whether the other two C. perfringens sialidase (NanJ and NanH) can also be trypsin-activated, the supernatants of nanI, nanJ and nanH complementing strains were similarly treated with trypsin and their sialidase activity was measured. The results obtained indicated that only the sialidase activity of NanI is enhanced by trypsin (Figure 10A). In contrast, NanJ and NanH sialidase activity slightly decreased after similar trypsin treatment (Figure 10A). Consistent with this result, trypsin increased the MDCK cell cytotoxicity of supernatants containing only NanI, but not supernatants containing only NanJ or NanH (Figure 10B).
During natural enterotoxemias, type D strains remain present in the intestines, where they contact host enterocytes [8]. Therefore, an experiment examined whether contact with cultured mammalian cells for 6 hr (including intestinal Caco-2 and HT-29 cells as well as MDCK, Vero and NT6 fibroblasts) might affect the exosialidase activity of BMC206. This study revealed that contact between BMC206 and Caco-2, HT-29, or MDCK cells resulted in increased culture supernatant sialidase activity, although this effect was weaker using MDCK cells. In the absence of BMC206 infection, supernatant sialidase activity for Caco-2, HT-29 and MDCK cell cultures did not increase above background (Figure 11A), suggesting that the increased culture supernatant sialidase activity measured in the BMC206-infected cultures was not from secreted mammalian sialidase activity. The results also detected no culture supernatant sialidase activity upon infection of these mammalian cells with the triple sialidase null mutant BMC207. In contrast to the stimulation of sialidase activity observed using Caco-2, HT-29 or MDCK cells, exosialidase activity did not increase after BMC206 infection of Vero cell or NT6 fibroblast cultures (Figure 11A).
The nature of this host cell-induced increase in sialidase activity was further investigated using a Caco-2 cell infection model (Figure 11B). This study detected no culture supernatant sialidase activity upon infection of the triple null mutant BMC207 with Caco-2 cells up to 6 h, consistent with the increased sialidase activity detected upon BMC206 infection of Caco-2 cells involving upregulated sialidase expression by BMC206 upon Caco-2 cell contact. Results obtained using the BMC2072 NanI complementing strain suggested this upregulation of sialidase activity upon contact of BMC206 with Caco-2 cells primarily involves increased NanI production (Figure 11B). To definitively resolve whether contact with Caco-2 cells upregulates NanI expression by BMC206, quantitative RT-PCR studies were performed (Figure 11C). Those studies clearly demonstrated a substantial increase in nanI transcription in the presence of Caco-2 cells.
When causing disease originating in the intestines, C. perfringens type D vegetative cells likely adhere to intestinal tissue to colonize and sustain toxin production [8]. Relatively little is known about C. perfringens adherence to host cells, particularly enterocytes during enteritis or enterotoxemia. However, since enterocytes are coated with a variety of sialic acid-containing glycoconjugates, sialic acid residues could conceivably modulate attachment of these bacteria to the intestines.
Therefore, experiments tested whether BMC206 or BMC207 can attach to cultured mammalian cells, including Caco-2 cells, HT-29 cells, MDCK cells, Vero cells and NT6 fibroblasts. As shown in Figure 12A and 12B, BMC206 attached well to Caco-2 and HT-29 cells, but substantially less well to the other surveyed host cells. BMC206 attachment to Caco-2 and HT-29 cells involved sialidase production since BMC207 failed to attach to those two mammalian cell lines (Figure 12A and 12B).
Additional studies were performed with the human enterocyte-like Caco-2 cell line as an in vitro model to examine whether sialidases play a role in CN3718 adherence to intestinal cells (Figure 13A and 13B). As already established (Figure 12A), BMC206 attached well to Caco-2 cells. In contrast, there was little or no attachment of BMC207 to these host cells under the same experimental conditions. However, the BMC2072 nanI complementing strain exhibited Caco-2 cell adhesion levels equal to, or exceeding those of, BMC206 (Figure 13A and 13B). In contrast, the Caco-2 adhesion ability of the nanJ and nanH complementing strains was not substantially increased over the attachment of BMC207 (Figure 13A and 13B).
A role for NanI sialidase in modifying the Caco-2 cell surface to increase the adherence of CN3718 vegetative cells to these host cells was further supported by results of an experiment where Caco-2 cells were pretreated with purified NanI. This pretreatment substantially increased the Caco-2 cell adherence of the BMC207 mutant (data not shown).
Phase-contrast and immunofluorescence microscopy analyses (Figure 13C) confirmed the strong adhesion of BMC206 or BMC2072 vegetative cells to Caco-2 cells. This microscopy detected many adherent bacteria attached to Caco-2 cells, often along their edges. In contrast, few (if any) BMC207 cells were adherent to Caco-2 cells.
A recent study [20] showed that a nanI mutation nearly completely abolished the sialidase activity of C. perfringens type A strain 13, which produces NanI and NanJ. However, unlike strain 13, many C. perfringens strains can also produce NanH [14]. Using a combination of nanI, nanJ or nanH single null mutants, a nanI/nanJ double null mutant, and a sialidases triple null mutant, the current study evaluated the relative contribution of each sialidase to the exosialidase activity and total sialidase activity of C. perfringens type D strain CN3718. These analyses revealed that, using the sialidase assay conditions employed in this study, i) CN3718 produces all three recognized C. perfringens sialidases but no additional unknown sialidases, ii) NanI is responsible for most exosialidase activity of this strain and iii) NanH is also a major contributor to the total sialidase activity possessed by CN3718. Furthermore, by demonstrating that CN3718 produces all three sialidases, the current results argue against proposals [17] that only myonecrosis strains of C. perfringens produce all three sialidases. Bioinformatics analyses of database sequences support carriage of three sialidase genes as the norm for C. perfringens strains, regardless of origin. In addition to our current CN3718 findings, analysis of the 8 completed or partially-completed C. perfringens genomes revealed that all three sialidase genes are present in 6 of those 8 isolates, which include several isolates from humans or livestock suffering from C. perfringens disease originating in the intestines. The two exceptions are strain 13 and SM101 that, respectively, lack nanH or both nanI and nanJ [14], [16].
Previous studies with other bacterial pathogens showed that sialidases can contribute to virulence in several ways, e.g., by exposing sites to facilitate more bacterial adhesion or toxin binding to host cells [19], [26]. Our results indicated that NanI sialidase increases the ETX sensitivity of MDCK cells. The mechanism of this enhancement was shown, for the first time, to involve an increase in toxin binding, suggesting NanI exposes additional ETX receptors on the host cell surface. Alternatively, NanI could modify host cell surface nonreceptors so they acquire ETX binding ability, similar to the situation in Vibrio cholerae, where a sialidase modifies glycolipids to increase cholera toxin binding [18]. We also demonstrated that increased ETX binding to NanI-exposed MDCK cells leads to more ETX complex formation, which translates to greater pore formation and host cell death [23]. Overall, these conclusions are consistent with, and explain for the first time, observations from a previous study reporting that C. perfringens sialidases can increase ETX cytotoxicity towards MDCK cells [21]. They are also consistent with previous results showing that soluble sialic acid cannot inhibit ETX binding to, or action on, MDCK cells [21]. However, the current results apparently contrast with another study [22] reporting that sialidases lower ETX binding to brain synaptosomal membrane vesicles. It is not immediately clear whether the varying conclusions amongst these studies are attributable to using MDCK cells vs. synaptosomal membranes or involve some other explanation.
Sialidases assist the ability of some pathogens to adhere to host cells or tissues. For example, sialidase contributions to adherence and colonization are well-documented for S. pneumoniae [19]. However, prior to the current work, only one study had addressed C. perfringens vegetative cell adherence to the mammalian intestinal epithelium, despite the likely importance of this process for pathogenesis; that earlier study suggested a putative C. perfringens collagen adhesion protein (CNA) might contribute to porcine enteritis by promoting adhesion of this bacterium to damaged intestinal tissue [27]. Our study now reports the first evidence that adherence of a C. perfringens strain to certain host cells, including enterocyte-like Caco-2 cells, is facilitated by sialidases, especially NanI. Inactivation of NanI sialidase production decreased adhesion of CN3718 strain by at least 20-fold and this effect was reversible by complementation. Furthermore, pretreating Caco-2 cells with purified NanI allowed adherence of the BMC207 mutant unable to produce any sialidase, indicating that NanI modifies the Caco-2 cell surface to render it favorable for CN3718 adherence. NanH and NanJ appeared to play a lesser role in facilitating C. perfringens adherence to Caco-2 cells based upon complementation results using the triple sialidase mutant. The varying ability of the three sialidases to enhance ETX binding or C. perfringens adherence likely reflects the reported enzymatic difference in sialidase kinetic properties and substrate specificities [17], although further comparative study of the properties of C. perfringens sialidases is warranted.
These in vitro findings open the possibility that sialidases, particularly NanI, contribute to virulence by facilitating C. perfringens adherence in the intestines, a hypothesis that should be tested in animals. In addition, future studies should identify the adhesins mediating CN3718 binding to certain host cells. Beyond the CNA study [27] mentioned in the preceding paragraph, the only other available information regarding molecular mechanisms of C. perfringens adhesion to host cells is a report implicating type IV pilin in type A strain 13 attachment to myoblasts [28]. Future studies will examine whether the type IV pilus or CNA protein contribute to CN3718 adherence to host cells.
A notable finding of the current work was that CN3718 adhesion is specific for certain mammalian cells. Notably, this strain adheres well to Caco-2 and HT-29 cells, both of which are intestinal cell lines. Much lower attachment of this strain was detected using cell lines of nonintestinal origin, including MDCK cells (canine kidney cells), Vero cells (African green monkey kidney cells) or NT6 rat fibroblasts cells. Furthermore, contact of CN3718 with the intestinal cells lines, and MDCK cells to a lesser degree, also stimulated nanI transcription. Collectively, these results suggest that CN3718, a type D disease strain, is well-adapted for sensing the presence of enterocytes and then responding by producing more sialidase to facilitate its specific attachment to these cells. However, a survey including many additional cell lines and C. perfringens strains are needed to test this hypothesis.
The current study also suggests that trypsin could be a greater contributor to type D disease than previously appreciated [29]. It is well established that, following secretion, the inactive ETX prototoxin must be activated by protease-induced removal of residues from the N- and C-terminus [29]. This prototoxin activation normally occurs in the gut of infected animals, where it likely involves intestinal proteases, including trypsin. Interestingly, the current study found that trypsin proteolytically activates NanI as well as ETX, thereby increasing sialidase activity. Since we also showed that trypsin activation of NanI enhances ETX binding and cytotoxicity towards MDCK cells, trypsin-activated NanI could similarly increase ETX action in vivo. Furthermore, since NanI was also determined to be important for CN3718 adherence to Caco-2 cells, trypsin activation of NanI could also contribute to type D disease by promoting C. perfringens colonization in the intestines. Chymotrypsin also activated NanI sialidase activity (data not shown), so that intestinal protease could further enhance disease.
Another possible virulence role for sialidases is suggested by our observation that various CN3718 sialidase mutants, particularly nanI mutants, exhibited significantly decreased ETX production levels despite wild-type growth properties (data not shown). This result might indicate that sialic acid signals ETX production, but initial attempts to prove this hypothesis have yielded inconclusive results (data not shown). Interestingly, previous studies showed that inactivating the nanI gene in strain 13 also affects toxin production levels [20]. However, for that strain, the nanI mutant showed increased production levels of alpha-toxin and perfringolysin O. Whether the increased production of alpha toxin and perfringolysin O by strain 13 involves a sialic acid signal should also be evaluated in the future.
While our findings suggest NanI plays, at minimum, two virulence roles for CN3718, i.e., enhancing both ETX binding and attachment of this strain to enterocytes, it remains unclear why this strain also produces NanJ and NanH sialidases. As mentioned, NanH and NanI have different temperature optimums, kinetic properties and substrate specificities [17]; the enzymatic characteristics of NanJ, which was only discovered during genome sequencing studies [14] remains uncharacterized. Possibly NanH and NanJ are important for growth or survival in specific environments such as soil or sewage or in other infections. Furthermore, NanJ possesses additional domains of unknown function that are missing from NanI [17]; those additional NanJ domains might contribute to growth or virulence under conditions that remain to be identified.
In summary, our in vitro results suggest that, while NanI sialidase does not appear to be involved in the pathogenesis of type A isolates during myonecrosis [20], it could contribute to type D enterotoxaemia and enteritis in at least two ways. First, NanI (particularly after trypsin activation in the intestines) could increase ETX binding and complex formation, thereby potentiating ETX cytotoxicity. The potential relevance of this effect for damage to non-intestinal target organs, such as brain and kidneys, is unclear since it has not yet been evaluated whether NanI sialidase can be absorbed into the circulation during type D enterotoxemias. Similarly, ETX causes limited damage in the intestines of sheep [8], so sialidase potentiation of ETX may have less importance during ovine type D disease. However, sialidases could possibly increase ETX binding to the intestines of sheep or goats and thus facilitate absorption of this toxin into the circulation. Moreover, ETX does substantially damage the caprine gastrointestinal tract [8], [30], so, by enhancing ETX intestinal toxicity, trypsin-activated NanI could directly contribute to caprine type D enteritis. A second possible virulence contribution of NanI may be to increase C. perfringens vegetative cell adherence to the intestinal epithelium, thereby contributing to type D infections by facilitating colonization of ETX-producing bacteria in the intestines. Lastly, our results suggest that contact with intestinal cells may stimulate C. perfringens to produce more NanI, thereby further potentiating (particularly after trypsin and chymotrypsin activation) ETX effects and bacterial adherence during disease. The relevance of these in vitro findings for possible sialidase virulence contributions should now be tested experimentally in animal models.
CN3718, an ETX-positive C. perfringens type D animal disease strain (Table 1), originated from the Burroughs-Wellcome collection and was obtained via Dr. R. G. Wilkinson [5]. NCTC8346, another type D animal disease isolate, was used for ETX toxin purification [5]. E.coli Top10 cells (Invitrogen) were used as the cloning host. Mutant, complementing strains, and plasmids used in this study are listed in Table 1.
Media used in this study for culturing C. perfringens included FTG medium (fluid thioglycolate medium; Difco Laboratories); TH medium (Bacto Todd Hewitt Broth [Becton-Dickinson], with 0.1% sodium thioglycolate [Sigma Aldrich]); TGY medium (3% tryptic soy broth [Becton-Dickinson], 2% glucose [Fisher scientific], 1% yeast extract [Becton-Dickinson], and 0.1% sodium thioglycolate [Sigma Aldrich]) and BHI agar plates (brain heart infusion, Becton-Dickinson). For culturing E. coli, Luria-Bertani (LB) broth (1% tryptone [Becton-Dickinson], 0.5% yeast extract [Becton-Dickinson], 1% NaCl [Fisher scientific] and LB agar (1.5% agar [Becton-Dickinson]) were used. All antibiotics used in this study were purchased from the Sigma-Aldrich Chemical Company or Fisher Scientific Company.
Purified C. perfringens neuraminidase (NanI) was purchased from Roche Applied Science. ETX prototoxin was purified to homogeneity using a previously described method [5]. Following the method described by the manufacturer, the purified ETX prototoxin was fluorescently-labeled using an Alexa Fluor 488 protein labeling kit (Invitrogen), creating AF488-pETX. For ETX detection by Western blots, an ETX-specific monoclonal antibody (5B7; kindly provided by Paul Hauer, Center for Veterinary Biologics, Ames, Iowa) was used as primary antibody, followed by rabbit anti-mouse immunoglobulin G (IgG)-peroxidase conjugate (Sigma) as a secondary antibody. A polyclonal rabbit antibody against C. perfringens neuraminidases was purchased from Thermo Scientific.
The nanJ, nanI and nanH genes of CN3718 were each inactivated by insertion of a group II intron via the Clostridium-modified TargeTron gene knock-out system [24]. Using intron insertion sites identified by the Sigma TargeTron algorithm, an ∼900 bp intron was targeted into the nanJ, nanI and nanH ORFs in a sense orientation. The intron insertion was targeted between nucleotides 657 and 658 of the nanJ ORF. The primers used for PCR targeting the nanJ intron were 657/658-IBS, 657/658-EBS1d and 657/658-EBS2 (Table 2). The intron insertion was targeted between nucleotides 730 and 731 of the nanI ORF using primers 730/731-IBS, 730/731-EBS1d and 730/731-EBS2 (Table 2). The intron insertion was targeted between nucleotides 707 and 708 of the nanH ORF. The primers used for PCR targeting the intron into nanH were 707/708-IBS, 707/708-EBS1d and 707/708-EBS2 (Table 2).
The 350 bp intron PCR products were inserted into pJIR750ai between the HindIII and BsrGI enzyme sites in order to construct nanJ, nanI and nanH-specific TargeTron mutagenesis plasmids. The resultant plasmids, named pJIR750nanJi, pJIR750nanIi and pJIR750nanHi respectively (Table 1), were individually electroporated into wild-type CN3718. This produced nanJ, nanI and nanH single null mutants named, respectively, BMC201, BMC202 and BMC203 (Table 1).
The CN3718 transformation efficiency was ∼240 transformants/µg plasmid DNA. Transformants were selected on BHI agar plates containing 15 µg/ml of chloramphenicol. Transformant colonies were then identified by colony PCR using primers nanJKOF and nanJKOR (Table 2) for screening nanJ-null mutants (BMC201), primers nanIKOF and nanIKOR (Table 2) for screening nanI-null mutants (BMC202), and primers nanHKOF and nanHKOR (Table 2) for screening nanH-null mutants (BMC203). Each reaction mixture was subjected to the following PCR amplification conditions: cycle 1, 95°C for 5 min; cycles 2 through 35, 95°C for 30 s, 55°C for 40 s, and 68°C for 90 s; and a final extension for 5 min at 68°C. An aliquot (20 µl) of each PCR sample was electrophoresed on a 1.5% agarose gel and then visualized by staining with ethidium bromide. The mutants were cured of the intron-carrying donor plasmid as described [25].
To prepare a nanI/nanJ double null mutant, pJIR750nanJi was electroporated into the BMC202 nanI null mutant strain, which had been cured of the pJIR750nanIi plasmid. Transformants were grown on BHI agar plates containing 15 µg/ml chloramphenicol and putative nanJ/nanI double null mutants were then identified by demonstrating the presence of introns in both the nanI and nanJ ORFs by PCR. The confirmed nanI/nanJ double null mutant (named BMC204, Table 1) was cured the intron-carrying donor plasmid pJIR750nanJi. The identity of this BMC204 nanI/nanJ double null mutant was then confirmed by Southern blotting, which demonstrated the presence of two introns in the mutant.
To prepare a nanI/nanJ/nanH triple null mutant, pJIR750nanHi was electroporated into the BMC205 double nanI/nanJ null mutant strain and transformants were then grown on BHI agar plates containing 15 µg/ml of chloramphenicol. Putative nanH null mutants were screened by PCR, as described above. The presence of introns in the nanI, nanJ and nanH genes of this mutant (named BMC205, Table 1) was confirmed by PCR and Southern blotting, following curing of the intron-carrying donor plasmid pJIR750nanHi.
An intron insertion, with a sense orientation, was targeted between nucleotides 330 and 331 of the etx ORF. The primers used for targeting this intron were 330/331-IBS, 330/331-EBS1d and 330/331-EBS2 (Table 2). The 350-bp PCR products were inserted into pJIR750ai between the HindIII and BsrGI enzyme sites in order to construct an etx-specific TargeTron plasmid. The resultant plasmid, named pJIR750etxi (Table 1), was electroporated into either wild-type CN3718 or BMC205 to inactivate their etx genes, which produced (respectively) etx null mutant strains named BMC206 and BMC207 (Table 1). The primers used for etx null mutant scanning were etxkoF and etxkoR (Table 2).
DNA was isolated from wild-type CN3718, the five sialidase null mutants (including the three single null mutants BMC201, BMC202 and BMC203, the double nanJ/nanI null mutant strain BMC204, and the triple sialidase null mutant strain BMC205) and the two etx null mutants of CN3718 and BMC205 (including the BMC206 null mutant that produces all three sialidases and the BMC207 null mutant unable to produce any sialidase) using the MasterPureTM Gram-Positive DNA Purification Kit (Epicenter). Each DNA was then digested with BsrGI or EcoRI overnight at 37°C and run on a 1% agarose gel. After the alkali transfer to a nylon membrane (Roche), the blot was hybridized with a digoxigenin-labeled, intron-specific probe as described previously [25]. This intron-specific probe was prepared using the primers KO-IBS and KO-EBS1d [25] and a PCR DIG Labeling Kit (Roche Applied Science) according to the manufacturer’s instructions.
DNA was isolated from CN3718 using the Master PureTM Gram-positive DNA purification kit. The primers nanJcomF and nanJcomR (Table 2) were used for nanJ complementing strain construction. The primers nanIcomF and nanIcomR (Table 2) were used for nanI complementing strain construction. The primers nanHcomF and nanHcomR (Table 2) were used for construction of a nanH complementing strain. The PCR reactions were set up as: 1 µl of each pair of primers (at a 0.5 µM final concentration), 1 µl of purified DNA template and 25 µl 2×Taq Long Range Mixture (NEB) were mixed together and ddH2O was added to reach a total volume of 50 µl. The reaction mixtures were then placed in a thermal cycler (Techne) and subjected to the following amplification conditions: 1 cycle of 95°C for 2 min, 35 cycles of 95°C for 30s, 55°C for 40s, and 65°C for 5 min, and a single extension of 65°C for 5 min. The resultant 4.9kb nanJ PCR product, 2.8 kb nanI PCR product, or 2.3 kb nanH PCR product were each separately cloned into the Invitrogen pCR2.1 TOPO vector according to the manufacturer’s instruction and inserts were then sequenced at the University of Pittsburgh Core Sequencing Facility. Using BamHI and SalI, the nanJ insert was removed from the TOPO vector and ligated into the pJIR751 C. perfringens/E. coli shuttle plasmid, forming a plasmid named pJIR751nanJcomp (Table 1). Using the same method, nanI and nanH inserts were separately cloned into pJIR750 and pJIR751 C. perfringens/E. coli shuttle plasmids, forming plasmids named pJIR750nanIcomp and pJIR751nanHcomp, respectively (Table 1). These plasmids were individually introduced, by standard electroporation techniques, into the BMC207 null mutant strain to create nanJ, nanI and nanH complementation strains named, respectively, BMC2071, BMC2072 and BMC2073 (Table 1).
A 0.2 ml aliquot of an overnight FTG culture of the wild-type, null mutants or complementing strains was inoculated into 10 ml of TH medium. To perform ETX and sialidase Western blots, supernatants of 8 h or overnight TH cultures were used. Samples were collected and each supernatant (or the sonicated whole culture) was mixed with SDS loading buffer and boiled for 5 min. Those mixtures were electrophoresed on a 12% polyacrylamide gel containing SDS for analyzing ETX or an 8% polyacrylamide gel containing SDS for analyzing sialidase proteins. The gels were then subjected to Western blotting using appropriate antibodies, as previously described [31].
To assay sialidase enzyme activity, a previously described protocol was used [20] Briefly, 20 µl of TH culture supernatant was added to 60 µl of 100 mM sodium acetate buffer (pH 7.2) in a microtiter tray. A 20 µl of aliquot 4 mM 5-bromo-4-chloro-3-indolyl-α-D-N-acetylneuraminic acid (Sigma) was then added, and the tray was incubated at 37°C for 30 min. The absorbance at 595 nm was then determined using a microplate reader (Bio-Rad).
A 0.1 ml aliquot of an FTG overnight culture of the wild-type strain, null mutants or complementing strains was transferred to TH medium, which was grown at 37°C overnight. A 0.2 ml aliquot of each TH overnight culture was then inoculated into 10 ml of pre-warmed TH medium and the OD600 of those cultures was measured at 37°C over time to determine vegetative growth, as described previously [32].
Madin-Darby Canine Kidney (MDCK) epithelial cells were cultured in a 1∶1 (v/v) mix of Dulbecco’s Modified Eagle’s Medium (DMEM, Sigma) and Nutrient Mixture F12 HAM (Sigma), supplemented with 3% fetal bovine serum (Mediatech), 100 µg/ml penicillin/streptomycin (Sigma) and 1% glutamine (Sigma). Caco-2 cells were maintained in Eagle minimal essential medium (Sigma) supplemented with 10% fetal bovine serum (Mediatech), 1% nonessential amino acids (Sigma), and 100 µg/ml of penicillin/streptomycin. Vero cells were cultured in M199 medium (Sigma) supplemented with 5% fetal bovine serum and 100 µg/ml of penicillin/streptomycin. HT-29 cells were cultured in RPM1-164 medium (Sigma) supplemented with 10% fetal bovine serum and 100 µg/ml of penicillin/streptomycin. NT6 fibroblast cells were culture in Alpha-MEM (Modified Eagle Medium) supplemented with 7.5% fetal bovine serum, 2 mM L-glutamine, 1× Non Essential Amino Acids, 1 mM sodium pyruvate, and 100 µg/ml of penicillin/streptomycin. All cells were grown at 37°C with 5% atmospheric CO2.
Prior to use in ETX cytotoxicity assays, purified ETX prototoxin was activated by trypsin. For this purpose, 50 µl aliquots of labeled or unlabeled prototoxin (2 mg/ml) were incubated with 12.5 µg of trypsin (Sigma)/µg prototoxin for 1 h at 37°C. After that incubation, trypsin inhibitor (Sigma) (1∶1 v/v) was added to remove trypsin activity. The same trypsin/trypsin inhibitor mix (no ETX) was used as a negative control for cytotoxicity assays. Confluent monolayer MDCK cells were incubated in the presence or absence of 5 µg/ml of activated ETX; in some experiments, that same amount of activated ETX was added to a 10 × concentrated TH supernatant culture for 1 hr at 37°C. After this incubation, cytotoxicity was measured using the LDH Cytotoxicity Detection Kit (Roche).
An aliquot of 0.005 U/ml NanI (Roche) or 8 h TH culture supernatants of BMC206, BMC207, BMC2071, BMC2072 or BMC2073 were incubated with 12.5 µg of trypsin (Sigma) or chymotrypsin (Sigma) for 1 h at 37°C. After that incubation with trypsin, trypsin inhibitor (Sigma) (1∶1 v/v) was added to remove trypsin activity from those samples. Sialidase activities were detected by the method described before [20].
Confluent monolayers of MDCK cells grown in 6 well cluster plates (Corning) were or were not pretreated with 0.005 U/ml or 0.001 U/ml of C. perfringens sialidase (Roche) for 30, 60, or 90 min. After this treatment, 5 µg/ml of AF488-ETX was added and the cells were then further incubated for 60 min at 37°C. Following this toxin challenge, the cells were harvested and washed twice with PBS buffer and the pellets were then resuspended in 50 µl of PBS. After treatment of those asample with 1 µl of Benzonase nuclease (Novagen) for 5 min at room temperature, 12 µl of 5×SDS loading buffer was added. These samples were then electrophoresed on a SDS-containing 8% polyacrylamide gel, in the dark. The resultant gel was imaged using a Typhoon 9400 variable mode imager (Amersham Biosciences), with fluorescence emission set to detect the Alexa Fluor 488 label using the green laser with wavelength 532 nm. For detection of the molecular weight markers, the red laser was used with a wavelength of 633 nm. Complex formation was then quantified using Imagequant version 5.2 (Molecular Dynamics).
To evaluate ETX binding, confluent monolayers of MDCK cells grown in 24 well cluster plates (Corning) were pretreated with a 0.005 U/ml or 0.001 U/ml of NanI sialidase at 37°C, and then challenged with 5 µg/ml of AF488-pETX suspended in 250 µl of buffer or culture supernatants (from BMC206, BMC207, BMC2071, BMC2072, BMC2073) that had been concentrated 10× using an Ultrafiltration centrifuge tube (Millipore). After 60 min at 37°C, the cells were harvested and washed twice in PBS. The washed pellets were then resuspended in 50 µl of RIPA buffer (Invitrogen) and added to 96-well plates, where fluorescence was quantified using a Multi-mode Microplate Reader (SynergyTM4, BioTek).
A 0.1 ml aliquot of an FTG overnight culture of BMC206, BMC207, BMC2071 or BMC2072 was transferred to TH medium, which was grown at 37°C overnight. A 0.1 ml aliquot of each TH overnight culture was then inoculated into fresh 10 ml TH medium. These overnight cultures were centrifuged and the pelleted cells were washed three times with HBSS buffer and then resuspended in HBSS buffer at 107/ml. Confluent monolayers of Caco-2 cells, MDCK cells, Vero cells, HT-29 cells, and NT-6 cells grown in 6 well cluster plates were washed three times with HBSS and then challenged with a 1 ml aliquot of the washed bacterial cell suspension. Following anaerobic incubation at 37°C, for 2, 4, or 6 h, the supernatants were removed, centrifuged and 60 µl supernatant aliquot were used to measure sialidase activity by the method described above, except the sample was incubated at room temperature overnight instead of 37°C for 30 min before the absorbance was read in order to increase assay sensitive. Control washed bacterial cell suspensions were treated similarly except for the absence of cells.
As described previously [25], total RNA was extracted from 1 ml of a 2, 4, or 6 h culture of BMC206 that had or had not been in contact with Caco2 cells. All RNA samples were treated with DNAaseI (Promega). The purified RNA was quantified by absorbance at 260 nm and stored in a -80°C freezer.
qRT-PCR reactions were performed on the purified RNA samples using the iScriptTM one-step RT-PCR kit with SYBR green (Bio-Rad). Briefly, 100 ng of each RNA sample was reverse-transcribed to cDNA at 50°C for 10 min. That cDNA was then used as template for PCR reactions with primers 16sF and 16sR (Table 2) targeting the 16sRNA as a housekeeping gene; primers nanIKOF and nanIKOR targeting the nanI gene, primers nanJKOF and nanJKOR targeting the nanJ gene, or primers nanHKOF and nanHKOR targeting the nanH gene (Table 2). All reactions were performed at AB Appled Biosystems Step One plus Real-Time PCR system.
A 1.5 ml aliquot of a TH overnight culture of BMC206, BMC207, BMC2071, BMC2072 or BMC2073 was centrifuged and the bacterial pellet was then washed three times with HBSS buffer. After resuspension of the washed pellet in 1.5 ml of HBSS buffer, these suspensions were serially diluted from 10-2 to 10-7 with sterile water and aliquots were spread onto BHI agar plates. Following an overnight anaerobic incubation at 37°C, the colonies arising on these plates were counted to determine the number of CFU added to the mammalian cell cultures. Monolayers of Caco-2, HT-29, Vero, MDCK or NT6 fibroblasts were then incubated under anaerobic conditions with a 100 µl aliquot of these washed bacteria for 2 h at 37°C. After this incubation, the monolayers were washed with HBSS three times and host cell-associated bacteria were retrieved by lysing the monolayers in distilled water. Aliquots of these suspensions were plated onto BHI agar plates. After overnight anaerobic incubation at 37°C, the colonies arising on the plates were used to calculate the number of adherent CFU.
Caco-2 cells were grown to confluency in an 8-well chamber slide (Fisher). A 1 µl aliquot of a BMC206, BMC207 and BMC2072 overnight TH culture was added to each chamber of the slide, and the slide was then incubated at 37°C for 2 h. Cells were washed three times with HBSS buffer. The slides were fixed in 4% formaldehyde in HBSS for 30 min at room temperature. The fixed cells were incubated with a 1∶500 dilution of anti-C. perfringens rabbit polyclonal antibody (Genway) in HBSS with 10% BSA for 2 h at room temperature. Those cells were then washed three times with HBSS buffer and incubated with 1∶500 Alexa Fluor-488 goat anti-rabbit IgG (Invitrogen) in HBSS with 10% BSA for 1 h at room temperature. After 3 more washes with HBSS buffer, the cells were incubated with 1∶200 VybrantTM Cell-labeling solutions (Molecular Probes) for 30 min at room temperature. Following a final three washes with HBSS buffer, the chambers were removed, and a coverslip was mounted with Fluoro-Gel (Electron Microscopy Sciences). Imaging was performed on a Zeiss Axioskop 40 immunofluorescence microscope.
|
10.1371/journal.pntd.0005436 | Different genotypes of Trypanosoma cruzi produce distinctive placental environment genetic response in chronic experimental infection | Congenital infection of Trypanosoma cruzi allows transmission of this parasite through generations. Despite the problematic that this entails, little is known about the placenta environment genetic response produced against infection. We performed functional genomics by microarray analysis in C57Bl/6J mice comparing placentas from uninfected animals and from animals infected with two different T. cruzi strains: K98, a clone of the non-lethal myotropic CA-I strain (TcI), and VD (TcVI), isolated from a human case of congenital infection. Analysis of networks by GeneMANIA of differentially expressed genes showed that “Secretory Granule” was a pathway down-regulated in both infected groups, whereas “Innate Immune Response” and “Response to Interferon-gamma” were pathways up-regulated in VD infection but not in K98. Applying another approach, the GSEA algorithm that detects small changes in predetermined gene sets, we found that metabolic processes, transcription and macromolecular transport were down-regulated in infected placentas environment and some pathways related to cascade signaling had opposite regulation: over-represented in VD and down-regulated in K98 group. We also have found a stronger tropism to the placental organ by VD strain, by detection of parasite DNA and RNA, suggesting living parasites. Our study is the first one to describe in a murine model the genetic response of placental environment to T. cruzi infection and suggests the development of a strong immune response, parasite genotype-dependent, to the detriment of cellular metabolism, which may contribute to control infection preventing the risk of congenital transmission.
| Congenital transmission of Trypanosoma cruzi, the causative agent of Chagas disease, remains a problem of global public health impact in endemic areas where vectorial and transfusional transmission have been controlled and in non-endemic countries due to migration movements. Little is known about how the parasite´s presence and genetic variability affect placental capacity to protect the fetus. This study explores, for the first time, the effects of placental environment infection by analyzing parasite persistence and gene expression using functional genomics and biological network analyses in mice infected by two strains of T. cruzi, with differential capacity of congenital transmission. The infection with the strain with a stronger placental tropism was associated to a higher degree of up-regulation in genes related to innate immunity and response to interferon-gamma. Our findings suggest that the placental environment exerts a strong immune response in detriment of cellular metabolism modulated by the parasite strain. These findings constitute a significant contribution to better understand the mechanisms causing congenital infection of T. cruzi.
| Maternal–fetal transmission of Trypanosoma cruzi, the etiological agent of Chagas disease, remains a public health problem that allows uncontrolled transmission of parasites from one generation to another, in endemic and non-endemic regions. Certainly, cases of congenital T. cruzi infection have been described in Japan, USA and Europe, especially in Spain [1, 2]. The risk of transmission is higher in the acute phase because of the great number of circulating parasites, but it may occur at any phase of the maternal disease [3]. Congenital infection has been associated with an increased risk of premature delivery, low birth weight, and more premature ruptures of the amniotic membranes, effects that may be related to placenta inflammation [4]. Studies in mice have reported also infertility and fetal growth retardation, associated or not with congenital infection, in both chronic and acute phase [5, 6, 7].
T. cruzi has been classified in six different discrete typing units (DTUs), named TcI to TcVI, according to biological, biochemical and genetic diversity [8]. Each DTU is formed by several parasite strains which are related to each other based on common molecular markers, but different to strains from other DTUs. These molecular markers currently used to define the T. cruzi-DTUs do not focus on the genes responsible for congenital transmission or pathogenicity of the parasite. In addition, except TcIV, the other five DTUs have been identified in human cases of congenital T. cruzi infection [3, 9]. Thus, the different genotypes of T. cruzi and their population characteristics, as parasite pathogenicity, virulence and tissue tropism may play an important role in congenital infection. However, until now it has not been found differences in the distributions of congenital cases and their respective parasite populations [3, 10–13]. Same genotypes are found in mothers and their infected newborns [11, 12, 14], although it was found a natural selection at clonal level in the parasite populations transmitted to the newborns [10, 12–15]. In murine models, previous studies comparing two different strains (K98 and RA, belonging to TcI and TcVI, respectively) showed differences in inflammatory compromise of the genital tract, the outcome of pregnancy and transmission of congenital infection [16]. Therefore, even though available data in humans suggested no association between particular T. cruzi genotypes with congenital infection, these findings cannot be omitted and deserve further study.
Because little is known about the genetic response to the infection by the most important barrier that T. cruzi faces to reach the fetus, the placenta, studies in infected mice might be a suitable model for understanding the potential role of T. cruzi genotypes on pregnancy and congenital infection.
Functional genomics is an important tool to study host-pathogen interactions, since it gives insight into the molecular mechanisms that control the onset of disease. The present study employed a transcriptomic approach combined with biological network analysis to highlight the differences between the responses of murine placenta environment to infection by two different T. cruzi strains, K98 a clone of the non-lethal myotropic CA-I strain, and VD (TcVI), a strain isolated from a human case of congenital infection.
Animal care was in accordance with institutional guidelines of the “Asociación Argentina para la Ciencia y Tecnología de Animales de Laboratorio” (AACyTAL) and the project was approved by the “Comité Institucional para el Cuidado y Uso de Animales de Laboratorio” (CICUAL) of the School of Medicine, University of Buenos Aires, Resol 2426/2015.
Two T. cruzi stocks were used: the myotropic clone K98 (TcI), a subpopulation derived from the CA-I strain, previously described by Mirkin [17] and, the monoclonal strain VD (TcVI) isolated by Risso from a case of congenital Chagas' disease [18]. These two strains have polar phenotypes: VD is more virulent and reaches its peak of parasitemia at 18–25 after infection whereas K98 displays a mild infection and its parasitemia peak is slightly delayed. Moreover, VD shows a slender shape in the bloodstream stage while K98 shows a stumpy shape [18]. DTU identification was done following the methodology reported by Burgos and coworkers [12] and monoclonality was verified by analysis of microsatellite loci as described Valadares and coworkers [19].
C57Bl/6J mice were supplied by the animal facilities of Facultad de Ciencias Exactas y Naturales and were housed at the Department of Microbiology, School of Medicine, University of Buenos Aires, respectively. To obtain chronically infected mice, 4 inbred females between 6 and 8 weeks old were inoculated by the intradermoplantar (idp) route with 500 bloodstream trypomastigotes of K98 or with 50 bloodstream trypomastigotes of VD, because higher inocula of VD are lethal for this mouse strain by the intraperitoneal route [18]. Parasitemia was measured weekly by counting in a Neubauer chamber the number of parasites obtained from the tail vein blood diluted (1/10) in red blood cell lysis solution (Tris–NH4 Cl 0.83% pH 7.2). Chronic phase was considered when no parasites were detected in chamber (around 3 weeks post-infection). At day 30 post-infection two females were housed with 1 uninfected male per cage. To assess the day of fertilization, the vaginal plug was macroscopically searched [16], being the day of vaginal plug appearance considered as day 0.5 of pregnancy. A third group of 4 uninfected females of the same age and weight were used as control.
On day 18.5±1 of pregnancy, dams were euthanized and samples of maternal blood were taken by cardiac puncture and stored in EDTA solution until processed. The fetuses were withdrawn and samples from skeletal muscle were conserved at −80°C, for DNA extraction and qPCR analysis.
Finally, entire placentas from the naturally mated crosses were removed, thoroughly washed with sterile PBS and placed in RNAlater solution (Applied Biosystems, Foster City, CA) until used. As traces of maternal blood and other tissues such as decidua cannot be discarded as part of the samples, we have termed them as placental environment.
We analyzed total RNA from placental environments (the most proximal placentas from uterine horn of each dam were selected) for global gene expression via the Illumina array platform and using the mouse WG-6 v2.0 Expression BeadChip, service provided by Macrogen, Seoul, Korea.
Two approaches were performed, the Over-Representation Analysis and the Gene Set Enrichment Analysis [20]. Statistical significance of the expression data was determined using LPE test and fold change (FC) in which the null hypothesis was that no difference exists among groups. False discovery rate (FDR) was controlled by adjusting p value using Benjamini-Hochberg algorithm. Gene-Enrichment and Functional Annotation analysis for significant probe list was performed using DAVID (http://david.abcc.ncifcrf.gov/home.jsp). The free and open-source gene function prediction service, GeneMANIA (http://www.genemania.org), was used along with the widely used large-scale network visualization and integration tool, Cytoscape [21] to formulate and visualize the resultant integrated gene network. A gene set analysis using the GSEA package Version 2.0 [22, 23] from the Broad Institute (MIT, Cambridge, MA) was also used to analyze the pattern of differential gene expression between the two groups. Gene set permutations were performed 1000 times for each analysis. The normalized enrichment score (NES) was calculated for each gene set. GSEA results with a nominal p < 0.05 were considered significant. Biological process of GO was the pathway database used (http://geneontology.org/).
To validate the results of the microarray, two placentas from each dam (one already used in the microarray assay) were analyzed by means of RT-qPCR, as described below (total sample size = 8 placentas/ group)”. The number of genes analyzed and the selection criteria followed the recommendations previously described [24, 25] in microarray analyses in order to obtain a good correlation between microarray findings and RT- qPCR data: 14 genes exhibiting at least 1.4 fold change and a p-value ≤ 0.0001 (Ccl3, Ccl4, Ccl7, Cd274, Cd3d, Cd8b1, Edn2, Gbp2, Gbp3, Gzmd, Igtp, Irgb10, Irgm1, H2-Aa, H2-Eb1). Cxcl1 and S100a9 were chosen because these genes were among the few ones up-regulated in K98 group and Gbp6 because it is involved in the secretory granule pathway. The expression levels were normalized to Gapdh.Primer sequences are listed in S1 Table. Moreover, in order to test expression at protein level, Western Blot was performed for CXCL1 and CD274 proteins. Briefly, placentas were homogenized in buffer 50 mM Tris-HCl, 150 mM NaCl, pH 7.4 (TBS) in the presence of 0.1% SDS and a protease inhibitor kit (Complete Mini EDTA-free, Roche, Germany). The homogenates were centrifuged at 17,000×g for 10 minutes at 4°C to remove insoluble material. Protein concentration of supernatants was determined by Bradford reagent (Sigma-Aldrich, St. Louis, MO, USA) and equal amount of protein (100 μg per lane) was resolved on a 10% SDS-PAGE. After blotting and blocking of nitrocellulose membranes with TBS, 0.1% Tween 20 (TBS-T) 3% BSA, they were incubated overnight at 4°C with rabbit anti-CXCL1 polyclonal Ab 1:2000 (PAI-2920, Thermo Fischer, Rockford. IL, USA) or rat anti-CD274 monoclonal Ab 1:1000 (10F.9G2, Bio X Cell, West Lebanon, NH, USA) in TBS-T 1% BSA. Blots were then incubated with horseradish peroxidase-conjugated anti-rabbit IgG 1:3000 (Vector Labs, UK) or horseradish peroxidase-conjugated anti-rat IgG 1:3000 (Sigma-Aldrich, St Louis, MO, USA), respectively. Immunoreactive proteins were revealed by enhanced chemiluminescence (Pierce ECL-Plus, Thermo Scientific, Rockford, IL, USA) according to manufacturer’s instructions. The bands were scanned and quantified using ImageJ software (version 1.410).
To reprobe same membranes with rabbit anti-GAPDH monoclonal Ab 1:5000 (14C10, Cell Signalling, Boston, MA, USA) as a loading control, the membranes first were incubated in stripping buffer (200 mM Glycine-HCl pH 2.2, 0.1% SDS, 1% Tween 20) twice for 10 minutes at room temperature while shaking, washed with TBS-T, and then blocked and immunodetected as describe above.
DNA from maternal blood, placental and fetal tissues was extracted using the High Pure PCR Template Preparation kit (Roche Diagnostics Corp., Indiana, USA) following manufacture instructions.
Total RNA was extracted from placentas using the TRIzol reagent (Invitrogen, Carlsbad, CA) according to manufacturer's instructions, treated with RQ1 RNase-Free DNase (Promega, Madison, USA) and stored at −80°C until used. RNA purity and integrity were evaluated by ND-1000 Spectrophotometer (NanoDrop, Wilmington, USA), Agilent 2100 Bioanalyzer (Agilent Technologies, Palo Alto, USA).
Determination of Satellite T. cruzi DNA loads in maternal, placental and fetal samples was performed by means of quantitative Real Time PCR, as previously reported [26] and normalized with murine β-Actin DNA (housekeeping gene) in the same sample (β-ACT Fw: 5´ CGGAACCGCTCATTGCC 3´ and β-ACT Rv: 5´ ACCCACACTGRGCCCATCTA 3´). Results were expressed as relative amplification respect to that of β-Actin DNA fragment.
Detection of Satellite T. cruzi DNA in fetal samples was performed by qualitative Real Time PCR [26] without using a standard curve for quantification.
The presence of viable parasites in placental tissue was evaluated by amplification of 18S T. cruzi RNA. For this purpose, 1 μg of RNA per sample was reversed transcribed using SuperScript II Reverse Transcriptase kit and random primer (Life technologies, Ontario, Canada). Reactions for quantitative reverse transcription PCR (RT-qPCR) were prepared with 0.15 μM forward and reverse primers (18S Fw: 5´ TGGAGATTATGGGGCAGT 3´ and 18S Rv: 5´ GTTCGTCTTGGTGCGGTCTA 3´), 1X FastStart Universal SYBR Green Master (Roche Diagnostics Corp., Indiana, USA) and each of the diluted template cDNAs (1:100 in DNAse free water). RT-qPCR was analyzed on the Applied Biosystems 7500 Real-Time PCR System using following cycling conditions, as recommended by the manufacturer (Applied Biosystem, California, USA): 95°C 10 min; 5 cycles of 95°C for 15 sec and 64°C 1 min, 35 cycles of 95°C for 15 sec and 63°C 1 min; melt from 60 to 95°C rising at 0.2°C per second. Gapdh (glyceraldehyde-3-phosphate dehydrogenase) was used as internal control to normalize mRNA levels using the following primers: GAPDH Fw: 5´ ACTCCCACTCTTCCA 3´ and GAPDH Rv: 5´ TCCACCACCCTGTTG 3´ and the standard curve method was applied.
Restriction fragment length polymorphism (RFLP)-PCR profiling was performed as described by Burgos [12] with 1 μg of purified minicircle amplicons that were digested with 1 U of MspI + RsaI + HinfI restriction enzymes for 4 h at 37°C. The digestion products were visualized after 10% PAGE and SYBR Gold Nucleic Acid Gel Stained (Invitrogen, Eugene, Oregon). Markers of 10 and 25 bp, DNA Step Ladder (Promega, Madison, USA), were included in runs.
Full nested-PCR targeted to sequences flanking microsatellite repeats for the loci TcTAC15, TcATT14, TcGAG10, TcCAA10 and TcTAT20, were carried out as previously described [19]. Each PCR was performed in a total volume of 15 μL using 1 U of Taq Platinum DNA Polymerase (Invitrogen, Carlsbad, CA). The first PCR rounds were carried out using 2 μL of DNA obtained from blood and placental tissue samples, whereas for the second PCR rounds 1 μL of the amplified products obtained in the first PCR round were used as DNA template. Amplifications were performed in a Veriti Thermal Cycler (Applied Biosystems, Foster City, CA). To determinate the allele sizes, 0.5 μL of PCR fluorescent products were analyzed in 6% denaturing polyacrylamide gels of an ALF DNA sequencer (GE Healthcare, Milwaukee, WI) and compared with fluorescent DNA fragments of 75–320 bp using Allelelocator software (GE Healthcare).
Statistical analyzes were performed using GraphPad Prism 6.01 and InfoStat version 2015. (http://www.infostat.com.ar). Mann-Whitney test and Kruskal-Wallis with multiple comparisons were employed for comparison between two groups and among three groups, respectively.
The data discussed in this publication have been deposited in the National Center for BiotechnologyInformation (NCBI) Gene Expression Omnibus (GEO) and are accessible through GEO Series accession number GSE85996 (http://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE85996).
We performed a microarray analysis to screen for genes whose expressions were altered in the placental environment upon infection with K98 or VD strains. Two different approaches were faced in order to determine the pathways involved. The first approach was “Over-Representation Analysis”, intended to detect protein and gene interactions. This analysis statistically evaluates the fraction of genes of a particular pathway among the complete set of genes that show changes in expression. The second approach used was “Gene Set Enrichment Analysis” (GSEA), which detects small changes in predetermined gene sets, in order to identify significant enrichment pathway-level effects. It was used to perform Gene Ontology (GO) analysis with those gene sets derived from the “Biological Process Ontology”.
We aimed to find out whether the differential gene expression detected in the placental environment of infected animals with VD and K98 strains was associated with organ parasite persistence. Indeed, parasite burden in VD infected animals was two times higher (Log (median) = 3.134; Q1 = 3.04 and Q3 = 3.37) than those from K98 infected mice (Log (Median) = 1.554; Q1 = 0.94 and Q3 = 1.93) (p = 0.0002) (Fig 5c). The median logarithmic parasitic loads observed in maternal blood where 1.965 (Q1 = 1.84 and Q3 = 2.00) for K98 and 1.058 (Q1 = 0.82 and Q3 = 1.10) for VD (p = 0.029) (Fig 5a). In K98 group, the parasite load in dams´ blood compared with that detected in placental environment was approximately three times higher, whereas in VD group the increment was 120 times higher than in bloodstream, suggesting placental tropism of VD (p < 0.0001).
In order to elucidate if T. cruzi DNA detected in the placental environment corresponded to living parasites and not to mere DNA, RT-qPCR of T. cruzi 18S RNA gene was designed and performed in these samples (Fig 5d). RT-qPCR results were in agreement with those observed in DNA samples from infected mice; in fact the expression level of T.cruzi 18S RNA in these samples infected with VD were significantly higher (Log (Median) = 1.941, Q1 = 0.78 and Q3 = 2.06) than in K98 group (Log (Median) = 0.275, Q1 = -0.18 and Q3 = 0.49) (p = 0.002), confirming placental tissue preferential site for VD spreading.
Minicircle signatures were characterized in bloodstream and placental environment (Fig 6). They showed identical intra-group profiles with Jaccard’s coefficients (JD) of 0 for K98 and VD in all tested paired samples (n = 8 for K98 and VD), and clearly distinctive for each strain group (inter-group JD = 0.875). Microsatellite loci polymorphism analysis confirmed the monoclonality of the parasite strains and showed identical profiles between paired bloodstream-placental samples (S7 Table).
We analyzed by qualitative Real Time-PCR skeletal tissue samples from fetuses withdrawn from dams infected by both parasite stocks (N = 27 and N = 24 from K98 and VD groups, respectively), obtaining detectable amplification in all samples, though with Ct values between 35 and 40 in all samples, indicative of very low parasitic burden. Furthermore, the fetuses belonging to the eight placentas used in the microarray validation experiments were tested by qPCR (Fig 5b). These samples presented median logarithmic values of parasitic load significantly lower than those obtained in placental tissues and maternal blood of both infection groups (p< 0.001 for both groups and tissue types; Fig 5). These findings suggest contamination with DNA traces from maternal tissues rather than congenital infection.
Chagas disease is a problem of global public health impact and congenital (mother to child) infection is a remaining problem in endemic countries and the most common form of transmission in non-endemic regions. Host response against T. cruzi infection is still being investigated and little is known about how the presence of this parasite affects the placenta that usually represents a barrier for T. cruzi to reach the fetus during pregnancy. Several difficulties arise when studying the association between T. cruzi genotypes and risk of congenital infection and the placental genetic response: small sample volume of neonatal blood, small sample sizes (as a result of reduced transmission rate), low sensitivity of conventional diagnostic and genotyping methods, and deficient follow-up [27]. In the present study, we explored the effects of the infection in the placental environment with two different genotypes of T. cruzi, by analyzing gene expression and parasite persistence in a murine model using microarray analysis. Several genes were affected in their expression levels, up and down, in association to infection and more remarkably in VD group, perhaps as a result of the higher placental tropism displayed by this strain. The present study also combined this transcriptomic approach with biological network analyses to highlight the differences between the responses of murine placental environment to the infection with different T. cruzi genotypes.
Our first approach of Over-Representation Analysis evaluated the fraction of genes in a particular pathway found among DEGs with FC ≥ |1.5|. One of the down-regulated networks shared by both infected groups was the “Secretory Granule”, in which granzymes play a fundamental role. A down-regulation of these proteins has been observed in fetal NK cells of umbilical cords from newborns congenitally infected with T. cruzi as a mechanism of modulating the immune response [28]. These might be one of the multiple strategies to escape or modulate antigen presentation and T-cell-mediated anti-T. cruzi response of the host leading to intracellular parasite persistence during chronic infection. It has been described that the maternal CD8 T-cell response to placental antigens and to pathogen antigens are independent pathways [29], so the up-regulation of several genes that participate in events of antigenic processing and presentation in the placental environment during VD infection, might indicate that it displays all its weapons to defend the fetus from infection. The IFN-γ response, an essential mechanism to control survival and proliferation of intracellular pathogens, was another up-regulated network observed in VD group. Among the up-regulated genes involved in that network, three members of the guanylate binding protein family (GBPs), showed FCs higher than 2. Interestingly, although it is known that T. cruzi and Toxoplasma gondii, both intracellular protozoans and producers of congenital infections, have different mechanisms of cellular invasion, it has been reported that T. gondii produces a recruitment of GBPs in vacuoles and, each GBP has a specific role in resistance, also depending on the virulence of the strains involved [30, 31].
We have performed another approach with a hypothesis of functional class scoring, since great changes in individual genes may have significant effects on pathways but also weaker and coordinated changes in sets of functionally related genes (i.e., pathways) can have significant impact. Applying GSEA and using the dataset of Biological Process of GO, we found that metabolic processes as well as transcription and macromolecular transport were in general down-regulated in the placental environment of infected animals and these data are consistent with fetal growth retardation observed in mice from dams infected by T. cruzi [5, 32].
Mammalian cell invasion by T. cruzi requires cell signaling and the pathways induced may be different depending on the parasite strain [33]. It has been reported that inhibition of protein kinases significantly inhibits the infection of macrophages by T. cruzi [34] and also the pro-kineticin receptors in mammalian cells, which are G protein-coupled receptors [35]. Moreover, it has been described an activation of host protein kinase C by RA strain (TcVI) that favors infection [36]. In this study, we found that T. cruzi infection by different genotypes had a distinct effect on protein kinase cascade as well as on the signaling pathway of G-protein coupled receptor protein, up-regulated by VD and down-regulated by K98 infected samples. So, the negative regulation of these pathways could explain the lower levels of placental infection observed in K98 group. Likewise, positive regulation by VD could explain the found high parasite loads.
Another interesting finding of our analysis was the down-regulation of apoptosis observed in the placental environment of animals infected by both strains. This prevention of apoptosis might be a mechanism used by the parasite to persist all along infection, reducing potential damage, as observed in cardiomyocytes [37].
Different placental tropism of T. cruzi strains has been described [38]. Our results show that VD displays stronger placental tropism than K98 strain, revealed as an increased in the amount of parasitic burden and 18S RNA expression there compared with dams´ blood parasitic loads. It is worth mentioning that VD strain was isolated from a human case of congenital infection [18]. In dams infected with K98, higher levels of bloodstream parasites were observed at around 48 days post-infection compared with mice infected with VD. This is in agreement with findings in C3H/HeN female mice infected with K98 (44 days post-inoculation) that showed higher parasitemias than mice infected with RA [16]. No significant differences in parasite burden were observed for K98 parasites in dams´ blood and the placenta, suggesting poor tropism, in contrast to VD, demonstrated by both qPCR and RT-qPCR, although a higher duplication rate in VD might not be discard. Both K98 and VD stocks were monoclonal, so no differences in genetic constitution between parasite populations were detected in bloodstream and placental environment by means of minicircle signatures and microsatellite loci polymorphism analyses (Fig 6 and S7 Table).
In order to detect congenital infection, we analyzed by qPCR fetuses from dams infected by both parasite strains, obtaining very low parasitic burden in all tested samples from both infection groups, below maternal bloodstream and placental loads (Fig 5a and 5b). This was indicative of DNA contamination from maternal tissues rather than of true congenital infection. In fact, K98 has been demonstrated not to cause congenital transmission in mice [16]. In agreement with our findings, it has been reported that PCR positivity in pups close to delivery may not be reliable [7]. Besides, we have tested by qPCR one month-old pups born to dams infected with the VD stock, with negative results, demonstrating no congenital infection (unpublished results). Altogether, these observations are in agreement with previous studies showing that congenital transmission in mice is a very rare event; changes in placental gene expression driven by T.cruzi infection appear efficient to control the parasite, precluding congenital transmission.
The placenta is a complex environment where mother and fetus coordinately interact and, during an infection, the presence of the pathogen adds complexity by altering the immunoregulatory circuits. To our knowledge, the present study is the first one to describe the genetic response of this environment to T. cruzi infection and reveals host pathways leading to generate a strong immune response, in order to limit the risk of congenital infection, to the detriment of cellular metabolism. Moreover, we have found that this effect may be modulated by the parasite genotype. New studies to deepen insight in this host response, at both the genetic and protein level, are necessary to better understand the mechanisms causing congenital infection of T. cruzi.
|
10.1371/journal.pntd.0001154 | Justified Concern or Exaggerated Fear: The Risk of Anaphylaxis in Percutaneous Treatment of Cystic Echinococcosis—A Systematic Literature Review | Percutaneous treatment (PT) emerged in the mid-1980s as an alternative to surgery for selected cases of abdominal cystic echinococcosis (CE). Despite its efficacy and widespread use, the puncture of echinococcal cysts is still far from being universally accepted. One of the main reasons for this reluctance is the perceived risk of anaphylaxis linked to PTs. To quantify the risk of anaphylactic reactions and lethal anaphylaxis with PT, we systematically searched MEDLINE for publications on PT of CE and reviewed the PT-related complications. After including 124 publications published between 1980 and 2010, we collected a total number of 5943 PT procedures on 5517 hepatic and non-hepatic echinococcal cysts. Overall, two cases of lethal anaphylaxis and 99 reversible anaphylactic reactions were reported. Lethal anaphylaxis occurred in 0.03% of PT procedures, corresponding to 0.04% of treated cysts, while reversible allergic reactions complicated 1.7% of PTs, corresponding to 1.8% of treated echinococcal cysts. Analysis of the literature shows that lethal anaphylaxis related to percutaneous treatment of CE is an extremely rare event and is observed no more frequently than drug-related anaphylactic side effects.
| The risk of anaphylactic shock is the objection most often raised by opponents of percutaneous treatments for cystic echinococcosis, but there are no updated figures on the actual occurrence of anaphylaxis as a complication of this treatment.
To assess the number of lethal and non-lethal anaphylactic reactions following percutaneous aspiration of echinococcal cysts, we systematically reviewed the literature published from 1980–2010.
The analysis of the available literature shows that the risk of severe anaphylactic reactions resulting from percutaneous treatment of echinococcal cysts has been widely exaggerated and the actual risk may be lower than that following administration of certain antibiotics.
Provided adequate stand-by resuscitation measures are available, each time an echinococcal cyst is punctured, fear of anaphylactic shock is no longer justified as an argument to avoid this therapeutic option.
| Human cystic echinococcosis (CE), caused by the larval stage of the cestode Echinococcus granulosus, is a cosmopolitan parasitic zoonosis, affecting mainly the liver (∼70%) and the lung (∼20%) of the human intermediate host. Clinical symptoms depend on the location, number, and size of the cysts. Until anthelminthic chemotherapy became available (mebendazole in the 1970s and albendazole in the early 1980s), surgery was the only treatment choice. The spectrum of therapeutic options was further extended in the mid-1980s when the increasing availability of modern imaging techniques, namely ultrasound, allowed the introduction of image-guided percutaneous treatment (PT) methods.
Over the years, various PTs have been developed, based on the classic PAIR (Puncture of the cyst, Aspiration of the cyst fluid, Injection of a scolicidal agent, and Re-aspiration of the cyst content) procedure [1], [2] with minor variations of the essential steps [3]–[5]. Different catheterization techniques allowing aspiration of the solid content of cysts have also been developed for those cyst stages that are often unresponsive to PAIR [6], [7].
Despite the wide use of PTs in the last two and a half decades, the fear of anaphylactic shock and dissemination due to the spillage of cystic fluid is still quoted by physicians favoring surgery for the treatment of CE [8], [9]. However, anaphylactic reactions in CE occur not only as a side effect of PT, but also of surgical treatment [10]–[17], result of accidental trauma [18]–[20] and even spontaneously [21]–[24].
To our knowledge there are no updated figures on the frequency of anaphylactic reactions, anaphylactic shock or lethal anaphylaxis following PTs of echinococcal cysts. To quantify the risk of allergic reactions and lethal anaphylaxis related to PT of echinococcal cysts we systematically reviewed the published literature.
We performed a PubMed (MEDLINE) search of the literature using the key words “echinococcal cysts”, “hydatid cysts”, “cystic echinococcosis”, “hydatidosis”, “PAIR”, and “percutaneous treatment” and reviewed the available references published between January 1980 and December 2009 for eligible publications (Figure 1).
The inclusion criteria were as follows:
All publications on PT of E. granulosus cysts with information about the number of treated cysts, the number of PT procedures and the occurrence of lethal complications were included. When the original article was not obtainable but the abstract containing the requested information was, the publication was included in the analysis.
In some cases repeated PTs of the same echinococcal cysts were performed during the course of treatment. In these instances, we collected the total number of treated echinococcal cysts and the corresponding total number of PT procedures.
To avoid multiple counting (duplication) of identical procedures and cases, follow-up publications on identical procedures and cases were traced and excluded (references S1).
In human CE, the liver is the organ most frequently affected. Therefore, we divided the collected cases and PT procedures according to the anatomical location into “hepatic cysts” (Table S1) and “extra-hepatic cysts” (Table S2).
When the exact anatomical locations of the cysts were not specified, the data was collected separately (Table 1).
Information about the reported complications was collected accordingly, differentiated into “lethal complications” and “reversible complications” and summarized (Tables 2, 3, 4, 5, 6, 7).
It was impossible to retrospectively grade the severity of the reversible anaphylactic reactions due to the lack of a standardized definition of the events.
If the authors labelled subjective severity levels of the observed anaphylactic reactions (e.g. “severe”, “moderate”, “mild”, “minor”) we collected, summarized and listed them accordingly. In addition to the evaluation and quantification of anaphylactic reactions, we also collected and summarized other PT related complications, to allow a representative overview of all PT relevant complications.
One hundred-twenty-four publications met our inclusion criteria, with a total number of 5943 PT procedures performed for the diagnosis or treatment of 5517 echinococcal cysts. Ninety-two publications that did not meet the inclusion criteria were excluded from the analysis. Four publications were follow-up publications on identical cases or case series and therefore excluded from analysis.
In all but three of the publications included, detailed information about the observed reversible complications were available. In one additional publication, the observed complications were specified but not quantified. These four publications were labeled in the tables accordingly (Tables S1, S3, 5, 6, 7).
For 863 cysts, information concerning the organ location involved was available, but exact number of cysts for each organ was not. For 17 cysts, information about the location was not available. The publications covering these 880 cysts were labeled in the tables accordingly (Table 1).
A detailed analysis of the observed complications in reference to size, stage, and exact location within the affected organs was impossible due to lack of details in the original publications. Overall, five lethal and 777 reversible complications were collected (Tables 2, 6). Of the five lethal complications, three were related to the PT procedure, while two fatalities occurred due to PT “unrelated causes”. Of the three PT related fatalities, two lethal anaphylactic shocks and one fatality “associated with the use of the method” were reported. Unfortunately, detailed information about the two fatalities due to “unrelated causes” [25], [26] and the fatality reported as “associated with the use of the method” [27] were not obtainable.
There were five fatal cases reported in 5943 performed PT procedures. This occurred while treating 5517 echinococcal cysts resulting in an overall fatality rate of 0.08% (5 in 5943) and 0.09% (5 in 5517) respectively (Table 2). The overall fatality rate due to lethal anaphylaxis is 0.03% (2 in 5943) and 0.04% (2 in 5517) respectively (Table 2).
Reversible complications were reported in 345 out of 3440 PT procedures (10%) for the treatment of 3232 liver echinococcal cysts (Table 3), in nine out of 175 PT procedures (5%) for the treatment of 142 extra-hepatic echinococcal cysts (Table 5) and in 423 out of 2328 PT procedures (18%) for the treatment of 2143 echinococcal cysts of unspecified anatomical location (Table 4).
In summary, 777 reversible complications were observed in 5943 PT procedures for the treatment of 5517 echinococcal cysts. Therefore, reversible complications were observed in 13% of all PT procedures, corresponding to 14% of all treated echinococcal cysts (Table 6).
The reversible complications fall into three categories: anaphylactic, potentially anaphylactic, and non-anaphylactic:
In total, 99 reversible anaphylactic reactions were reported in 5943 PT procedures for the treatment of 5517 echinococcal cysts. Therefore, reversible allergic reactions complicated 1.7% of all PT procedures, corresponding to 1.8% of all treated echinococcal cysts (Table 7).
The potentially anaphylactic complications include “fever”, “hypotensive reaction”, “vaso-vagal-reaction”, and “nausea and vomiting”. In total, 128 potentially anaphylactic reactions were reported during 5943 PT procedures (2.1%) for the treatment of 5517 (2.3%) echinococcal cysts (Table 6).
Non-anaphylactic complications – ranging from frequently observed “biliary fistulas” to very rare events such as “active arterial hemorrhage”, “intracystic bleeding” or “gallbladder hemorrhage” – were reported in 550 cases during 5943 PT procedures (9.3%) for the treatment of 5517 (10%) echinococcal cysts.
Allergic reactions and anaphylaxis are IgE-mediated immediate hypersensitivity reactions that occur when antigen-specific IgE, bound to Fc receptors on mast cells and basophils, are cross linked by the antigen, activating the cells to rapidly release a variety of mediators such as histamine, enzymes and lipid mediators [28].
While anaphylactic reactions and allergic symptoms are usually observed in cases of treatment-related rupture of echinococcal cysts, they may also occur spontaneously. The symptoms vary from mild urticaria to anaphylactic shock [29]. The presence of specific IgE in serum of patients is a well known feature of CE with levels varying according to cyst number, location, morphology, disease severity, and cyst stage [30], [31].
Despite 75% of CE patients having detectable levels of specific IgE and histamine release by circulating basophils in response to E. granulosus, antigens can be detected in 100% of patients [32]. Consequently, allergic reactions are rare and unpredictable. So far, the predictive value of IgE titers (or of IgG4 titers, considered “anti-anaphylactic” isotypes) for the development of allergic reactions has not been investigated.
Echinococcus allergens have mainly been studied with the aim of improving the performance of diagnostic tests. Three conserved proteins have been identified (EgEF-1β/δ, EA21 and Eg2HSP70), by screening of an E. granulosus cDNA library with IgE from patients with and without cutaneous allergic manifestations showing significantly different IgE-binding reactivity between groups [33], [34], [35]. Nevertheless, the identification of such reactivity by a patient's IgE as a predictive factor for the development of anaphylaxis has never been investigated. Another appealing, still unexplored possibility, is the use of these allergens for desensitization therapy. The control of CE-related allergic reactions relies on the administration of vasoactive agents (e.g. epinephrine) and corticosteroids. Although a study reported less severe hemodynamic alterations in surgical patients pre-treated with histamine H1 plus H2 receptor blockers [36], the usefulness of any pre-operative treatment for the prevention of anaphylaxis has never been demonstrated.
The pathogenesis of anaphylactic reactions in CE is still unclear but commonly explained by the disruption of the integrity of the cyst wall with spillage and translocation of allergenic cyst contents into the host's circulation. Despite this, rupture of echinococcal cysts does not always or necessarily lead to anaphylactic reactions. In a series of 24 patients with proven rupture of echinococcal cysts (12 patients with liver cysts and 12 patients with lung cysts) only four patients (16.7%) had symptoms or a history of allergic reactions [37]. The same observation has been made during surgery of echinococcal cysts, were apparent spillage of cyst fluid – despite all precautions taken – is reported to occur in 5% to 10% of cases, without this necessarily leading to anaphylaxis [38].
In our review, we found an incidence of three anaphylactic fatalities per 10,000 PT procedures (0.03%). To put this figure in perspective, one may consider other conditions where treatment entails the risks of lethal anaphylaxis:
Overall, we found a frequency of 1.67 reversible anaphylactic reactions per 100 PT procedures of echinococcal cysts (1.67%) (Table 7). The majority of these reversible anaphylactic reactions were allergic skin reactions (urticaria, rash, pruritus), reported in 1.1 per 100 PT procedures (1.1%) (Table 7).
Again, to put these figures in perspective, we consulted the literature on drug-related allergic skin reactions: in a large surveillance program on drug-induced allergic cutaneous reactions – including 15,238 consecutive inpatients – Bigby et al. found antibiotics to be associated with the highest risk. Among the 51 drugs studied, allergic cutaneous reactions were observed in 1.8% to 5% of all treated patients (penicillin G: 1.8%, erythromycin: 2%, semisynthetic penicillins: 2.1%, cephalosporins: 2.1%, ampicillin: 3.3%, trimethoprim-sulfamethoxazole: 3.4%, amoxicillin: 5%). In the same study, 2.2% of patients receiving blood products presented allergic cutaneous reactions [43].
One problem with allergic reactions from the puncture or surgery of echinococcal cysts is that the exact pathophysiological cause and correlation with consecutive symptoms remains unclear. Some peri-interventional complications reported as “fever” (111 cases), “hypotensive reaction” (15 cases), “vaso-vagal-reaction” (1 case), and “nausea and vomiting” (1 case) (Table 6) might represent allergic reactions. If this were to be the case, the risk of reversible allergic reactions might be as high as 3.8 per 100 PT procedures (3.8%). Even though some of these cases might represent anaphylactic reactions, it can be assumed that most of the “fever” events (111 of the 128 potentially anaphylactic reactions) are due to infections, as post-interventional “cavity infections” and “abscesses” account for 60 of the total 550 non-allergic reversible complications (Table 6).
Additionally, the concept of anaphylaxis awaits a stricter definition, as there is no consensus on exactly how to define it along with considerable disagreement about its prevalence, diagnosis and management [44], [45].
The retrospective evaluation of publications on PT related complications is certainly limited by a number of factors such as non-uniform definitions of anaphylactic events, the merging of data from different kind of studies – covering different PT methods in different settings and dealing with a different composition of clinical cases – and the denominator issue. Due to the retrospective nature of our review and because we can only analyze published data, a publication bias can also be at work. It can be argued that severe events (e.g. severe anaphylaxis) might be more likely be be published. But one could counter that events assumed to be common (especially the often quoted PT related anaphylaxis) might not as readily be published. Nevertheless, we consider the analysis of the existing published literature a justified approach as no other source of more accurate data is currently available.
Future work in this area is needed to investigate the pathophysiology of anaphylactic reactions in CE and to prospectively study the potential relationship between clinical variables such as location, number, size, stage of the cyst, and risk of anaphylactic reactions. While large, well-designed clinical trials are needed to develop treatment algorithms stratified by cyst stage and available level of health care resources, the analysis of the available literature shows that the traditional fear of lethal anaphylaxis and allergic reactions in PT of echinococcal cysts has been exaggerated by the critics of PT. Provided adequate stand-by resuscitation measures are available, each time an echinococcal cyst is punctured, fear of anaphylactic shock is no longer justified as an argument to avoid this therapeutic option.
A necessary evolution in the clinical management of CE will be the comparative evaluation of different PT and surgical methods in certain situations.
While surgery legitimately maintains a central role in complicated cysts (rupture, biliary fistulas, compression of vital structures, bacterial superinfection, haemorrhage), cysts at high risk of rupture, or large cysts with many daughter vesicles, that are not suitable for percutaneous treatment approaches, PT has shown to be a safe and effective alternative for many patients with suitable cysts. What is needed now are evidence-based criteria to allocate the patient to the most appropriate treatment option according to the specific situation.
|
10.1371/journal.pntd.0000348 | Leishmanicidal Metabolites from Cochliobolus sp., an Endophytic Fungus Isolated from Piptadenia adiantoides (Fabaceae) | Protozoan parasites belonging to genera Leishmania and Trypanosoma are the etiological agents of severe neglected tropical diseases (NTDs) that cause enormous social and economic impact in many countries of tropical and sub-tropical areas of the world. In our screening program for new drug leads from natural sources, we found that the crude extract of the endophytic fungus Cochliobolus sp. (UFMGCB-555) could kill 90% of the amastigote-like forms of Leishmania amazonensis and inhibit by 100% Ellman's reagent reduction in the trypanothione reductase (TryR) assay, when tested at 20 µg mL−1. UFMGCB-555 was isolated from the plant Piptadenia adiantoides J.F. Macbr (Fabaceae) and identified based on the sequence of the internally transcribed spacer (ITS) regions of its ribosomal DNA. The chromatographic fractionation of the extract was guided by the TryR assay and resulted in the isolation of cochlioquinone A and isocochlioquinone A. Both compounds were active in the assay with L. amazonensis, disclosing EC50 values (effective concentrations required to kill 50% of the parasite) of 1.7 µM (95% confidence interval = 1.6 to 1.9 µM) and 4.1 µM (95% confidence interval = 3.6 to 4.7 µM), respectively. These compounds were not active against three human cancer cell lines (MCF-7, TK-10, and UACC-62), indicating some degree of selectivity towards the parasites. These results suggest that cochlioquinones are attractive lead compounds that deserve further investigation aiming at developing new drugs to treat leishmaniasis. The findings also reinforce the role of endophytic fungi as an important source of compounds with potential to enter the pipeline for drug development against NTDs.
| Protozoans belonging to genera Leishmania and Trypanosoma are single-cell organisms that can infect humans and cause disfiguring lesions and debilitating or fatal diseases, with enormous social and economic impact in many tropical and sub-tropical areas of the world. The drugs currently available to treat the different forms of leishmaniasis and trypanosomiasis were introduced many decades ago and have significant drawbacks, especially in terms of efficacy, length of treatment, route of administration, toxicity, and cost. In our screening program for new natural products with leishmanicidal activity, we found that the crude extract of a fungus living within the plant Piptadenia adiantoides could kill 90% of the amastigote-like forms of Leishmania amazonensis. The bioassay-guided fractionation of the extract resulted in the isolation of cochlioquinone A and isocochlioquinone A, which showed EC50 values (effective concentrations required to kill 50% of the parasite) of 1.7 µM and 4.1 µM, respectively. These compounds were not active against three human cancer cell lines (MCF-7/mammary, TK-10/renal, and UACC-62/melanoma), indicating some degree of selectivity towards the parasites. Our results suggest that cochlioquinones may serve as starting points for developing new drugs to treat leishmaniasis and reinforce the role of endophytic fungi as an important source of natural products with relevant biological activities.
| Protozoan parasites belonging to the genera Leishmania and Trypanosoma (order Kinetoplastida, family Trypanosomatidae) occurs in the tropical and sub-tropical areas of the world, where they cause severe diseases with huge medical, social, and economic impact to millions of people [1]. All diseases caused by these parasites are among the Neglected Tropical Diseases (NTDs) listed by the World Health Organization [1]. Different species of Leishmania affects over 12 million people and puts over 350 million people at risk in 88 countries; Trypanosoma cruzi infects approximately 8 million and puts 100 million at risk in Central and South America, and T. brucei infects 60 million people in 36 sub-Saharan African countries [2]. The drugs currently available to treat the different forms of leishmaniasis and trypanosomiasis were introduced many decades ago and have significant drawbacks, especially in terms of efficacy, length of treatment, route of administration, toxicity, and cost [2]. To complicate the situation, there is no new drug being developed by the major pharmaceutical industries for these diseases [3].
It is well known that plant-associated microorganisms produce a variety of metabolites with novel structures and interesting biological activities [4],[5],[6],[7]. Indeed, some medicinal properties and biological activities initially attributed to plant species were found latter to be due to the secondary metabolites produced by their endophytic microorganisms [8]. With the aim to discover new drug leads for some NTDs, we have been bioprospecting the Brazilian flora and mycota using bioassays which includes the protozoan parasite Leishmania amazonensis and the enzyme trypanothione reductase (TryR). This flavoenzyme is part of a complex enzymatic system present in protozoans of the order Kinetoplastida that help them to survive under oxidative stress [9]. More important, TryR was shown to be essential for the growth and survival of these parasites, and was validated as a drug target for the discovery and design of new leishmanicidal and trypanocidal drugs [9],[10],[11],[12]. In our bioprospecting program, we have prepared hundreds of extracts of endophytic fungi isolated from plants growing in different Brazilian biomes (unpublished results). The isolate UFMGCB-555, obtained from Piptadenia adiantoides J.F. Macbr (Fabaceae), showed strong activity in the assay with TryR and L. amazonensis and was chosen for investigation aiming at its bioactive components. In this report we describe the molecular taxonomic identification of UFMGCB-555 and the isolation and identification of two leishmanicidal compounds from its extract.
The endophytic fungus was isolated from Piptadenia adiantoides J.F. Macbr (Fabaceae) and deposited at Coleção de Microrganismos e Células da Universidade Federal de Minas Gerais under the code UFMGCB-555. The fungus was grown in potato dextrose agar (PDA) medium for 5 days at 28±2°C. Five millimeter diameter plugs of this culture were then inoculated at the center of 160 Petri dishes (90 mm diameter), each containing 20 mL of malt extract agar (MEA) medium (1% glucose, 1% malt extract, 0.1% peptone, and 1.5% agar). The plates were incubated at 28±2°C for 14 days. After this period, a small aliquot of the biomass was used for extraction of DNA and the remaining material extracted with ethyl acetate for the isolation of the fungal secondary metabolites. Five plates which were not inoculated with the fungus were subjected to the same protocol to serve as control of the culture medium.
The DNA was extracted according to the protocol described by de Hoog [13]. The ribosomal DNA internal transcribed spacer domains (rDNA-ITS) were amplified using the primers ITS1 (5′-TCCGTAGG-TGAACCTGCGG-3′) and ITS4 (5′-TCCTCCGCTTATTGATATGC-3′), as previously described [14]. Five sequences were generated using MEGABACE (Amersham Biosciences, USA) which were used to feed PHRED-PHRAP software [15] in order to find the consensus sequence. The sequence thus obtained was compared with those deposited in the GenBank using the software BLASTn [16] to identify the fungus to the genus level. The sequence was deposited in the GenBank (accession number EU684269). Phylogenetic relationships were calculated using the version 4.0 of the software MEGA [17]. The phylogenetic tree was constructed using the neighbor joining algorithm with bootstrap values calculated from 1000 replicate runs. The Kimura 2-parameter model [18] was used to estimate the evolutionary distances.
The fungus culture material from 160 Petri dishes (approximately 3 liters) was transferred to a six-liter Erlenmeyer containing 3.5 L of ethyl acetate and left in contact for 48 at room temperature. After decantation, the organic phase was filtered and the solvent evaporated under vacuum in a rotary-evaporator at 45°C. Residual solvent was eliminated in a vacuum centrifuge at 40°C, and the crude extract thus obtained was stored at −40°C until use. A similar extraction procedure was carried out using the five plates containing medium only to generate a control extract of the medium.
The initial fractionation step involved high-speed counter-current chromatography (HSCCC), using a Pharma-Tech chromatograph model CCC-1000 equipped with three multilayer coils totaling a volume of 850 mL. With the rotor stopped, the coils were filled with the lower aqueous phase of the biphasic mixture of hexane-ethyl acetate-methanol-water (1.2∶0.8∶1∶1). The coils were then rotated at 1000 r.p.m. and the upper phase was pumped at a flow-rate of 5 mL min−1 in tail-to-head direction until hydrodynamic equilibrium was reached, that is, until only the mobile phase was flowing out of the column. Part of the extract (800 mg) was then dissolved in 10 mL of equal parts of the upper and lower phases of the biphasic solvent mixture and this mixture injected into the column. A total of 158 fractions of 15 mL each was collected. They were pooled into 40 groups based on their similarity, as assessed by thin-layer chromatography (TLC) on silica gel plates (Merck). Group 10 was fractionated on reversed-phase high performance liquid chromatography (HPLC) using a semi preparative column (250×20 mm) filled with RP-18 (octadecyl) silica gel with 5 µm average particle size. The separation was run using a gradient of methanol-water from 70 to 100% in 50 min, at a flow rate of 10 mL min−1. The column effluent was monitored with a UV detector set at a wavelength of 220 nm. Two pure compounds, 1 (14 mg) and 2 (2.4 mg) were isolated.
Proton (1H) and carbon (13C) nuclear magnetic resonance (NMR) spectra (Figure S1), Distortionless Enhancement by Polarization Transfer (DEPT), Heteronuclear Multiple-Quantum Coherence (HMQC), and Heteronuclear Multiple Bond Correlation (HMBC) experiments were run on a Bruker DRX 400 spectrometer using the pulse programs provided by the manufacturer. The substances were dissolved in perdeuterated solvents containing 0.1% tetramethylsilane as the internal chemical shift standard. The data obtained from these spectra are summarized in the Tables 2–4. Mass spectra (MS) were acquired on a Thermo Finnigan LCQ-Advantage spectrometer equipped with an electrospray ion (ESI) source. Solutions of the compounds at 200 µg mL−1 in MeOH-H2O (1∶1) were infused at 25 µL min−1, and the positive and negative mass spectra acquired with a m/z range between 50 and 1000 daltons. The cone voltages were optimized for positive and negative ion analysis in the range between 25 and 50 V. The capillary voltages were set at 4.5 kV in positive ion mode and −3.1 kV in negative ion mode. In the MS/MS experiments, the parent ion isolation width was 3.8 daltons and the normalized collision energy was set at 30% for both compounds 1. Fifty scans from 150 to 600 daltons were collected to generate the averaged spectra.
The TryR microtitre plate assay procedure based on in situ Ellman's-reagent-mediated regeneration of trypanothione, described by Hamilton et al. [19], was used during the screening and the bioassay-guided fractionation protocols. The assay was performed in 96-well plates (Costar 9017, Corning, USA) using Hepes buffer (40 mM, pH 7.5) with 1 mM EDTA. Each assay well (250 µL) contained enzyme (1 mU), trypanothione (1 µM) and NADPH (200 µM). The extracts, fractions and pure compounds were added to the above mixture and incubated at 30°C during 30 min. After this period, Ellman's reagent [5,5′-dithiobis(2-nitrobenzoic acid) –DTNB] was added (70 µM) and the absorbance (Abs) measured at the wavelength of 412 nm in the kinetic mode during the time (t) of 10 minutes at every 10 seconds. The slope of the curve ΔAbs/Δt is proportional to DTNB reduction which, in absence of competing reactions, is proportional to the enzyme activity. The inhibition of the coupled system was calculated as the ratio between slope ΔAbs/Δt) of the experimental wells and that of the controls without drug, that is, percent inhibition = [1−slopeexp/slopecontr]×100.
The classical assay based on the measurement of NADPH consumption was performed in 1 mL cuvettes (1 cm light path length) at 27°C using a Beckman DU spectrophotometer. Each cuvette contained 500 µL Hepes buffer (40 mM, pH 7.5),1 mM EDTA, 4 mU enzyme, 200 µM NADPH, 50 µM cochlioquinone A, and 100 µM trypanothione. The enzyme, NADPH and the compound were pre-incubated for 5 minutes at 27°C. The absorbance measurement at the wavelength of 340 nm was started in the kinetic mode (1 reading per second) for about 30 seconds. The sample compartment cover was opened during 3 to 4 seconds, just enough for the quick addition of the substrate. After closing, the absorbance measurement was continued for another 30 seconds. The initial reaction velocity (v0) was calculated by the instrument software using the first 5–20 data points after the substrate addition. These points should fit a strait line with R2 (squared correlation coefficient) greater than 0.99. To estimate the effect of the compound on enzyme activity, the v0 values obtained in the experiments with and without cochlioquinone A were compared. Each experiment was repeated three times.
The effect of the extract and isolated compounds on the survival and growth of the human cancer cell lines UACC-62 (melanoma), MCF-7 (breast), and TK-10 (renal), was determined using a colorimetric method developed at the National Cancer Institute-USA [20],[21]. Briefly, the cells were inoculated in 96-well plates and incubated at 37°C for 24 h in 5% CO2 atmosphere. The solutions of the test samples were added to the culture wells to attain the desired concentrations, and the plates incubated for further 48 h. Trichloroacetic acid was added to each well to precipitate the proteins, which were stained with sulforhodamine B. After washing out the unbound dye, the stained protein was dissolved with 10 mM Tris, and the absorbance measured at the wavelength of 515 nm. Results were calculated using the absorbance measured in the test-wells (T) in comparison with that of the control wells corresponding to the initial cell inoculum (Ti) and cells grown for 48 h without drug (Tf), using the formula: [(T−Ti)/(Tf−Ti)]×100. This formula allows the quantification of both growth inhibition (values between zero and 100) and cell death (values smaller than zero). Each sample was tested in duplicate in two independent experiments.
Promastigotes of L. amazonensis (strain IFLA/BR/196/PH-8) were obtained from lesions of experimentally infected hamsters. The parasites were incubated for 9 days at 26°C in Schneider's medium buffered at pH 7.2. The promastigotes were then stimulated to differentiate into amastigote-like forms by rising the incubation temperature to 32°C and lowering the pH of the medium to 6.0. After 7 days under these conditions 90% of the parasites were in the amastigote-like forms. The parasite concentration was adjusted to 1×108 cells mL−1, and 90 µL added to each well of 96-well plates, followed by 10 µL of the solutions containing the samples and control drug (0.2 µg mL−1 Amphotericin B - Fungisone Bristol-Myers Squibb, Brazil). The plates were incubated at 32°C for 72 h and the number of parasites estimated using the MTT (methyl thiazolyl tetrazolium)-based colorimetric assay [22]. The results were calculated from the measured absorbancies using the formula [1−(Absexp/Abscontr)]×100, which express the percentage of parasite death in relation to the controls without drug. All samples were tested in duplicate and the experiments repeated at least once. Experiments to determine the dose response curves and the EC50 (effective concentration to kill 50% of the parasites) were run as above, using 1∶2 serial dilutions of the test compounds to reach the appropriate concentrations. The experiments were run in duplicate and repeated at least once.
The antimicrobial activity of the samples was evaluated using the following microorganisms: Candida albicans ATCC18804, C. krusei ATCC2159, Staphylococcus aureus ATCC12600, Escherichia coli ATCC25922 and Cryptococcus neoformans ATCC32608. The yeasts were grown on agar Sabouraud (Difco) at 28°C for 24 h, and their inocula were adjusted in saline solution to Mac Farland optical density scale 1 [23] before seeding into plates containing agar Sabouraud. The bacteria were grown in Brain Heart Infusion (BHI, Difco) in 6-mL tubes, and their concentration adjusted to 103 to 104 cells mL−1 before inoculation into plates containing agar BHI. In each plate, five clean filter-paper disks with 6 mm diameter were placed equidistant from each other on the surface of the medium. Solutions of the samples at 10 mg mL−1 were prepared in 1% aqueous dimethyl sulfoxide (1% aq. DMSO). Five microliters of these solutions, corresponding to 50 µg of the sample, were applied to the paper disks, and the plates incubated at 37°C for a period of 24 to 48 h. Aqueous solutions of amphotericin B and chloramphenicol at 10 mg mL-1 were used as positive controls for yeasts and bacteria, respectively. Solvent control consisted of 1% aq. DMSO. The sample was considered active if it caused a growth inhibition halo around the disk to which it was applied.
The software GraphPad Prism version 4.03 was used to calculate the EC50 values using the non-linear curve fitting of two ore more independent experimental datasets to a four-parameter logistic dose-response model. No constraints were applied to the curve fitting calculations.
A small aliquot (1 g) of the fungus culture was used to isolate the fungal DNA for sequencing of the ITS domains. These sequences were used for the taxonomic identification of the fungus and for the elucidation of its phylogenetic relationships (Figure 1). The fungus UFMGCB-555 was then identified to the genus level as Cochliobolus sp. and showed a close phylogenetic relationship with C. melinidis (Genebank access number AF452445), from which its consensus sequence differs by only 2.7% (12 nucleotides, 70% bootstrap value).
The culture material (3 liters) yielded 2 g of the crude extract, representing a yield of approximately 0.6 g L−1 of culture broth. This extract showed activity in different bioassays (Table 1) when tested at 20 µg mL−1, but was completely inactive against C. albicans, C. krusei, S. aureus, E. coli and C. neoformans when tested at 50 µg per disk (data not shown). The extract of the control (sterile) culture medium was not active in these assays.
Approximately 800 mg of the crude extract was subjected to counter-current chromatography in an HSCCC to afford 158 fractions, which were pooled into 40 groups (Figure 2). After testing all groups in the TryR assay, only Group 10 and Groups 24–40 showed some activity. Groups 24–40 were not studied further because they presented low masses and were constituted of complex and instable mixtures of highly polar compounds. Fraction 10 (100 mg) was further fractionated using reversed-phase HPLC to yield compounds 1 (14 mg; 1.4% w/w of the crude extract) and 2 (2.4 mg; 0.24% w/w of the crude extract).
Both compounds strongly inhibited the growth of amastigote-like forms of L. (L.) amazonensis, with EC50 values (effective concentration to kill 50% of the parasites) of 1.7 µM (95% confidence interval = 1.5 to 1.9 µM) and 4.1 µM (95% confidence interval = 3.6 to 4.7 µM), respectively (Figure 3). However, only 1 was active in the DTNB-coupled TryR assay. Furthermore, while the crude extract was toxic to three human cancer cell lines used in this work (MCF-7, TK-10, and UACC-62), neither Group 10 nor compounds 1 and 2 showed activity against these lineages (Table 1). The compounds were also inactive against the five pathogenic microorganisms investigated (data not shown).
The electrospray ionization mass spectra (ESI-MS) of both compounds exhibited quasi-molecular sodiated ion peaks [M+Na]+ with m/z 555 daltons in positive ion mode, and [M-H]− with m/z 531 daltons in negative ion mode, indicating they both have molecular weight of 532 g mol−1 (Figure 4). The peak integration areas in the 1H NMR spectra (Table 2 and Table 3, Figure S1) indicated the presence of 44 hydrogen atoms, while the 13C NMR spectra (Table 4) showed 30 signals for each compound. Edited DEPT sub-spectra showed signals due eight methyl groups, one belonging to an acetyl group, and five methylene carbon atoms for both compounds. Signals due to eight methyne carbon atoms, four of them oxygenated, were observed for compound 1, while seven signals of methyne carbon atoms, three of them oxygenated, were detected for in the spectra of compound 2. Altogether, these data is compatible with the molecular formula C30H44O8, with an index of hydrogen deficiency of 18, corresponding to 9 unsaturations. The analysis of the 1H and 13C spectra allow to infer that both compounds have five double bonds in their structures, with the remaining unsaturations attributed to the presence of four rings. A quinonoid ring in 1 was suggested by the presence of the signals at 181.64 and 188.86 δ, attributed to carbonyl groups, together with four signals between 134.32 and 151.40 δ in the 13C NMR spectrum. A phenol moiety in 2 was indicated by the presence of six signals between 107.0 and 181.5 δ in the 13C NMR spectrum. A signal at 198.5 δ suggested the presence of a ketone carbonyl in 2. The connections between carbon and hydrogen atoms in the structures were established based on the analysis of the two-dimensional NMR experiments (COSY, HMQC and HMBC - Table 2 and Table 3). The spectral data of 1 and 2 are in agreement with those published in the literature for cochlioquinone A and isocochlioquinone, respectively (Figure 5).
Several endophytic fungi were isolated from the P. adiantoides, a plant species selected due to the activity of its extract in a panel of assays used to screen the Brazilian flora for bioactive natural products (unpublished results). Among the fungi isolated from this plant, the isolate UFMGCB-555 showed strong activity in the assays with TryR and L. amazonensis. Using molecular taxonomy techniques, we were able to identify this fungus as Cochliobolus sp. (Pleosporaceae, Ascomycota). This genus comprises approximately 50 species occurring all over the world [24], many of which can parasitize plants and cause considerable agricultural losses [25]. Some species of Bipolaris, the anamorphic state of Cochliobolus, are also the etiologic agent of several human diseases, such as sinusitis, ocular infections, peritonitis, and meningoencephalitis [26],[27]. However, in the present work we looked for the ability of Cochliobolus UFMGCB-555 to produce secondary metabolites with biological or pharmaceutical potential, especially for neglected tropical diseases.
To guide the fractionation process aiming at the isolation of the active compounds of this extract, we decided to use the TryR assay developed by Hamilton et al. [19]. Besides being more economic due to low consumption of the expensive substrate trypanothione, this assay is simpler, safer and faster to perform than the assay with L. amazonensis. Thus, the bioassay with the parasite was used only at the end of the isolation procedure, with the pure compounds. Using this strategy we arrived at a fraction containing two major substances, one active (1) and the other inactive (2) in the TryR assay (Table 1). Both compounds were, however, active against L. amazonensis, showing EC50 values of 1.7 µM and 4.1 µM, respectively (Figure 3).
After detailed analysis of the ESI-MS and NMR spectra, and comparison of our data with those published in the literature [28], compounds 1 and 2 were identified as cochlioquinone A and isocochlioquinone, respectively (Figure 5). However, slightly different interpretation of some NMR signals, as compared with those described by Miyagawa [28], are noteworthy: a) the hydrogen at C-5 in both compounds (Table 2 and Table 3–position 5) couples with the three hydrogen atoms bound to C-27 and with the hydrogen at C-4, thus appearing as a double quartet, with coupling constants J = 7.0 and 5.1 Hz, and not as quintet, as described by Miyagawa; b) the two hydrogen atoms at C-2 in compound 1 (Table 2–position 2) show distinct signals centered at 1.45 δ and 1.08 δ and not one centered at 1.35 δ; c) the same is true for compound 2 (Table 3–position 2), where the corresponding signals are observed at 1.45 and 1.08 δ and not only at 1.25 δ, and finally; d) the two hydrogen atoms at C-20 in compound 2 (Table 3–position 20) resonate at 1.46 δ and 1.74 δ and not only at 1.60 δ as described by those authors.
Cochlioquinones and related compounds are known to occur in fungi from different genera, including Cochliobolus and Bipolaris [29],[30],[31], Helminthosporium [32], Drechslera [33], and Stachybotrys [30]. However, this is the first report on the leishmanicidal activity of compounds belonging to this class of natural products.
It is known that quinonoid compounds with free positions in the quinone ring are susceptible to nucleophilic attack by thiol groups to form stable Michael addition products [34]. In this regard, after 1 was unequivocally identified as cochlioquinone A, its activity in the Ellman's-reagent-mediated TryR assay was questioned due to the presence of a quinonoid ring in its structure. Thus, rather than directly inhibiting the enzyme in the assay, cochlioquinone A could be capturing the reduced substrate resulting in the interruption of the chemical reduction of Ellman's reagent (Figure 6). To confirm this possibility, we performed the assay using the classical protocol based on the measurement of NADPH consumption [35] while using an excess trypanothione (100 µM) in comparison to cochlioquinone A (50 µM). Indeed, under these conditions no inhibition of the enzyme activity could be observed (data not shown).
In view of the above results, we can conclude that cochlioquinone A and isocochlioquinone A are exerting their leishmanicidal effect by hitting targets other than TryR within the parasite. A literature search disclosed the following information: a) cochlioquinone A is a competitive inhibitor of the ivermectin binding site in Caenorhabdites elegans, with an inhibition constant of 30 µM [36]; b) ivermectin is also active in vivo against different species of Leishmania [37],[38]; c) the mode of action of ivermectin in nematodes is related to its high affinity to glutamate-gated chloride channels, causing an increase in the permeability of the cell membrane to chloride ions [39]; d) the TDR database of potential drug targets for NDTs (http://tdrtargets.org) reveals four genes of L. major expressing putative chloride ions transporters (LmjF01.0180, LmjF04.1000, LmjF32.3370, and LmjF33.1060); e) at least three of these genes have orthologs in C. elegans (C07H4.2, R07B7.1), while LmjF01.0180 has an ortholog also in T. cruzi (Tc00.1047053504797.140). Based on these pieces of information it is plausible to speculate that chloride ions transporters may also serve as a target for cochlioquinone A and related compounds in Leishmania and Trypanosoma. This hypothesis needs further experimental evidences to be confirmed or refuted.
Concerning isocochlioquinone A, the literature [40] show that it can inhibit the growth of the malaria-causing protozoan parasite Plasmodium falciparum with IC50 values of 1.4 µg mL−1 for the K1 strain (resistant to chloroquine and pyrimethamine), and 3.3 µg mL−1 for the NF 54 strain (susceptible to standard antimalarials). These values are close to the EC50 shown against L. amazonensis in the present work.
Besides the significant activity against L. amazonensis, our data indicate that 1 and 2 present some degree of selectivity, as they were inactive in the assays with three human cancer cell lines (Table 1) and five pathogenic microorganisms (data not shown) used in this investigation. The low toxicity of 1 and 2 to mammalian cell lines reported in this work is in agreement with recently reported data showing that isocochlioquinone A has only a small effect on HeLa and KB cells [29],[41]. Another study [33] showed that when cochlioquinone A was tested in vitro against different kinases it showed selective activity against diacylglycerol kinase, both in vitro and in whole cell assay employing BW5147 T cell lymphoma lineage. Related compounds, such as cochlioquinone A1, exhibited selective toxicity towards bovine aortic endothelial cell when compared with normal and cancer cell lines [29] leading the authors to suggest that it may serve for developing new therapeutic agents for angiogenesis-related diseases.
The activity in the low micro molecular range towards L. amazonensis and the selectivity of 1 and 2 are reported here for the first time and justify further investigations on compounds of this class to assess their in vitro and in vivo effect on parasites of the genera Leishmania and Trypanosoma. Finally, by disclosing the leishmanicidal activity of two secondary metabolites from an endophytic fungus, the present work reinforces the role of these organisms as an important source of drug lead candidates for the development of new chemotherapeutic agents for NTDs.
|
10.1371/journal.pbio.0050056 | Auditory Short-Term Memory Behaves Like Visual Short-Term Memory | Are the information processing steps that support short-term sensory memory common to all the senses? Systematic, psychophysical comparison requires identical experimental paradigms and comparable stimuli, which can be challenging to obtain across modalities. Participants performed a recognition memory task with auditory and visual stimuli that were comparable in complexity and in their neural representations at early stages of cortical processing. The visual stimuli were static and moving Gaussian-windowed, oriented, sinusoidal gratings (Gabor patches); the auditory stimuli were broadband sounds whose frequency content varied sinusoidally over time (moving ripples). Parallel effects on recognition memory were seen for number of items to be remembered, retention interval, and serial position. Further, regardless of modality, predicting an item's recognizability requires taking account of (1) the probe's similarity to the remembered list items (summed similarity), and (2) the similarity between the items in memory (inter-item homogeneity). A model incorporating both these factors gives a good fit to recognition memory data for auditory as well as visual stimuli. In addition, we present the first demonstration of the orthogonality of summed similarity and inter-item homogeneity effects. These data imply that auditory and visual representations undergo very similar transformations while they are encoded and retrieved from memory.
| Memories are not exact representations of the past. But can we say that all our senses are equally reliable (or unreliable) sources for memory? We performed a series of experiments to test that proposition. Sound and light are processed by different receptors and neural pathways in the brain. Previous comparisons of auditory and visual memory have done little to place on equal footing the stimuli that will be remembered, limiting the ability to truly compare the two processes. However, using current knowledge of how these sensations are represented in the nervous system, we created auditory and visual stimuli of similar complexity and that undergo similar initial processing by the nervous system. We then used these well-matched stimuli to examine memory for studied lists of either auditory or visual items. Using behavioral measures and a computational model for list memory, we show that memory representations are altered similarly for both hearing and vision. We found that auditory and visual memory exhibit striking parallels in terms of how memory is affected by all the parameters we changed in this experiment. These results imply that auditory and visual short-term memory employ similar mechanisms.
| In the past decade, cognitive science has spawned some powerful computational models for both the large-scale and detailed structure of many fundamental phenomena, including categorization and recognition. These models have enjoyed considerable success, particularly in accounting for recognition of simple visual stimuli, such as sinusoidal gratings and chromatic patches [1–3], and more complex visual stimuli, such as realistic synthetic human faces [4]. By exploiting stimuli whose properties can be easily manipulated, but resist consistent verbal rehearsal strategies [5], researchers can formulate and test detailed predictions about visual recognition memory.
To date, this effort has focused on vision, raising the possibility that the properties of recognition memory revealed thus far might be modality specific and therefore of limited generality. There are several prerequisites that must be satisfied before another sensory modality can be addressed in a comparable fashion. First, a suitable task must be found; second, a family of stimuli must be identified that can be parametrically varied along dimensions thought to be encoded in memory. In addition, baseline memory performance must be comparable across modalities, and the effect of early perceptual processing on the stimulus representations must be similar. Failure to satisfy any of these prerequisites would undermine inter-modal comparisons of memory.
We decided to use Sternberg's recognition memory task, which had been used previously with visual stimuli and whose properties were well understood [6]. We then identified a family of auditory stimuli—moving ripple sounds—whose attributes resembled ones that had proven useful in modeling visual recognition memory. These auditory stimuli vary sinusoidally in both time and in frequency content, and are generated by superimposing sets of tones whose intensities are sinusoidally modulated. Until now, these stimuli have been mainly used to characterize the spectro-temporal response fields of neurons in mammalian primary auditory cortex [7–9], but because their spectro-temporal properties resemble those of human speech [7,10], moving ripple stimuli are well suited to probe human speech perception and memory with minimal contamination by semantic properties or by the strong boundaries between existing perceptual categories [11].
This selection of stimuli was influenced by previous attempts to compare auditory and visual memory. Some of those attempts used auditory and visual stimuli that differed substantially in their early sensory processing, but shared semantic representations [12]. For example, Conrad and Hull's classic study compared memory for a list of digits presented either visually or as spoken items [13]. Initial processing differs tremendously for the two types of inputs, indicating that differences in memory may be due to the divergent initial processing. Further, with stimuli like these, once the items have been encoded into verbal form for storage in memory, shared semantic processes may obscure any fundamental differences in memory for the two modalities. Other experiments use stimuli that arguably are free from semantic influences, yet still fail to equate the early stages of processing required by the stimuli [14].
We examined short-term memory for auditory and visual stimuli whose early sensory processing is comparable. Finding comparable stimuli across modalities is difficult, as it may initially seem incontrovertible that the brain operates differently upon auditory and visual inputs. Certainly the initial stages of processing by the modalities' respective receptors differ from one another in many ways. However, the transformations performed by the nervous system on the information generated by the auditory and visual receptors appear to be very similar [7,15]. Starting from each modality's sensory receptors and continuing to the modality's respective processing networks within the cerebral cortex, analogs between hearing and vision have been noted by several researchers [7,9,16,17]. To take a few examples, adjacent sensory receptors in the cochlea of the ear detect neighboring frequencies of sound the same way adjacent sensory receptors in the retina of the eye respond to light from neighboring locations in space. This analogy extends to the retinotopic/tonotopic structure and receptive fields of auditory and visual cortex.
Both moving ripples and Gabor patches vary sinusoidally along the dimensions that primary sensory neurons encode. These stimuli are described in Figure 1. Moving further along the processing hierarchy, it appears that primary auditory cortex responds to moving ripple stimuli analogously to the way primary visual cortex responds to Gabor patches: a few neurons respond robustly to the stimulus, but most are relatively quiet [9]. The sets of stimuli, therefore, are very well matched in terms of early sensory processing. In addition, to decrease reliance upon verbal rehearsal, these unfamiliar stimuli can be varied continuously, and do not support readily available verbal or semantic labels [5]. So we should expect results to be minimally influenced by semantic relationships among stimuli.
Finally, to promote comparability in the difficulty of the memory task with auditory or visual stimuli, we adopted a strategy introduced by Zhou and colleagues [18]. Recognizing that the similarity relationships among visual stimuli strongly influenced recognition memory, those researchers adjusted each participants' memory test stimuli according to that participant's discrimination threshold. Their aim was to minimize individual differences on the memory task. We took the procedure one step further, adjusting stimuli separately within each modality according to each participant's discrimination threshold for that modality. This was meant to equate for both auditory and visual modalities the powerful influence that similarity exerts on memory.
We present the results of two experiments. Experiment 1 assessed several basic properties of recognition memory for ripple stimuli and memory for Gabor patches; Experiment 2 used ripple stimuli to isolate the effects of summed probe-item similarity and inter-item homogeneity. The design of Experiment 2 was meant to orthogonalize these two potential influences on recognition memory, allowing the effects of summed similarity and inter-item homogeneity to be explored independently. A previously proposed model for visual memory, the Noisy Exemplar Model (NEMo) was fit to the data [1]. Because so many trials were required for each case, and because the NEMo has been shown previously to fit data for visual stimuli quite well [1], only auditory stimuli were used in Experiment 2.
Experiment 1 measured short-term recognition memory for moving ripple stimuli and for both moving as well as stationary Gabor patches. We used a variant of Sternberg's recognition task [6,19]. On each trial, one to four stimuli were sequentially presented, followed after some retention interval by a probe. The participants' task was to identify whether the probe matched any of the items presented in the list, pressing a button to indicate their choice. The use of the Sternberg paradigm for auditory stimuli allows comparisons to the many studies that have used the same paradigm with visual stimuli [1,19,20].
Both moving and static visual Gabor patches were tested because although moving Gabor patches change in time similarly to the ripple sounds, their stationary counterparts have been extensively studied in psychophysical examinations of memory [1]. We examined several basic properties of short-term memory for auditory and visual stimuli: the effect of the number of stimuli that must be remembered (list length), the interval over which those stimuli must be remembered (retention interval), and the serial position of the stimulus matching a probe.
Each participant's data from trials of a given list length and retention interval were averaged to obtain a proportion correct for that combination of conditions. These were compared across participants using standard parametric statistics. Proportion correct measures were used rather than, for example, d′ measures because in this case, the assumption that variances associated with target (probe matches a list item) and lure (probe does not match a list item) trials are identical is probably not defensible, as the range of summed probe-item similarities for target trials is much smaller than for lures (as by definition, Target trials always include a stimulus that is identical to the probe, with a similarity equal to 1) [21].
Previous studies have shown that short-term recognition memory for visual stimuli can be understood using the NEMo introduced by Kahana and Sekuler [1]. Experiment 2 directly tested this model's predictions for memory for moving ripple sounds, and compared these results to previous results obtained with visual stimuli. This experiment was crafted so that the key assumptions of the model, effects of inter-item homogeneity and summed similarity, could be explored in a model-free way, while also allowing data to be fit to NEMo for a more quantitative assessment of these effects. The next section explains the logic of the experimental design.
Figures 5 and 6 show that inter-item homogeneity and summed probe-item similarity both affect memory for complex sounds. This result is analogous to that observed for visual stimuli [1,2,4,18]. By fitting the same computational models to these auditory data and visual memory data, we can more sensitively examine whether the cognitive processing undergone by auditory and visual representations are similar.
The data for Experiment 2 were fit to the models described in the Methods section: a three-parameter model that does not take into account inter-item homogeneity, a four-parameter model that adopts values describing perceptual similarity based on participants' performance when list length is 1 (Figure 7), and a five-parameter model including inter-item homogeneity effects, and not assuming that perceptual similarity can be based on participants performance for list length 1.
Table 1 shows the parameter values produced by model fits to the combined data for 12 participants. Fits were made for individual participants as well, and the parameters are similar in the individual participant fits and the fits to the average. The τ and σ parameters showed most variability across participants. Models in which the τ parameter was estimated from an independent dataset in which study lists comprised just one item are indicated. The value for A calculated from data with list length 1 was 0.93 for the data averaged over participants, with individual participant values ranging from 0.88 to 1.02. The value for τ calculated from data with list length 1 ranged from 0.43 to 1.41 across participants. When τ was allowed to vary in the five-parameter model, the value ranged from 0.97 to 3 (the maximum of the allowed range) across participants. The parameter τ and the criterion C have somewhat of a reciprocal relationship mathematically, and so their values depend on one another: as C decreases, τ increases.
Interestingly, the α parameter was not significantly less than 1, indicating that in this experiment, when participants had to remember only two stimuli, both stimuli were remembered equally well. When participants must maintain more stimuli in memory, however, they are more likely to forget stimuli presented earlier in the list, as is shown in Figure 4.
Note that the β parameter remained negative and with a similar value regardless of model. Note that in both the four- and five-parameter models, β ∼ −1. This result is similar to that found by Kahana and Sekuler [1].
Results indicated that models must incorporate inter-item homogeneity in order to fit the data well. The three-parameter model that did not incorporate inter-item homogeneity (as shown in Figure 8A) accounted for only 51% of the variance (r2), and had an Akaike information criterion (AIC) value (see Methods) of 1,010 (higher AIC values indicate worse fit [27]). On the other hand, the four-parameter model accounted for 78% of the variance (r2), and had a considerably lower AIC value of 652. The five-parameter model, allowing τ to vary according to the list length 2 data, accounted for 81% of the variance (r2) and had a slightly higher AIC value of 696, indicating that the addition of this extra parameter does not make the model more generalizable.
The ripple stimuli used here share many similarities with visual grating stimuli. Grating stimuli have long been a fixture of psychophysical experiments because they can be used to explore some properties of vision that are thought to be fundamental: spatial location, luminance, orientation, and spatial frequency. Similarly, the ripple sounds used in the present study can be used to examine some fundamental properties of hearing: frequency spectrum, sound level, and temporal frequency. The experiments presented here make use of the similarities to explore whether the fundamental information processing steps in vision and hearing are similar.
Moving ripple stimuli and visual gratings are processed by the nervous system in analogous ways, and therefore represent an important class of stimuli for comparing memory in the visual and auditory domains. Both auditory and visual cortical receptive fields have characteristic center-surround properties [7,9,15]. Further, edge detection in visual cortex appears to have an analog in auditory cortex [28]. Relatedly, both auditory and visual systems appear to exploit “sparse” coding [29]: when presented with stimuli of the appropriate type, individual cells respond very strongly to one example of the stimulus type and less strongly to other examples. In the visual modality, single primary visual cortical cells show large responses and specific tuning for oriented sine-wave gratings, or Gabor patches [9,30]. In the auditory modality, single primary auditory cortical cells show large responses and specific tuning for moving ripple stimuli [7,9,15].
Thus, early stages of cortical processing seem to treat Gabor patches and moving auditory ripples in an analogous fashion. Although a number of studies have examined recognition memory for Gabor patches [1,3], comparable tests of memory for auditory ripple stimuli have been lacking until now.
Parametrically manipulable stimuli were used in order to explore how memory alters the representation of stimuli. By using an auditory stimulus set for which early processing is similar to the visual gratings used here and in myriad previous studies (e.g., [1,20,30]), comparisons between memory effects in the two modalities can be made. Our results indicate that these auditory stimuli are processed in a way that is quite analogous to visual gratings. In Experiment 1, we directly tested properties of memory between the two modalities, and found little or no difference depending on stimulus type in how memory is affected by list length, retention interval, or serial position. The overall mean proportions correct were larger for the auditory stimuli than the visual stimuli, but the change with each of these variables was similar regardless of the stimulus type. In Experiment 2, we tested the hypothesis that a quantitative model for visual memory, NEMo, would fit the data for auditory memory better than other models. Indeed, NEMo fit the auditory memory data quite well, as shown in Figure 8, and the two major assumptions of the model proved true for auditory stimuli just as they had for visual stimuli: summed probe-item similarity and inter-item homogeneity each contribute to a participant's probability of responding that Yes, an item has been seen before.
In our hands, direct comparison between auditory and visual memory revealed the two to be strikingly similar. The list length manipulation effectively changed the memory load participants had to bear, and has been used in experiments on vision [6] and hearing [11]. The current experiment reveals that the effect of load does not depend on the modality of the stimulus by comparing the stimulus types using the same participants and same experimental paradigm.
The effects of retention interval on recognition memory are also quite similar across stimulus types, as seen in Figure 3. Memory for auditory and visual stimuli decreased only modestly with retention interval. This result is consistent with previous studies of visual memory [20].
The effects of serial position on recognition memory were found to be quite similar across modality, as seen in Figure 4. Although this is consistent with some studies [31], there is an apparent contradiction in the literature: some researchers have found serial position curves of different shapes for auditory and visual experiments [32]. Many such experiments rely on auditory stimuli that are phonological in nature, and others use different experimental paradigms or stimulus types for auditory and visual experiments. A study by Ward and colleagues [31] implies that the auditory versus visual difference seen in other studies can be explained by the differing experimental methods used. When experimental methods are held constant, little or no serial position difference was seen between the two modalities, consistent with our data. Although there was no significant interaction of the effect of serial position with stimulus type, there is a trend toward a larger recency effect for the auditory stimuli than for the visual stimuli (Figure 4). The origin of this recency effect has been debated [33]. One idea put forward by Baddeley and Hitch [33] implies that the recency effect may be due to implicit learning of the items (similar to priming) followed by explicit retrieval of the residual memory.
Many studies using various types of stimuli in free-recall tasks have shown a “primacy effect” in which serial position 1 shows a better proportion correct than serial position 2 [34]. No such primacy effect is evident in our data, as can be seen in Figure 6. The lack of a prominent primacy effect is consistent with some previous experiments using this paradigm [1,35], whereas other experiments using the same paradigm, but different stimuli, have found modest primacy effects [36]. Previous experiments have shown that these effects are sensitive to the delay between the stimulus items and probe [14,37]. The absence of a primacy effect may be due to specifics of timing, the difficulty of rehearsing these stimuli, or an interaction of the stimuli and recognition memory task employed.
Figure 5 shows that summed probe-item similarity correlates very strongly with whether a probe will be judged as new. Because the similarity of the probe to the closest item is identical in each pair, the data imply that participants use information from all stimuli when making a judgment, not just information about the stimulus closest to the probe [1,25]. This gives credence to an exemplar model of memory, rather than a prototype model [25], and is entirely consistent with the results found in the visual domain [1,4,18].
These data indicate strongly that inter-item homogeneity plays a role in memory for sounds. When items in a list are more similar to each other, participants are less likely to say that a probe was a member of the list. This result was robust through direct data comparison (Figure 6) as well as by model fitting, which gave a more sensitive measure of the effect of inter-item homogeneity. As noted earlier, these results are consistent with experiments that examine memory for visual stimuli, including gratings and faces [1,4]. In fact, some older experiments using sound stimuli are consistent with this inter-item homogeneity effect. In one experiment, participants were required to remember a tone stimulus during presentation of distracter tones, and performed much worse when the distracter tones were presented both higher and lower than the remembered stimuli (low homogeneity between the remembered tone and the distracters), as opposed to the case when distracters were presented only higher or only lower than the remembered stimulus (higher overall homogeneity between the remembered tone and distracters) [38]. The similarity across stimulus type implies that the origin of the inter-item homogeneity effect is a process common to both auditory and visual memory.
The strikingly similar patterns of memory observed for the auditory and visual stimuli imply that the information-processing steps involved in memory for each stimulus type are similar. Previous research has shown that sensory-specific cortex is re-activated during memory for a sensation [39,40]. Further, lesions of some auditory-specific cortex results in impairment specifically to auditory memory [41]. The current data imply that the effects of inter-item homogeneity and summed probe-item similarity on memory either arise from non-sensory–specific cortex, or that the mechanisms in each sensory-specific region are very similar.
The data presented here show that memory for visual and auditory stimuli obey many of the same principles. In both modalities, recognition performance changes in similar ways in response to variation in list length, retention interval, and serial position. Further, memory performance depends not only on the summed similarity between a probe and the remembered items, but also on the similarity of remembered items to one another. Memory performance data for both modalities are fit well by the NEMo. These results imply that auditory and visual short-term memory employ similar mechanisms.
Previous studies have examined how auditory and visual items are encoded into memory, implicating some structures in both visual and auditory working memory [42,43]. Behaviorally, visual and auditory stimuli can interfere with each other, indicating some shared processing [44]. On the other hand, some memory information is processed in sensory-specific cortex, indicating that the transformations performed on such information may differ between modalities [39,40,45]. Our data imply that, regardless of whether the processing is performed by the same brain area or not, similar processing is performed on auditory and visual stimuli as they are maintained and retrieved from memory.
For centuries, people have pondered possible parallels between their experiences of light and sound [17]. Belief that the two modalities were parallel probably influenced Sir Isaac Newton's conclusion that the visible spectrum contained seven colors, the same number of tone intervals in a musical octave [46]. (Newton observed: “And possibly colour may be distinguished into its principle degrees, red, orange, yellow, green, blue, indigo and deep violet, on the same ground that sound within an eighth is graduated into tones.” [46]) Today, 300 years after Newton, understanding of the neural signals supporting vision and hearing has advanced sufficiently that we have been able to formulate and test hypotheses about fundamental relationships between the characteristics of short-term memory for each modality.
Moving ripple sounds: moving ripple stimuli varied sinusoidally in both time (with a period w cycles per second [cps]) and frequency content (with a period Ω cycles per octave). The sounds were generated by superimposing sounds at many frequencies whose intensity at any time, and for any frequency (f), was defined by
where g = log(f/f0), t is time, ψ is the phase of the ripple, and D is modulation depth. (D0 represents the baseline intensity, and is set to 1 in the equation to avoid negative intensity values.) f0 is the lowest allowed frequency. In these experiments, the parameter space was simplified by allowing only one parameter (w) to vary. Other parameters took the following fixed values: Ω = 1, D0 = 0.9, f0 = 200 Hz, and ψ was varied randomly between 0 and π/2 for each stimulus. Frequencies ranged over three octaves above f0 , that is, from 200 to 1,600 Hz. Choices for these parameters were made so that a range of stimuli with parameters close to these could be discriminated, as suggested by existing psychophysical data [10,47,48], and in pilot experiments of our own.
Each stimulus contained 20 logarithmically spaced frequencies per octave. Levels for each frequency were identical, but psychophysical loudness varied. However, the same group of frequencies was used for every stimulus, so the time-averaged loudness should be nearly identical for each of the stimuli. Equation 1 describes for each frequency f, a sinusoidal modulation of the level around some mean, at a rate of w cps. This produces a spectral profile that drifts in time, so that different frequencies are at their peaks at different times. Figure 1 illustrates the dynamic spectrum of a moving ripple, with modulation in both time (w, horizontal axis) and frequency content (Ω, vertical axis). For all stimuli, duration was fixed at 1 s. The level of the stimulus was ramped on and off gradually and linearly over 10 ms at the beginning and end of each stimulus. Frequencies at the spectral edges of the stimulus were treated identically to frequencies in the middle of the frequency range. Two examples of auditory stimuli with different w values are given in Audios S1 and S2, and correspond to the stimuli schematized in Figure 1A and 1B.
Visual stimuli: visual stimuli were Gabor patches, created and displayed using Matlab and extensions from the Psychtoolbox [49]. The CRT monitor was calibrated using Eye-One Match hardware and software from GretagMacbeth (http://www.gretagmacbeth.com/index.htm). The Gabor patches' mean luminance matched that of the background; the peak contrast of a Gabor patch was 0.2. Patches were windowed with a two-dimensional Gaussian envelope with a standard deviation of 1.4 degrees. Before windowing, the visual stimuli were generated according to the following equation:
where s represents the luminance of the stimulus at any y (vertical) position and time, t. Note that these stimuli were aligned horizontally and moved only vertically; the luminance did not change with horizontal position. ψ is the phase of the grating, which varied randomly between 0 and π/2 for each stimulus. D is modulation depth. (D0 is the mean luminance, set to a mid-gray level on the monitor.) In these experiments, the parameter space was simplified by allowing only one parameter to vary at a time. In blocks with moving gratings, the wv parameter varied; in blocks with static gratings, the spatial frequency, Ωv, parameter varied. Other parameters took the following fixed values: D0 = 0.9 and f0 = 200 Hz.
All moving gratings had a spatial frequency, Ωv, of 0.72 cycles per degree, and moved with speeds that ranged upward from 1.5 cps (2.1 degrees per second). For static gratings, stimuli did not move (wv = 0), and had spatial frequencies, Ωv, with a minimum of 0.36 cycles per degree. An example of a moving grating is shown in Video S1, and an example of a static grating is shown in Figure 1C. Parameter values were chosen based on pilot experiments and previous data so that a range of stimuli with parameters near these would be discriminable.
Stimuli were tailored to each participant in an initial session, JND thresholds to achieve 70% correct were estimated using the QUEST algorithm [50] as implemented in the Psychtoolbox [49]. Participants were presented with two stimuli sequentially and responded indicating which stimulus was “faster” (in the case of moving ripples or moving gratings) or “thinner” (in the case of stationary gratings). Thresholds for each stimulus type were estimated in separate blocks. These JND values were used to create an array of ten stimuli for each participant, in which each stimulus differed from its nearest neighbor by one JND. All stimuli were chosen from this array, and were thus separated from one another by an integer number of JNDs.
The timing of stimulus presentation during threshold measurements was the same as that used in the later memory tests for a list with a single item. Stimuli were thus individually tailored for each participant, so that the task was of similar difficulty for all participants, and somewhat similar difficulty across modality [18]. The lowest value that each stimulus could take was the same for all participants. Other stimulus values were allowed to vary by participant in order to equate discriminability across participants. In Experiment 1, for the static grating stimulus type, in which the spatial frequency, Ωv, changed, the lowest Ωv value was 0.36 cycles per degree. For the moving grating stimulus type, in which temporal frequency, wv, changed, the lowest wv value was 0.025 cps. For the moving ripple sounds, the lowest possible ripple velocity, w, was 6 cps. In Experiment 2, the lowest ripple velocity, w, was 7 cps.
In order to minimize the possibility that participants could memorize all stimuli, a second, “jittered” set of stimuli was created and then used on half the trials chosen randomly. This list of stimuli started at 0.5 JND above the base value, and increased in units of 1 JND to create a second array of ten stimuli. For data analysis, we do not distinguish between trials on which the two arrays were used.
We experimentally manipulate the physical difference between any two stimuli, here measured in JND. However, the perceptual similarity is traditionally referred to in models that take perception into account. Therefore, when discussing physical stimuli, we refer to their difference (in JND), but later, when discussing fits to models, it is the related perceptual similarity that is relevant.
Participants: participants for all experiments were between the ages of 18 to 30 y, and were recruited from the Brandeis student population. They participated for payment of $8 per session plus a performance-based bonus. Using a MAICO MA39 audiometer, participants' hearing thresholds were measured at 250, 500, 750, 1,000, 2,000, 3,000, 4,000, and 6,000 Hz. Each participant had normal or better hearing, that is, thresholds under 20 dBHL (decibels hearing level) at each frequency.
Fourteen participants participated in seven total sessions each. In an initial session, hearing was tested and vision was tested to be 20/20 or better (using a Snellen eye chart), participants performed 30 practice trials for each stimulus type, and JND thresholds were measured at a 70% accuracy level for each stimulus type. Each of the subsequent six sessions lasted approximately 1 h, and consisted of 504 trials. A session began with 15 practice trials, whose results were not included in subsequent data analysis. For each participant, successive sessions were separated by at least 3 h, and all sessions were completed within 2.5 wk.
Apparatus and sound levels: participants listened to ripple sounds through Sennheiser Pro HD 280 headphones. All stimuli were produced by Apple Macintosh iMac computers and Matlab, using extensions from the Psychtoolbox [49]. Sound levels for this system were measured using a Knowles electronic mannequin for acoustic research, in order to define the stimulus intensity at the participant's eardrum. Levels for all stimuli in Experiment 2 were 79 dBSPL (decibels sound pressure level), well above our participants' hearing thresholds, and levels for stimuli in Experiment 1 were similar (with the same code and hardware settings, but a different computer).
This experiment examined and compared some basic characteristics of short-term memory for moving ripple sounds and for Gabor patches. Using Sternberg's recognition memory paradigm, we examined recognition's dependence on the number of items to be remembered, the interval over which the items had to be retained, and the serial position of the to-be-remembered item [6]. The experiment used static visual gratings (in which the spatial frequency of the gratings, Ωv, varied), moving visual gratings (in which the speed of the gratings, wv, varied), and moving ripple sounds (in which the temporal frequency, w, of the ripples varied).
Stimulus presentation: trials were presented in blocks such that only one stimulus type (moving ripple sounds, static gratings, or moving visual gratings) was presented per block. During presentation of either visual or auditory list stimuli, participants fixated on a “+” in the center of a computer screen. Each stimulus, auditory or visual, lasted for 1 s. After the last item from a list was presented, a short beep sounded, and the “+” was replaced by the text “...”, indicating that the participant should wait for the probe. The text “?” was presented onscreen during presentation of the probe (for sound stimuli only) and after the probe presentation, before the participant made a response. Participants were instructed to be as quick and accurate with responses as possible. Stimuli were presented in blocks of 84 trials of a given stimulus type. Six total blocks were presented per session. The first two trials of each block were not used for analysis to allow for task-switching effects.
Stimuli for each list were chosen from a set created as described above for each participant based on their own JND threshold. Trials with different list lengths and retention intervals were randomly interleaved. Twenty-four trials of each possible serial position were presented to each participant, for each stimulus type. Effect of retention interval was examined by having participants perform trials in which a single stimulus was followed by a probe, after a retention interval of 0.6, 1.9, 3.2, 4.5, or 9.7 s; 24 trials of each retention interval were performed by each participant. Equal numbers of trials in which the probe matched a list item (target), and trials in which the probe did not match (lure) were performed.
Trials were self-paced, with each beginning only when participants indicated with a key press that they were ready. Participants were alerted with a high or low tone whether they got the current trial correct or incorrect, and were updated after each trial as to their percent correct. For every percentage point above 70%, participants received an extra $0.25 reward above their base payment of $56.
Participants and stimulus presentation: on each trial, a list of one or two ripple stimuli (s1, s2) were presented, followed by a probe (p). As in Experiment 1, the participants' task was to identify whether the probe stimulus matched any of the items presented in the list, and press a button to indicate a choice. During list presentation, participants fixated on a “+” in the center of a computer screen. This was replaced by a “?” during the presentation of the probe item. Twelve participants participated in each of eight sessions, following an initial session in which hearing was tested, JND thresholds for the w parameter (cps) were measured, and 200 practice trials were performed. Sessions were approximately 1 h each, and consisted of 586 trials. At the beginning of every session, each participant completed at least 30 practice trials that were excluded from data analysis. Each session began at least 6 h from the previous session, and all sessions were completed within 3 wk. All other details are as described for auditory stimuli in Experiment 1.
Summed probe-item similarity: in order to examine the effect of summed probe-item similarity independently of other confounds, such as the similarity of the probe to the closest item or the inter-item homogeneity, stimulus conditions were created that varied summed probe-item similarity while other factors were held constant. Two pairs of conditions were created that were similar in all respects, but the summed probe-item similarity varied between the two conditions in the pair.
Figure 5A shows the relationships between stimuli for each condition. All figures indicate relationships between stimuli in terms of their differences in units of JND.
Pairs of conditions (labeled a & b on one side, and c & d on the other) were created with identical inter-item homogeneities, and identical similarities between the probe and the item closest to it. However, each pair has one low and one high summed probe-item similarity (pair a & b, for example, both have inter-item difference = 2 JND, but summed probe-item differences of 2 and 4 JND units, respectively).
Figures 5 and 6 indicate only the relationship among the stimuli in units of JND, not their physical values. Part A in these figures illustrates the case when s1 < s2, equally often s1 > s2. Also in conditions b and d, the probe, p, is equally likely to be greater than or less than the stimuli s1 and s2. The conditions as shown in the figures do not specify exactly the stimulus values for a trial. Eight cases of each condition were chosen randomly from all possible configurations that satisfy the condition, given ten stimuli in the array. This made 64 lure cases. Twenty repetitions of each case were performed by each participant, interleaved among the other trial types. For each lure case, analogous target cases were created where the probe matched one of the stimuli. Each target case matched a different lure case in either inter-item homogeneity (in conditions a–d), or summed probe-item similarity (in conditions e–h, explained below).
Inter-item homogeneity: stimulus conditions with high and low inter-item homogeneity were created according to Figure 6A, which follows the same conventions as Figure 5A. Relationships between stimuli for each condition are shown in terms of their physical differences, in units of JND. Two sets of paired high and low homogeneity conditions were created; both members of a pair had the same inter-item homogeneity and similarity between the stimulus and the closest probe.
Computational modeling of results: fitting computational models to experimental data can help determine what information processing steps are involved in short-term memory. Previous experiments in the visual domain found that a NEMo, including effects of summed probe-item similarity as well as inter-item homogeneity, fit data for short-term visual memory well [1,2,4].
The NEMo model was applied only to the data from the 128 auditory memory cases whose list length was two items, because only those trials incorporated information about inter-item homogeneity, important to the model. The NEMo assumes that given a list of L items and a probe item, p, the participant will respond that “Yes, the probe is a member of the list” if the quantity:
exceeds a threshold criterion value, C.
The first term depends on the summed similarity between the probe and the items on the list. α is defined as 1 for the most recent stimulus; its value for a less-recent stimulus determines the degree of forgetting of that stimulus. It should take on values less than 1 if the earlier item is forgotten more readily. η, as defined in Equation 4, measures the perceptual similarity between any two stimuli, as a function of τ, which defines how quickly perceptual similarity drops with physical distance:
The parameter A in Equation 4 defines the maximum similarity between two stimuli. ε defines the noise in the memory representation of the stimulus (hence the label “Noisy Exemplar”). The parameter ε is a normally distributed random variable with variance σ2. Note that the similarities incorporated in the model depend on the noisy values of the remembered stimuli.
The second term in Equation 3 involves the homogeneity of the list, that is, the similarity between the remembered list items. β is a parameter determining the direction and amplitude of the effect of list homogeneity. If β < 0, as was found in earlier experiments using visual stimuli, a given lure will be more tempting when s1 and s2 are widely separated; conversely, if β > 0, a lure will be less tempting when s1 and s2 are widely separated. If β = 0, the model does not depend on inter-item homogeneity, and is a close variant of Nosofsky's Generalized Context Model [24]. The parameter A, as defined in Equation 4, was set to 1. This model allows five parameters, σ, α, β, C, and τ, to vary.
Two additional similar models were also examined. A second model assumes that the similarity between items can be predicted from participants' probability of confusing two items in a trial of list length 1. This model adopts values for τ and A for each participant based on the fit of Equation 4 to their data with list length 1. This model is identical to that above, but simpler, allowing only four parameters (σ, α, β, and C) to vary based on the data with list length 2.
A third model is identical to the second, but in it, β is forced to be 0, which means that only three parameters are free to vary: σ, α, and C. Note that this last model does not take into account any possible influence of inter-item homogeneity. Models are labeled according to the number of parameters varied in each: five, four, and three.
Model fits: models were fit to participants' accuracy data by means of a genetic algorithm. Such a method was chosen because it is robust to the presence of local minima [51]. The parameter spaces involved in this experiment are relatively complex, so the genetic algorithm approach was particularly attractive. To summarize our implementation of a genetic algorithm, 3,000 “individuals” were generated, each a vector of randomly chosen values for each of model's parameters. The ranges for each parameter were: 0 < σ < 5, −3 < τ < 3, 0 < α <1, −2 < β < 2, and 0 < C < 2. Three thousand trials were simulated for each individual, each with a randomly chosen value for ε given the parameter σ. When the value in Expression 3 exceeds C, the simulation produced a Yes response. The proportion of Yes responses for each case was calculated. The fitness of each individual was computed by calculating the log likelihood that the predicted and observed data came from the same distribution. Log likelihood was chosen because it is more robust to non-normal data than is a least-squares error method [27]. The 10% most fit individuals are maintained to the next generation. These act as “parents” to the next generation: the parameters for the 3,000 individuals of the next generation come from combinations of pairs of parents and mutations. This procedure was repeated for 25 generations. Best-fit parameters typically did not change past the 20th generation, indicating stable parameter values had been obtained.
Model comparison: in order to compare the three models described above, the predicted data and observed data were plotted against each other, and a measurement of the variance accounted for by the model, r2, was calculated. However, when comparing two models with different complexities, for example, with different numbers of parameters, the important distinction between models is their generalizability to new data, that is, the likelihood that the model will fit another set of similar data. The AIC is a measure of model fitness that takes into account both how well the data fit the model and the number of parameters in the model. See the work of Myung et al. [27] for more information about AIC and calculation techniques. Thus, both the AIC and r2 values were used to discriminate between different models.
|
10.1371/journal.pcbi.1000623 | A Stochastic Model for Microtubule Motors Describes the In Vivo Cytoplasmic Transport of Human Adenovirus | Cytoplasmic transport of organelles, nucleic acids and proteins on microtubules is usually bidirectional with dynein and kinesin motors mediating the delivery of cargoes in the cytoplasm. Here we combine live cell microscopy, single virus tracking and trajectory segmentation to systematically identify the parameters of a stochastic computational model of cargo transport by molecular motors on microtubules. The model parameters are identified using an evolutionary optimization algorithm to minimize the Kullback-Leibler divergence between the in silico and the in vivo run length and velocity distributions of the viruses on microtubules. The present stochastic model suggests that bidirectional transport of human adenoviruses can be explained without explicit motor coordination. The model enables the prediction of the number of motors active on the viral cargo during microtubule-dependent motions as well as the number of motor binding sites, with the protein hexon as the binding site for the motors.
| Molecular motors, due to their transportation function, are essential to the cell, but they are often hijacked by viruses to reach their replication site. Imaging of virus trajectories provides information about the patterns of virus transport in the cytoplasm, leading to improved understanding of the underlying mechanisms. In turn improved understanding may suggest actions that can be taken to interfere with the transport of pathogens in the cell. In this work we use in vivo imaging of virus trajectories to develop a computational model of virus transport in the cell. The model parameters are identified by an optimization procedure to minimize the discrepancy between in vivo and in silico trajectories. The model explains the in vivo trajectories as the result of a stochastic interaction between motors. Furthermore it enables predictions on the number of motors and binding sites on pathogens, quantities that are difficult to obtain experimentally. Beyond the understanding of mechanisms involved in pathogen transport, the present paper introduces a systematic parameter identification algorithm for stochastic models using in vivo imaging. The discrete and noisy characteristics of biological systems have led to increased attention in stochastic models and this work provides a methodology for their systematic development.
| The function of eukaryotic cells relies on the transport of macromolecules and organelles throughout the cytoplasm. Pathogenic viruses can exploit a cell's cytoplasmic transport mechanisms [1],[2] in order to reach their site of replication. Cytoplasmic transport involves three types of molecular motors. Kinesin and dynein motors use microtubule tracks to move cargo throughout the cytoplasm, while myosin motors interact with actin filaments to move their cargoes [3],[4]. Microtubule based transport is usually bidirectional and its mechanism can be explained by the exclusive binding of dynein and kinesin motors to the cargo, motor cooperation and regulation, or a stochastic tug-of-war [5]–[8]. Exclusive binding of motors has not been reported in cells, while in systems with cooperating motors, additional factors such as on/off switches or coordinators between motors have been postulated for bidirectional transport of large cargo, such as vesicles [7]. The mechanism of bidirectional motor transport by non-coordinated motors of opposite polarity has been the basis of tug-of-war models [7],[9].
In this work we propose a stochastic model for motor transport on microtubules and we systematically identify its parameters using virus trajectories obtained by in vivo imaging (Fig. 1). Trajectories are obtained by live cell microscopy of fluorescently labelled human adenovirus type 2 (Ad2) using confocal microscopy. Motility information extracted through single virus tracking [10], and trajectory segmentation [11] are implemented in order to study the properties of virus transport by employing a systems identification process [12] for a stochastic model of cargo transport on microtubules.
The small number of motor proteins involved in microtubule transport implies a system where the fluctuations in the behavior of motors and the randomness of molecular reactions are essential characteristics [13] suggesting a stochastic modeling of the governing processes. Here we propose a stochastic representation of the main events involved in motor transport, namely stepping along microtubules and binding and unbinding of molecular motors to the cargo.
The proposed model has six parameters, namely the binding, unbinding and stepping rates of plus-end and minus-end motors (herein presumed to be dynein and kinesin, respectively). The step sizes of the motors were set to −8/+8 nm for dynein/kinesin as suggested by the results of single molecule experiments [14],[15]. We note that we do not impose any geometrical information on the motors and their binding sites on the virus capsid. The motor protein binding sites on the adenovirus capsid are not known even though a recent cryo-EM image of the structure of the human adenovirus type 2 temperature sensitive mutant revealed the organization of the surface of the virus capsid [16].
The six model parameters are inferred through a system identification process using the velocity and displacement distributions of segmented trajectories as the cost function of our optimization. An evolutionary algorithm, capable of handling noisy cost functions, is used to obtain the rates that minimize the distance between the velocity and displacement distributions of the in silico and in vivo trajectories.
The velocity distribution in virus trajectories has led to several suggestions regarding the cooperation or lack thereof between molecular motors. High velocities, in the order of a few microns per second, were observed for intracellular viruses (Fig. 2E) [17]. Similar high speeds have been observed for vesicles moving along microtubules such as peroxisomes [18] and endosomes [19]. These velocities are above the maximum velocities measured for single motors without load (3 µm/s for dynein, [14]; 0.4 µm/s for kinesin-1, [20]; 3 µm/s for kinesin-1, [21]; 0.8 µm/s for kinesin-1, [22]; 0.8 µm/s kinesin-1 and 0.5 µm/s kinesin-2, [23] in in vitro experiments. It has also been reported for drosophila lipid droplets, that multiple processive motors do not move cargoes faster [24]. It is likely that yet unknown mechanisms account for the high velocities measured in in vivo biological systems. These mechanisms may involve motors which are able to increase their velocities up to few microns per second or motors are able to act additively to achieve higher speeds. Both assumptions have not been experimentally validated or discarded in in vivo experiments. Additive behaviour of motors is an underlying assumption in our model (Fig. 2A). The additive behaviour is inherent to the Stochastic Simulation Algorithm [25] used herein to simulate the model, since the time step to the next event depends on the total propensity (numbers and event rates). The proposed stochastic model does not impose any explicit coordination between motor proteins, e.g. a switching mechanism that selects a certain motor protein type to be active.
We emphasize that our model does not aim at a mechanistic description at the motor level. Forces are known to affect motor properties, but it is not clear how they are distributed among multiple motors [26]. Furthermore while it is possible to obtain data relating forces for certain motors in vitro, there is no such data for in vivo experiments. In the present model the forces between molecular motors and cargo are implicitly taken into account through the binding/unbinding/stepping rates of the stochastic models.
The simulation of the stochastic model produces cargo trajectories that depend on the parameter settings. The model contains no a-priori assumptions on the existence of either a tug-of-war or coordination between molecular motors. In turn, the model parameters are systematically identified with a derandomized evolution strategy that minimizes the difference between the length and velocity distributions of directed motions (fast microtubule dependent runs [11]) produced by the model and those of fluorescently labelled human adenovirus type 2 (Ad2) as measured by confocal microscopy at 25 Hz temporal resolution. The two-dimensional virus trajectories are extracted by a single particle tracking algorithm [10] (Fig. 1A, B). Directed motions along microtubules are classified by trajectory segmentation [11] and the distance travelled along the microtubule is determined as a function of time (1D trajectory shown in Fig. 1C). The same analysis is applied to trajectories obtained by the simulation of our model using the Stochastic Simulation Algorithm (SSA) [27]. These trajectories are also subsequently segmented to classify directed motions [11]. In turn an optimization algorithm is employed to identify the parameters of the stochastic model [28].
Here the six model parameters (binding, unbinding and stepping for both kinesin and dynein, Fig. 2A) were identified by minimizing the Symmetric Kullback-Leibler divergence between the in silico and in vivo length and velocity distributions using an Evolution Strategy with Covariance Matrix Adaptation (CMA-ES) [29] (Fig. 2B,C). The proposed de-randomized optimization algorithm is an essential aspect of our method. CMA-ES samples the six-dimensional multivariate normal distribution involving the parameters of this problem at each iteration and it encodes relations between the parameters of the model and the objective that is being optimized without requiring explicitly the gradients of the cost function [29]. The CMA-ES is a method capable of optimizing noisy cost functions (such as those from the present stochastic model) and its efficiency, reliability and robustness were demonstrated over a number of benchmark problems [30],[31]. The CMA-ES is particularly suitable to this optimization problem as it is know to perform best [29] in problems that are low dimensional (here six parameters), inherently noisy (here a stochastic model), multimodal and computationally expensive (for each parameter set thousands of trajectories are generated and segmented to collect reliable statistics). The algorithm identifies an optimal set of parameters and at the same time provides a sensitivity analysis of the model. The standard deviations of the six principal axes are shown to converge (Fig. 3), thus yielding a minimum (Text S1).
After the convergence of the optimization process (Fig. 3) we found that the directed motion length and velocity distributions of the in silico trajectories, under the optimal set of parameters, matched with high accuracy the experimental data (Fig. 2D, E).
The maximum number of motors attached to the viral cargo is limited by the number of binding sites on the virus. The present model enables predictions on the number of motor binding sites on the viral capsid, a quantity that is difficult to determine experimentally but important for understanding the mechanisms of transport. We first estimated the number (between 2 and 20) of motor binding sites on the virus by an optimization procedure (Fig. 2F). In models with 6–16 binding sites, the cost function values were almost constant around the minimum value obtained for 14 binding sites (Text S1). For less than 6 motor binding sites, the optimization process did not converge to the experimentally observed directed motion length and velocity distributions. Above 16 binding sites, an unbalanced configuration of motors was feasible only at low binding and unbinding rates, and yielded largely unidirectional trajectories due to infrequent motor binding to the virus. We concluded that 14 common binding sites for dynein and kinesin correspond optimally to the experimental data.
Since it is not possible to differentiate between common and separate binding sites, we additionally investigated the possibility that the experimental data support separate binding sites for the different motors. We optimized a model where dynein and kinesin have distinct binding sites, namely 4+4, 5+5, 6+6, 7+7 binding sites, and various permutations thereof (Fig. 2F, Text S1), and found that an equal number of motor binding sites was optimal in all cases. This is consistent with the observation that center and periphery directed length and velocity distributions were almost symmetric (Fig. 2D, E). We note that the optimal number of binding sites, i.e., 14, is the same for the models with common and separate binding sites (Fig. 2F, black curve).
Molecular motors carrying cargo on microtubules operate as individuals or as an ensemble. We found that, on average, during virus directed motions, 1.56±0.56 dynein or kinesin (for minus-end and plus-end directed motions, respectively) motors, and 0.15±0.22 motors of opposite polarity were bound to the virus (Fig. 4A). The probability of binding more than four motors to one virus particle was below 10−3, and most often only one type of motor was bound (Fig. 4A, B). These data are in agreement with low number of motors estimated on vesicular cargo in squid axoplasm by cryo-EM [7]. For other organelles, the estimates for motor numbers range from a few to dozens based on immunological detections in chemically fixed cells.
In order to quantify the correlation between the number of bound motors and the directed motion length, the Sample Pearson Product Moment correlation coefficient (with a range of 0 to 1, where 1 is maximal correlation) between motor numbers and directed motion length was computed to be 0.51 for dynein and 0.49 for kinesin for minus-end and plus-end directed motions, respectively. This implies a weak correlation between the number of bound motors and the directed motion length, showing that long runs do not necessarily require many motors, as two or three already account for lengths in the order of micrometers (Fig. 4B). This result is consistent with the recent in vitro observation that two motors are sufficient to move a cargo over several micrometers [32].
The velocities, derived from optimized stepping rates, for single dynein and kinesin motors were 866 nm/s and 833 nm/s, respectively, consistent with values reported for dynein and conventional kinesin-1 or kinesin-2 [21]–[23],[33]. Although kinesins are currently not known to be involved in cytoplasmic transport of adenovirus [1], the model makes a clear prediction for a plus end directed motor in cytoplasmic transport of adenovirus.
Our findings indicate that microtubule-based motility of adenovirus requires a low number of bound motors compared to the number of binding sites on the capsid. This allows configurations where only one motor type is bound, and thereby produce directed motions. Low numbers of motors allow fast switches between directions and therefore, bidirectional motion. Importantly, the binding and unbinding rates were much smaller than the stepping rates, which is key for directed motion runs (Fig. 2C). Small perturbations of binding and unbinding rates greatly affect the model dynamics (Text S1). The susceptibility of motor based cargo transport to these parameters has been reported in other theoretical studies [26] and hints to a possible mechanism to regulate the run length of the motors [32].
The present results enabled an assessment on the virus binding sites for motor proteins. The outer surface of adenovirus particles is composed of five polypeptides, three of which are still present on cytosolic viruses that have undergone stepwise disassembly [34]. Cytosolic particles contain the major protein hexon, a large fraction of the pentameric penton base at the icosahedral vertex, and protein IX (pIX), which stabilizes hexon. By considering the size (90 nm in diameter) and icosahedral geometry of the virus and the cylindrical microtubule (25 nm in diameter), we can postulate that the maximum number of microtubule motor-capsid interactions occurs along the edge of a capsid facet, in this case on hexon (Fig. 5A, B). This arrangement implies that 9 hexon trimers are aligned with the microtubule, giving a maximum of 27 hexon binding sites for the motors. This is above the value of 14 binding sites predicted from the simulations. If we assume, however, that the motor protein binding sites are located at the interface of two trimeric hexons, one microtubule filament could cover 1–15 sites (Fig. 5B, red lines), which is within the predicted range of 6–16. In addition to hexon, 6 to 8 binding sites were available for pIX, and less than 5 for penton base which detaches to a significant extent from the incoming virions before reaching the cytosol [34]. We analyzed trajectories of pIX-deficient adenoviruses to distinguish between hexon and pIX binding sites for motor proteins [35]. The directed motion length and velocity distributions of pIX-deleted adenovirus were similar to those from wild-type viruses without significant deviations or asymmetries, indicating that pIX may not provide a binding site for microtubule dependent motors during cytoplasmic transport (Fig. 5C, D). Therefore, we predict that hexon harbours the binding sites for dynein and kinesin motors.
In this study, we use in vivo imaging to identify a stochastic model of cargo transport by molecular motors on microtubules. The model parameters were systematically identified using live imaging data of virus trajectories and a de-randomized optimization algorithm to minimize the Kullback-Leibler divergence between the length and velocity distributions of adenovirus directed motions on microtubules with the in silico trajectories produced by the model. The model accounts for directed motions at µm/s speeds, processive stepping over hundreds of nanometres, and periods of stationary behaviour. The results show that the stochastic model can result in bidirectional support without an explicit coordination mechanism.
In our work kinetic rates of a stochastic model are determined via an evolutionary optimization approach using experimental data. The identified model enables a number of predictions. First, it shows that one to four motors are active on virus particles during microtubule-dependent motions, although the number of motor binding sites is estimated to be 6–16. The observation that the cost function value is constant within this range suggests that the virus may align with the microtubule in different orientations (Fig. 5B) and still preserve its motility. This range is consistent with the maximum of 15 hexon trimer-trimer interfaces along the edge of a capsid facet. The low number of motors involved in directed motions supports an emerging concept from wet lab experiments and in silico simulations, that key events of cell functions are in many cases executed by only a few polypeptides [36].
Second, if equal numbers of opposite motors are attached, the cargo oscillates and eventually stops, or remains confined to small areas. This may be an important mechanism for fine-tuning the subcellular velocity to achieve localized delivery of the cargo. We anticipate that viral transport is tuned by the binding and unbinding rates of motors to microtubules or the cargo, rather than by additional regulatory factors. Such tuning could be cell-type specific [17], and could control the number of engaged motors and motor configuration, and also provide specific segregation or guidance cues for traffic. In support of this, it has been suggested that the microtubule binding protein Tau can fine-tune the distance that the cargo travels by reducing microtubule binding of kinesin in distal parts of neuronal axons [7],[37]. In addition, motor properties can be tuned by post-translational modifications, such as phosphorylation of dynein or kinesin binding partners, which could affect their enzymatic functions and hence their stepping rates [7].
We close by noting that besides the results on motor transport on microtubules, the algorithm taken here is in line with reverse engineering and systems identification approaches [28], [38]–[40] which are gaining significance as discovery and model validation tools in systems biology. The CMA-ES algorithm is capable of handling noisy and multimodal cost functions that are inherently associated with stochastic models. The CMA-ES optimization algorithm along with image analysis of in vivo systems can be a robust process to help identify parameters of stochastic models employed in several areas of systems biology.
HeLa cells were grown to 30% confluency on 18 mm glass cover slips (Menzel Glaser) and kept in Hanks buffered salt solution containing 0.5% BSA (Sigma) and 1 mg/ml ascorbic acid (Sigma). Adenovirus serotype 2 and protein IX deficient adenoviruses were grown, isolated, and labeled with atto565 (Atto-tec, Germany) as described by Nakano and Greber in [41] and Suomalainen et al. in [17].
HeLa cells were infected with Ad-atto565 and imaged between 30 and 90 minutes post infection at 25 Hz. Flat regions of the cell were chosen for imaging in order to minimize the cytoplasmic volume above the imaging plane. The center of the cell was detected by differential interference contrast imaging to assign directionality to the virus motions. Images were recorded using a spinning disc confocal microscope (Olympus IX81) fitted with an UplanApo100x objective of N.A. 1.35 on a back-illuminated monochrome Cascade 512 EM-CCD camera (Photometrics) containing a 512×512 pixel chip (with 16×16 micrometer large pixels).
For the computational methods see Text S1.
|
10.1371/journal.pcbi.1002895 | Activity in Inferior Parietal and Medial Prefrontal Cortex Signals the Accumulation of Evidence in a Probability Learning Task | In an uncertain environment, probabilities are key to predicting future events and making adaptive choices. However, little is known about how humans learn such probabilities and where and how they are encoded in the brain, especially when they concern more than two outcomes. During functional magnetic resonance imaging (fMRI), young adults learned the probabilities of uncertain stimuli through repetitive sampling. Stimuli represented payoffs and participants had to predict their occurrence to maximize their earnings. Choices indicated loss and risk aversion but unbiased estimation of probabilities. BOLD response in medial prefrontal cortex and angular gyri increased linearly with the probability of the currently observed stimulus, untainted by its value. Connectivity analyses during rest and task revealed that these regions belonged to the default mode network. The activation of past outcomes in memory is evoked as a possible mechanism to explain the engagement of the default mode network in probability learning. A BOLD response relating to value was detected only at decision time, mainly in striatum. It is concluded that activity in inferior parietal and medial prefrontal cortex reflects the amount of evidence accumulated in favor of competing and uncertain outcomes.
| In order to make adaptive choices, people need to gather evidence to predict what will happen next. In general, the more frequently an event is observed, the more likely it will occur in the future. Thus the probability of an event is useful to predict its future occurrence. Previous studies have identified regions in the brain that react to rewarding or surprising events, but not to likely events. In the present study, participants had to predict payoffs by observing their repeated occurrence. Functional imaging showed that brain activity in inferior parietal and medial prefrontal cortex increased if the currently observed payoff had been seen many times before. This suggests that these two cortical regions accumulate evidence to predict future events. Further analyses revealed that they belonged to the larger default mode network. This network is involved in introspection and remembering. The inferior parietal and medial prefrontal cortex might thus support the prediction of future events by activating memories of past events.
| In an uncertain environment, probabilities are crucial information because they improve prediction of future events. For humans, information about the likelihood of events can be described with abstract symbols, for instance with a verbal sentence or a pie chart. But in many situations, probabilities are learned through experience by observing the occurrence of events [1]. Animals can only learn probabilities through experience as they have no access to language. Thus understanding how information about probabilities is acquired in the brain is a fundamental question in decision neuroscience for both humans and animals.
In the present study, we focused on the probability of events independently of their value. The motivation came from the observation that people can memorize, make predictions, and decide in the absence of immediate reinforcements. This ability to build a representation of the environment independently of the rewards to be received is made explicit in model-based reinforcement learning [2]. In addition, a separate estimation of probability and value is necessary to ensure rational choices [3], [4]. This principle called “probabilistic sophistication” might seem counter-intuitive because probabilities are combined with values to estimate expected value in many decision models (e.g., expected utility). Nevertheless, the fact that probabilities and values are multiplied does not contradict the necessity to estimate them independently. The concept of probabilistic sophistication is illustrated in Fig. 1.
In psychology, there has been a long tradition of research showing that people can learn the probabilities of stimuli with no value like neutral words or symbols [5]–[7]. In neuroscience, this type of inference has been studied with categorization tasks [8]. In the weather prediction task for instance, participants have to predict the occurrence of two probabilistic outcomes through trial and error (e.g., sunshine or rain). The probability of the outcome is conditional on a set of four symbols. When comparing this task to a control condition, authors have found activation in a large network including the medial and lateral prefrontal cortex, inferior parietal cortex, posterior cingulate cortex and striatum [9]–[11]. A limitation of these studies is the use of a subtraction instead of a parametric approach. It is thus unknown if regions in the brain encode the probability of the outcome in this task.
Following a parametric approach, authors have observed a larger BOLD response in striatum and ventro medial prefrontal cortex when the probability of an anticipated reward increased [12]–[14]. These results have been interpreted in terms of value because for a single and uncertain reward, probability and expected value are positively correlated. Other regions of the brain, like the parietal cortex and the amygdala, have been found to increase with the probability of an upcoming punishment [15], [16]. To support probabilistic sophistication however, one has to identify structures in the brain which encode probability independently of value.
The effect of value can be controlled for by relating brain activity to the probabilities of events and making sure these probabilities do not covary with reward expectation. An event can be defined as a stimulus, its omission, a feed-back and so on. In reinforcement learning studies, authors have shown a larger BOLD response to the occurrence of rare events [17]–[19]. Activity has generally been found in the lateral parietal and prefrontal cortex. Using EEG, numerous studies have shown an enhanced brain response (P300) to rare target in the odd ball paradigm [20], [21]. It should be noted that in these fMRI and EEG studies, brain activity was not always related to the probability of the outcome, but to other measures like surprise or “state” prediction error (one minus the estimated probability of the outcome). However, these measures are highly and negatively correlated with probability. If the surprise or state prediction error is large, the outcome probability is low.
The brain response to rare events has been explained by associative learning theory (as a prediction error) [18], [19] or statistical inference (as a Bayesian surprise) [17], [21], [22]. In a learning context however, we are not aware of experiments showing a positive correlation between brain activity and event probabilities. This is surprising because several models explain choices as the result of evidence accumulation [23], [24]. When the environment is stable (probabilities do not change overtime), the past occurrence of an uncertain stimulus constitutes evidence for its future occurrence.
In a perceptual decision-making task, the agent has to make a decision based on a noisy signal. Several studies in monkeys have shown that the firing rate of neurons in the lateral intraparietal cortex increased over time as a function of the proportion of dots moving in the same direction [25], [26] and this pattern is well explained with artificial neural networks [27]. In these studies, accumulation of evidence is observed by recording the firing rates of neurons with a specific response field [26], [28]. With fMRI, the researcher only has access to the activation of a large population of neurons and evidence accumulated by neurons of one response field (e.g left direction) might cancel out the evidence accumulated by neurons sensitive to a different response field (e.g. right direction).
In a probability learning task evidence is not presented simultaneously but one after another. This offers the opportunity to relate brain activity to the characteristic of the currently observed evidence (which serves as a probe). If the evidence has been observed many times, retrieval models based on accumulation processes predict a stronger brain response because the probe matches numerous traces of past outcomes in memory [29], [30]. In neuroscience, it has been proposed that the inferior parietal cortex plays the role of a mnemonic accumulator because this area is more activated during the successful recognition of old versus new items [31]–[33]. Other regions of the default mode network (medial temporal lobes, medial prefrontal cortex, posterior cingulate cortex) have been found to be more activated for objects which are easily associated to a specific context compared to objects eliciting weak association [34], [35]. According to the principle of an accumulation of evidence in memory, a positive BOLD response can be expected for likely events, particularly in the default mode network.
Overall, studies have identified regions in the brain where activity increased with the probability of a single and random reward. BOLD response related to reward probability has been observed, mainly in striatum [12], [36] and ventro medial prefrontal cortex [13], [37]. However, when the effect of value was controlled for, an increase of activity in response to unlikely outcomes has been found in lateral parietal and lateral prefrontal cortex [17]–[19]. As such, previous studies on learning have shown that the brain reacts to rewarding or rare events but not to likely events. This conflicts with models of perception and memory [27], [29] where activity increases with the accumulation of evidence. In a probability learning task, we found that activity in bilateral inferior parietal and medial prefrontal cortex increased for events which had been observed many times and were likely to occur again.
We developed a task where evidence for future outcomes were balls drawn from a bin. The bin contained balls of different colors and each color was associated to a payoff (Fig. 2a). The composition of the bin was hidden, therefore payoff probabilities were unknown to the participants. However, they had the opportunity to learn these probabilities by observing 10 to 14 drawings from the bin. Balls were sampled one after another with replacement and shown in the center of the bin. The sample payoff was displayed but not the color of the ball. Thus colors were hidden states and payoffs were stimuli (Top insert, Fig. 2b). Colors could be inferred from payoffs because the color-payoff association was shown to the participants in the periphery of the bin. After the 10 to 14 draws, this association changed while color probabilities remained constant. In this resampling phase, 10 to 14 balls were drawn again.
At the end of the sampling and resampling phases, participants had to decide to buy or not a gamble for a certain price. After their choice, the payoff was determined by drawing an additional ball from the bin. If participants decided to buy, they earned the price minus the payoff (this net payoff could be negative). If they decided to pass, the net payoff was 0. To maximize their earning, it was optimal for them to predict the payoffs (stimuli) based on the colors (hidden states). Participants learned the probability of 2, 5, or 10 payoffs (Histograms, Fig. 2b).
For the main analysis, brain activity in the sampling stages was regressed on the probability of the currently observed stimulus, that is the probability of seeing the evidence. Probabilities of stimuli were orthogonal to their values and the value to be expected at the end of the sampling or resampling period. For instance, if a red ball was associated with a low payoff, sampling a red ball increased the probability of seeing this low payoff, but it decreased the expected payoff. The independence between probability and value was obtained by randomly assigning payoffs to colors.
To decide whether to buy the gamble for a certain price or to pass it, participants had to predict the gamble payoff at the end of each sampling stage. Analysis was conducted to determine from gamble choices whether participants estimated probabilities based on the color or payoff history (Fig. S1 in Text S1). If inference is based on colors (hidden states), it can be concluded that people are able to make abstraction of rewards when estimating the likelihood of future outcomes. When extracting probabilities from choices, one needs to control for attitudes towards uncertainty. We did so in a generally accepted way, using prospect theory, which allows one to separately control for loss aversion (“losses loom larger than gains”) and differential risk attitudes.
When a new bin was introduced, initial beliefs were set to equiprobable priors, and subsequent updating was assumed to follow Bayes' law. The posterior stimulus probability increased with the accumulation of evidence. At the end of each sampling stage, probabilities and payoff magnitudes (net of the price) were combined to compute gamble expected value according to prospect theory principles. The decision to buy was predicted from valuation with a logistic regression. We compared models with the Bayesian Information Criterion (BIC), because it can be used with non-nested models and limits the risk of over-fit by penalizing free parameters (a lower BIC is better). Models are indexed by the number of free parameter (M2, M3, etc.) and are presented in the Methods section below.
The most efficient model (the best compromise between parsimony and fitting) was a model with payoff probabilities calculated conditional on the colors of the balls drawn since the presentation of a new bin (model M4, BIC = 1250, Table 1). Colors were hidden states but could be inferred from the observed payoffs. It was more efficient than a model with payoff probabilities calculated conditional on the payoffs observed since the beginning of the sampling or resampling stage (M4a, BIC = 1271). In this model, participants ignore colors and have to re-estimate probabilities after the payoff-color association changed. Further analyses at the individual level showed that model M4 offered a better fit than M4a for all participants (section Behavioral choice & Fig. S2 in Text S1). Thus all participants appeared to be “sophisticated”: their inference was indirect, based on the hidden states behind the observed payoffs, rather than on the observed payoffs directly. Model M4 was more efficient than a model which did not update payoff probabilities (M4b, BIC = 1354). In this latter model, participants used equiprobable payoffs to make decisions (absence of learning).
Model M4 included a prospect theory value function. It was more efficient than a simpler model using a linear value function (M2, BIC = 1312). The shape of the estimated non-linear function revealed diminishing sensitivity for large payoffs (either positive or negative) and greater importance of losses compared to gains. Decision parameters found in previous studies are reported in the footnote of Table 1. Loss aversion was close to the estimation by Tversky and Kahneman in a study made on decisions from description [38]. Diminishing sensitivity was more pronounced in the present study (Fig. 3a).
Model M4 was more efficient than a model that included a prospect theory probability weighting function (M5, BIC = 1251). Indeed, the later model led to a quasi-linear function. We also report the estimation found in decisions from description by Tversky and Kahneman [38] and in decisions from experience by Hau et al. [39]. Probability weighting appears to be minimal in decisions from experience (Fig. 3b).
Finally a reinforcement learning algorithm was estimated (M3). The first payoff observed at the beginning of the sampling or resampling period defined the initial forecast. Then each new payoff was compared to the previous forecast to compute a prediction error. This prediction error multiplied by a learning rate was added to the previous forecast to find the new forecast. This model used a linear value function and bypassed probabilities in order to directly estimate the expected payoff of the gamble. Results showed it had the lowest efficiency (BIC = 1428). Thus probability-based models offered a better explanation of decisions than a reinforcement algorithm.
In sum, the analysis of choices indicated that participants were loss and risk averse (non-linear value function), but there was no indication of a distortion of probabilities (linear probability weighting function). Supporting the principle of probabilistic sophistication, participants learned probabilities based on the hidden states and not simply the observed payoffs. A reinforcement learning model tracking payoffs to estimate gamble values performed worse than any of the probability-based models.
The threshold for significance was set at , uncorrected, cluster voxels for voxel-based analyses (including the identification of ROIs). False Discovery Rates (FDRs) are reported in tables, to gauge the risk of false-positive results. Coordinates are given in MNI space [mm]. The threshold was set to to analyze mean activation in ROIs. Circularity in ROI analysis was avoided with cross-validation. Subject variability was modeled as a random factor in voxel-based and ROI analyses. Details on the GLMs (GLM1, GLM2, etc.) are given in the Brain analysis section of Text S1.
The display of a new payoff in the center of the bin is referred as a stimulus and gives evidence for future outcomes. Stimuli were defined as 1 [s] events and led to a significant activation in the occipital cortex and bilateral hippocampus (GLM1, Table S1 in Text S1).
Probabilities inferred from the history of the hidden states were entered as a covariate to modulate the effect of stimuli (model M4). For instance in Fig. 2a (Resampling), when “68” was displayed the probability of seeing “68” was used as a parametric covariate. When “46” was displayed the probability of seeing “46” was used instead. Thus analyses were conducted with the probability of the currently observed stimulus. Probabilities estimated with model M4 ranged from 0.08 to 0.92. The probability of the stimulus did not correlate with its associated payoff magnitude (, ) or the gamble expected payoff (, ). This is because payoffs were randomly assigned to colors which in turn were randomly assigned to probabilities. For example, in Fig. 2a (Resampling), the stimulus “46”, which was the lowest payoff in the bin, could have a low or high probability of occurrence because it was randomly assigned to the color blue. There is thus no confound between probability and value in our design.
Results showed a positive and significant effect of stimulus probabilities medially in the prefrontal cortex and bilaterally in the angular gyrus (Fig. 4a+b, Table S2 in Text S1). In these regions, brain activity increased when a stimulus was likely to be observed (confirmatory signal). A negative and significant effect of probabilities was also observed in the occipital, superior parietal, and middle frontal gyrus. Activity in the middle frontal gyrus was posterior and reached the precentral gyrus (Fig. S3a in Text S1). A significant effect was also observed in the bilateral hippocampus (Table S3 in Text S1). In these regions, brain activity increased when a stimulus – learned to be rare – was observed (surprise signal). Activation related to stimulus probabilities survived correction for multiple comparisons, except in the hippocampus (FWE, ).
When a payoff was displayed in the sampling or resampling period, it generated a prediction error. This prediction error was calculated with model M4 as the change in expected value before and after the new payoff was revealed. Results indicated that prediction errors did not increase or decrease the effect of stimuli on the brain (GLM2, no table or figure was reported for this non-significant covariate). Thus, the brain encoded stimulus probabilities but not values during the learning phase. We used the sign of the prediction error to define positive and negative stimuli (GLM3). Results showed BOLD response to probabilities in the angular gyri and medial prefrontal cortex for both positive and negative stimuli (Tables S4 & S5 in Text S1). Thus, the encoding of probabilities was comparable for positive and negative stimuli. In short, it appeared that during the learning phase the brain ignored values (probabilistic sophistication) and focused on probabilities.
A BOLD response to unlikely stimuli has been reported in the literature [17], [19]. The BOLD response to likely stimuli is novel. We will focus on this positive correlation in the rest of the results. ROIs were defined as the cluster of voxel significantly activated for a given variable of interest (GLM4) and mean activations in ROIs (GLM5) were further analyzed with mixed effect regressions. ROIs analyses revealed that BOLD response in angular gyrus and medial prefrontal cortex was better explained by stimulus probabilities inferred from the hidden states (model M4) rather than the observed payoffs (model M4a), in line with the behavioral results. The interaction of probabilities with ROI location was not significant. The effect of stimulus probabilities is thus similar in the three ROIs (Table S6 in Text S1, see [40] for the necessity to test interactions before making simple contrasts). The number of colors did not interact with stimulus probabilities, meaning that the effect of probabilities was not influenced by the number of states (Table S6 in Text S1). When the effect of probabilities was estimated for each number of states, it was found to be significant for 2 (), 5 (), and 10 states (, Table S7 in Text S1).
Analysis of choices revealed that expected utility was linear relative to probabilities. We also tested whether BOLD responses in the three ROIs increased linearly with probabilities. The first model included only an intercept and yielded to a BIC of 78707. In the second model, we added a linear effect for stimulus probabilities. This linear effect of probabilities was significant (Table S8 in Text S1) and the BIC fell to 78491, showing an increase in efficiency. Finally, a non linear weighting function was added. The linear effect was again significant (). In contrast with the inverted-S shape of prospect theory, there was a slight diminution of sensitivity for probabilities close to 0 and 1 (Fig. 3c) but this non-linear effect was not significant (, , Table S9 in Text S1). The BIC of this model was 78542, showing a decrease in efficiency compared to the previous model. Similar results were found when the non-linear model was tested on each ROI separately (no table was reported for the separate analyses). Thus, we found no evidence for a non-linear encoding of stimulus probabilities when learned from experience.
Connectivity analysis was conducted to explore the functional link between the ROIs encoding stimulus probabilities and the rest of the brain. Each of the ROIs was taken as a seed region in three separate analyses. Results showed that during the learning phase (compared to the resting phase) the correlation increased between each ROI and voxels in the two other ROIs. Correlations also increased between each ROI and the posterior cingulate (no table or figure was reported for the separate analyses). Connections with posterior cingulate cortex were also observed when voxels of the three ROIs were merged to define a single seed region (Fig. 4c, GLM6, Table S10 in Text S1).
Comparing the active phase (when choices were made without knowing the outcome in advance) to the control phase (when the outcome was known before making choices), significant activity was observed in the occipital cortex, suggesting that visual exploration of the bin was more intense when the outcome was unknown (GLM1, Table S11 in Text S1). In addition, BOLD responses in the right anterior insula and bilateral caudate were significant. These regions have been involved in risky decision-making, which is present in the active phase but not in the control phase [41].
During the learning phase, we have reported above how brain activity changed as a function of probabilities of a specific stimulus. This approach was possible because stimuli (payoffs) were presented one at a time. But the approach cannot be used when the participants deliberated on their choice because all possible outcomes should be contemplated at once. To study the link between brain activity and probabilities during the decision epoch, a measure that summarizes the set of outcome probabilities should be used instead. Here, we chose entropy, which measures the uncertainty reflected in a set of probabilities. Entropy increases as the probability distribution approaches the uniform distribution.
During the deliberation preceding the decision to buy or pass, the expected gamble value (net of the price) was related to activity in the caudate and spread to other regions in the brain (Fig. S4a, Table S12 in Text S1). At the same period, expected value interacted with outcome entropy in the bilateral insula (Fig. 5a & Table S13 in Text S1). An ROI analysis revealed a main and positive effect of expected value in insula. The interaction showed that this effect of value was stronger when the outcome entropy was high (Fig. S4b+c & Table S14 in Text S1). Thus, the insula seems to be especially sensitive to the value of gambles with uncertain outcomes.
In order to quantify uncertainty regarding choices, we computed the entropy of the probabilities (and complementary probabilities) that participants bought into the gamble (as predicted by model M4). Voxel-based analyses showed a significant effect of choice entropy in dorsal anterior cingulate cortex (Fig. 5b & Table S15 in Text S1). ROI analyses were conducted to further test the double dissociation between outcome and choice entropy in insula and anterior cingulate. Results indicated that choice entropy was specifically encoded in anterior cingulate and not insula. The dissociation was not significant for outcome entropy. Finally, BOLD response in the bilateral striatum (putamen and caudate) was related to the net payoff revealed at the end of each decision phase (Fig. 5c & Table S16 in Text S1).
The three ROIs found to encode stimulus probabilities along with the posterior cingulate are all key regions of the default mode network. Regions forming the default network have two characteristics: (1) their spontaneous activity is correlated when people are at rest, (2) they are deactivated during tasks requiring attention to external stimuli [42]. The default mode network includes the inferior parietal cortex, the posterior cingulate cortex, the medial prefrontal cortex, the lateral temporal cortex, and the hippocampus.
To test the involvement of the default network in the present study, we explored the spontaneous correlations between ROIs encoding probabilities and the whole brain during the resting phase (threshold ). Results showed a functional link between each ROI and voxels in the other two. Each ROI was also connected to activity in the posterior cingulate cortex (no table or figure was reported for the separate analyses). Connections with the posterior cingulate cortex were also observed when voxels of the three ROIs were merged to define a single seed region (Fig. 6a, GLM7, Table S17 in Text S1).
Baseline activity was compared between the decision-making task (learning and decision phases) and the resting phase. Results showed an increase of BOLD response in occipital, superior parietal cortex, supplementary motor areas, and lateral prefrontal cortex (red voxels, Fig. 6b, GLM1, S18 in Text S1). Decreased activity was found bilaterally in angular gyrus, supramarginal gyrus, and middle and superior temporal gyri. Decreased activity was also observed in cingulate and medial prefrontal cortex (blue voxels, Fig. 6b, Table S19 in Text S1, similar results were found when comparing the resting phase to the learning phase only). There was a substantial overlap between the task-negative network and voxels encoding stimulus probabilities during the learning period (Fig. 6c). The results indicated that regions reacting to likely events belonged to the default network. On the other hand, there was a substantial overlap between the task-positive network and voxels reacting to rare stimuli (Fig. S3b in Text S1). For a schematic of the main findings, see Fig. 7.
In a complex and uncertain environment, probabilities are essential for predicting future events. To make consistent choices, it is necessary for a decision maker to separate the chances of objective events (“it will snow”) from the values that could potentially be attached to those events (“we can go skiing”). With such a strategy the decision maker can make inference in the absence of immediate reinforcements and quickly adjust his predictions when the reinforcing values of events change [43]. Here, we developed a paradigm in which the probability and value of stimuli were statistically independent. This allowed us to identify the regions in the brain encoding event probabilities and exclude a confound with value.
Analysis of brain activity during the learning period revealed both positive and negative correlations with stimulus probabilities. BOLD response in angular gyrus and medial prefrontal cortex increased for stimuli which had been observed many times in the current trial. This relationship was significant in conditions with 2, 5, and 10 different stimuli. This shows that the brain can keep track of the probabilities of multiple events. Comparison with the resting state condition and connectivity analyses indicated that these regions belonged to the default mode network and that their baseline activity decreased during the task (task-negative network). A negative correlation between stimulus probabilities was observed in the occipital, superior parietal and lateral prefrontal cortex. Here BOLD response increased for improbable stimuli. These regions were more activated during the task (task-positive network). Before a choice was made activity in striatum and insula increased with gamble expected value. The effect of value in insula was stronger when outcome entropy was high, that is when the future was uncertain. Choice entropy which reflects decision uncertainty was preferentially associated to a BOLD response in dorsal anterior cingulate. After the decision, activity in the striatum increased with the net payoff.
The principle of a separation between probability and value, namely probabilistic sophistication [3], [4], was supported by several results. In the learning period, the effect of probabilities was significant for both positive and negative events and the main effect of probability did not interact with event value. No significant effect of value was observed during the learning period. These results suggest that when reinforcements are delayed, the brain focuses on event probabilities while abstracting from rewards. It is only during the decision period that activation in relation with value was observed.
These results are relevant for the debate on value-based and model-based reinforcement learning [18]. In value-based reinforcement learning, the agent learns the expected value associated to a situation or action by updating his forecast with a reward prediction error. Through this process, the agent acquires information about value but remains ignorant of probabilities. This stands in sharp contrast with model-based reinforcement learning. There, in order to forecast future rewards, the agent forms a representation of how the world “behaves” and this can be done by learning the probabilities of all events given the current situation (i.e. “state transition probabilities”, [44]). A neural signature of probabilities but not value was observed during the learning period. In addition, model comparison showed that choices were better explained by a probability rather than a reinforcement learning algorithm. Thus both behavioral and biological data favored model-based over value-based reinforcement learning in our task. By showing both positive and negative BOLD response to event probabilities, the present study add to the previous literature on model-based reinforcement learning [45].
The functionality of the regions encoding probabilities deserves further discussion. Based on prior literature and our own results, we would argue that activation correlating with rare stimuli in occipital cortex reflects the visual exploration of the bin, while that in parietal and middle frontal gyrus captures the attention triggered by surprising events. Activation in hippocampus enhances encoding of rare stimuli. In contrast, the positive correlation between probabilities and activation in angular gyrus and medial prefrontal cortex would reflect the reactivation and reinforcement of past events in memory.
BOLD activity increased in the occipital cortex for rare stimuli and this could be due to visual exploration. When a rare payoff was sampled, its probability of occurrence increased. This might incite participants to identify its associated color and re-evaluate its relative importance by looking at all the colors and payoffs in the periphery of the bin. Previous studies have shown increased activation in the occipital cortex for visually incongruent stimuli and this effect seems to generalize to rare events in our study [46]. Rare events were also related to activation in the superior parietal cortex and middle frontal gyrus and this could be explained by attentional processes. Indeed, these regions were more activated during the task compared to the resting period (task-positive network) and have been related to attention or the oddball effect [17], [47]. Activity in the hippocampus was observed when a new stimulus was displayed, and the effect was stronger when the stimulus was rare. Lesions to the middle temporal area and the hippocampus can lead to amnesia and the inability to retain new information [48]. The BOLD response in the hippocampus suggests that rare events benefit from a better encoding when they occur. This is consistent with behavioural studies showing that surprising stimuli are better memorized [49].
Activity in the default network has been found to increase during tasks of theory of mind, mind wandering, and memory [50], [51]. On the contrary, it has been found to decrease when participants pay attention to external stimuli (task-negative network). This was also the case in our task because participants had to pay attention to the sampled payoffs. While controlling for the effect of the task, activity increased in several regions of the task-positive network when a rare stimuli was displayed. On the contrary, a BOLD response in angular gyrus and medial prefrontal cortex increased for frequent stimuli. A possible explanation for this novel result is that frequent stimuli attract less attention and hence allow for more resting-state introspection, the role traditionally assigned to the default mode network. This switch would be consistent with optimal use of the limited amount of energy available in the brain [52]. However, the switch would have to take place within the time frame of display of our stimuli (1 [s]). If this interpretation is indeed true, our findings would amount to evidence for high-frequency switching between elemental states of the brain, namely, attention and rest. An alternative explanation is that activity in angular gyrus and medial prefrontal cortex reflects a distinct process, namely, evidence accumulation. This process can be modeled as we did, in terms of learning of probabilities. A drift-diffusion approach [53] could be used instead, though this approach is fundamentally the same.
A cognitive mechanism that could explain the positive correlation between stimulus probability and brain activity is memory [29]. Cognitive psychologists have developed models centered on memory to explain how people judge the likelihood of events [54]. In these models, each outcome is encoded as a trace in memory. An event will be considered as probable if many traces are retrieved from memory in response to a probe (the payoff in our task). Neuroimaging studies have confirmed the involvement of parietal and medial prefrontal cortex in memory: activity in these areas predicts the successful recognition of items [31], [55]. Furthermore, many studies have demonstrated involvement of the default mode network in memory tasks [56], [57]. As a consequence, the positive brain response to stimulus probabilities in the angular gyrus and medial prefrontal cortex might reflect the activation and reinforcement of memory traces in reaction to a probe. This hypothesis is compatible with an “attention to memory” model developed in neuroscience [32], [58]. In this model, activation in the inferior parietal cortex reflects the attention captured by information retrieved from memory. Still, because fMRI can only recover correlation, other approaches like TMS are needed to determine the causal role of angular gyrus and medial prefrontal cortex in the acquisition and retrieval of event probabilities.
In decision neuroscience, the posterior cingulate and ventro-medial prefrontal are often involved in the judgement of value [59], [60]. For instance the ventro-medial prefrontal cortex is more activated when participants see the image of food they like [61], [62]. A recent study has shown that the time spent watching an item increased the likelihood to choose it and this type of behavior was well formalized by drift-diffusion models [63]. Preferences depend on the sensory characteristics of goods, but they are also shaped by our past experience and memories [64]. The reactivation of memory traces could partially explain why a key structure to evaluate the value of goods, the ventro medial prefrontal cortex, belongs to the default mode network and not to a task-positive or saliency network like the striatum [65]. Accordingly, a recent study has shown that affective value and associative processing shared a common substrate in medial prefrontal cortex [66].
Our study sheds new light on decision making under uncertainty when uncertainty is described as opposed to experienced [1]. In a task where decisions were based on experience, BOLD response to uncertain stimuli increased linearly with their probabilities of occurrence. This was confirmed in behavior: choices exhibited no bias is assessment of probabilities, in contrast to decision making based on description of available gambles [67]. Our results therefore cast doubt on the generalizability of probability weighting in prospect theory to decision making from experience [68].
In addition to a better understanding of the neural foundation of probability learning, the present study brings new knowledge concerning the representation of various kind of uncertainties in the brain [69]. Previous studies have linked uncertainty to activity in the insula [70], [71], but also in the anterior cingulate cortex [72]. In the present study, we found that BOLD response in insula and dorsal anterior cingulate were related to different forms of uncertainty. Activity in the anterior cingulate correlated with choice entropy which reflected uncertainty in making a choice. The later interpretation matches previous studies reporting BOLD response in dorsal anterior cingulate when a conflict existed between several responses [73] (difficulty of choice).
Activity in the insula increased with the gamble expected value and this effect was more pronounced when the outcome entropy was high. Entropy corresponds to the notion of expected uncertainty discussed by A. Yu and P. Dayan [74]. It is a function of probabilities only and thus does not depend on the value associated with the stimuli. An agent separating probability from value would favor entropy over reward volatility to estimate risk. When the outcome is a single and uncertain payoff, its standard deviation and entropy covary. This might explain why previous studies have found activation related to payoff standard-deviation in the insula [71]. Another possibility is that the insula becomes sensitive to entropy when participants learn state probabilities as in the present study and rely less on summary statistics like payoff mean and variance [75].
The general view that emerges from our study is that the brain does not only react to rewarding or surprising events, but also to likely events. When people observed uncertain stimuli, the average activity in the default mode network decreased compared to a resting condition. Nevertheless, the functional connectivity in this network increased and stimulus probabilities were positively correlated with BOLD response in angular gyrus and medial prefrontal cortex. Thus activity in these two regions signalled the accumulation of evidence (confirmatory signal). Brain response to uncertain stimuli increased linearly in probability and there was no evidence of probability weighting in choices. Further research is needed to test if the brain response to likely events reflects an activation of memory traces (internal world) or a lack of attention to the environment (external world).
Twenty-five students from the Université de Lausanne and the Ecole Polytechnique Fédérale de Lausanne were enrolled in the study. One participant was removed from the analysis because of significant head movements. Another, because her decisions to buy the gamble were random. The analyzed sample included 23 participants (10 women, 13 men; median age = 22, min = 19, max = 30; all right handed). The study took place at the University Hospital of Lausanne and was approved by its institutional review board. At the end of the experiment, students received 1/10 of their net play money in real currency, in addition to a 10 Frs (Swiss francs) participation reward.
To explain the task, the investigator read the instructions aloud and students played with one demonstration bin. They completed the task in a 3 Tesla MRI scanner. During the functional image acquisition, participants watched the display through goggles and indicated their decision to buy or to pass the gamble by pressing the left or right button of a response box. Participants learned probabilities and made decisions on 9 different bins. After bin 3 and 6, a resting phase of 60 [s] was introduced.
Payoffs were determined by the colors of balls drawn from a bin. Bins contained balls of different colors, with same-colored balls yielding the same payoff. The composition of the bin was hidden therefore probabilities were unknown. The time line for one example bin is shown in Fig. 2a. During the first sampling stage, 10 to 14 balls were drawn from the bin one after another and the associated payoff was displayed for 1 [s] at the center of the screen. Balls were drawn with replacement. Only stimuli representing the payoffs were shown. Colors were hidden states. These states could be inferred from the colored balls displayed in the periphery of the bin.
This learning phase was followed by a decision phase. The participant had to decide whether to buy the gamble or not for a certain price. After each choice, an additional ball was drawn. If the participant bought the gamble, he earned the payoff written on the ball minus the price. Otherwise, the payoff was 0 and the play money remained unchanged. Four choices were made without knowing the outcome in advance (active condition) and two choices were made while knowing the outcome in advance (control condition). For each of the six choices, a different price was posted. Prices were drawn from a uniform distribution between the minimum and maximum payoffs. After each decision, a message indicating that the gamble was bought or passed was shown (but the payoff was not shown to limit learning in the decision period). The total net payoff of the current decision period was displayed after the six choices.
The learning and decision phases were repeated with the same bin after changing the color-payoff association. That is, color probabilities remained unchanged (same composition of the bin), but each color was associated with a new payoff. It was thus adaptive to learn probabilities based on the color of the past drawings. In Fig. 2, the color-payoff association in panel a is reproduced in the top insert of panel b. For instance, red was associated with 57 in the sampling stage and with 67 in the resampling stage. Bins contained balls of 2, 5, or 10 different colors. The probabilities used to generate the drawings are represented by the histograms in Fig. 2b. Because balls were drawn with replacement, these probabilities remained constant during the sampling and resampling stages.
Nine different bins were presented in the task. Importantly, colors were randomly assigned to probabilities at the beginning of each new bin. Payoffs were randomly assigned to colors at the beginning of each sampling stage. As a consequence, payoff probabilities are orthogonal to payoff magnitudes and expected payoff. Uncertainty at decision time increased with both the number of possible payoffs and the payoff standard-deviation. To disentangle their effects, these two factors were manipulated independently (Fig. S5). See section Task in Text S1 for more details.
|
10.1371/journal.ppat.1006221 | The thermodynamics of Pr55Gag-RNA interaction regulate the assembly of HIV | The interactions that occur during HIV Pr55Gag oligomerization and genomic RNA packaging are essential elements that facilitate HIV assembly. However, mechanistic details of these interactions are not clearly defined. Here, we overcome previous limitations in producing large quantities of full-length recombinant Pr55Gag that is required for isothermal titration calorimetry (ITC) studies, and we have revealed the thermodynamic properties of HIV assembly for the first time. Thermodynamic analysis showed that the binding between RNA and HIV Pr55Gag is an energetically favourable reaction (ΔG<0) that is further enhanced by the oligomerization of Pr55Gag. The change in enthalpy (ΔH) widens sequentially from: (1) Pr55Gag-Psi RNA binding during HIV genome selection; to (2) Pr55Gag-Guanosine Uridine (GU)-containing RNA binding in cytoplasm/plasma membrane; and then to (3) Pr55Gag-Adenosine(A)-containing RNA binding in immature HIV. These data imply the stepwise increments of heat being released during HIV biogenesis may help to facilitate the process of viral assembly. By mimicking the interactions between A-containing RNA and oligomeric Pr55Gag in immature HIV, it was noted that a p6 domain truncated Pr50Gag Δp6 is less efficient than full-length Pr55Gag in this thermodynamic process. These data suggest a potential unknown role of p6 in Pr55Gag-Pr55Gag oligomerization and/or Pr55Gag-RNA interaction during HIV assembly. Our data provide direct evidence on how nucleic acid sequences and the oligomeric state of Pr55Gag regulate HIV assembly.
| Formation of any virus particle will require energy, yet the precise biophysical properties that drive the formation of HIV particles remain undefined. Isothermal titration calorimetry (ITC) is a biophysical technique that is the gold standard to reveal parameters governing biochemical and biophysical reaction. However, ITC requires large amount of proteins for analysis. As large quantities of full-length recombinant HIV Pr55Gag proteins have not been available in the past 30 years due to technical limitation, a comprehensive thermodynamic analysis of full-length HIV Pr55Gag has not been possible. Here, we have generated sufficient amount of full-length recombinant HIV Pr55Gag protein for isothermal titration calorimetry analysis. Our analyses have shown that the major interactions amongst HIV proteins and RNA sequences during viral assembly are energetically favourable reactions. In other words, HIV Pr55Gag proteins and viral RNA have evolved to overcome the energy barrier for virus formation by utilising energy obtained from protein-RNA interactions in order to facilitate the viral assembly process. Furthermore, HIV also use the oligomeric states of HIV Pr55Gag proteins and the RNA sequences as means to regulate the viral assembly process.
| The assembly of the HIV particle is orchestrated by the HIV-1 Gag precursor protein (Pr55Gag). As the major structural protein forming the virus, it is essential that Pr55Gag packs tightly with other Pr55Gag molecules to shape the virus shell while also encapsulating the genomic RNA required for its replication. Fittingly, Pr55Gag is comprised of multi-functional domains that facilitate key interactions involved in the early stages of the assembly process. The nucleocapsid (NC) domain is responsible for genomic RNA packaging and the capsid (CA) domain mediates the key interactions to promote Pr55Gag-Pr55Gag oligomerization (for review see [1–3]). While both Pr55Gag-nucleic acid and Pr55Gag-Pr55Gag interactions have been identified as major contributors to the assembly process [4], the mechanistic details of how they regulate HIV assembly are not well defined, specifically in the context of full length Pr55Gag.
It is generally accepted that nucleic acid binds to Pr55Gag and acts as a scaffold for Pr55Gag oligomerization [5–7], although short oligonucleotides (10 nucleotides or less) are not sufficient to bridge multiple Pr55Gag molecules to support assembly [8, 9]. The p6 C-terminus truncated recombinant Gag protein (Pr50GagΔp6), with the addition of nucleic acid and lipid-mimicking inositol phosphate, has previously been shown to assemble in vitro, forming virus-like particles (VLPs) in the absence of other viral proteins [7, 8]. Cellular imaging and immunoprecipitation studies with virion producing cells have suggested that nucleic acid initially binds to cytosolic Pr55Gag, therefore promoting the formation of low order oligomers of Pr55Gag in the cytosol; higher order Pr55Gag oligomers only occur upon Pr55Gag binding to the plasma membrane [10–13]. However, cytoplasmic HIV-1 RNA has also been shown to traffic to the membrane by passive diffusion independent of Pr55Gag, suggesting that the viral RNA genome may bind to Pr55Gag molecules on the plasma membrane where a high local concentration of Pr55Gag has already been achieved for oligomerization [14]. At the plasma membrane, the assembly of Pr55Gag has also led to the reorganization of the lipid membrane, such as the nano-clustering of phosphatidyl inositol (4,5) biphosphate lipid [15]. Clearly, the relationships between nucleic acid binding and Pr55Gag oligomerization that lead to virus assembly are still unclear. Moreover, the energy requirements and thermodynamic properties that drive these two basic components to facilitate the formation of the viral particle are currently not known. In this regard, a thermodynamic analysis of the assembly process will provide critical information (such as, the affinity and the stoichiometry between ligands and substrates, plus the energetics involved in these reactions) to define the mechanisms of the process. Specifically, enthalpic and entropic components of the binding reaction derived from the analysis will enable us to gain insight into the mechanisms of the interactions that take place.
Apart from the packaging signal (Psi) of the genomic RNA that is involved in the encapsidation of viral genome, the contributions from the rest of the viral RNA sequence to the assembly process is not well defined. Early studies have indicated that the mature nucleocapsid domain on its own displays preferential binding to selected nucleic acid sequence motifs in addition to the Psi packaging sequences [9, 16, 17]. Studies comparing Psi and non-Psi RNA binding to C-terminal truncated Pr55Gag protein reported differences in the electrostatic and non-electrostatic interactions of Pr55Gag proteins with different RNAs [18], and we have shown how Pr55Gag selected between unspliced and spliced HIV RNA in vitro for viral assembly [19, 20]. Recent cross-linking immunoprecipitation experiments have shown that different RNA motifs interact with Pr55Gag at different stages of virus biogenesis [21]. However, the functional significance and the mechanisms of these specific Pr55Gag-RNA interactions in regulating the assembly of Pr55Gag have yet to be shown.
A key bottleneck in acquiring fundamental information about the mechanistic details of Pr55Gag assembly has been the limited availability of full-length recombinant Gag protein for analysis, as large quantities of full-length recombinant Pr55Gag are difficult to produce [8, 22]. Here, we have produced large quantities of recombinant Pr55Gag and use in vitro systems to study the early steps of HIV assembly. Using the entire form of recombinant Pr55Gag in isothermal titration calorimetry (ITC) studies, we present the first thermodynamic analysis of how HIV Pr55Gag and RNA regulate assembly. More specifically, we directly demonstrate that: (1) the specific interaction between Pr55Gag and Adenosine (A)-containing RNA motifs; and (2) the oligomerization of Pr55Gag; are energetically favourable reactions that facilitate virion particle formation. Our findings benchmark the thermodynamic regulations of retroviral assembly, and provide a platform to interrogate, and ultimately to reveal, structural and biophysical properties between HIV Pr55Gag and host cell factors during viral replication.
The binding of RNA to Pr55Gag and the oligomerization of Pr55Gag molecules are two of the basic components of viral assembly. However, a thermodynamic analysis unpacking how these two components drive the assembly process has never been conducted in the context of full length Pr55Gag. We produced sufficient quantities of recombinant full-length Pr55Gag (S1 Fig) and validated its capacity to oligomerize and form VLPs in the presence of RNA in vitro, similar to previous observations with the truncated Pr50GagΔp6 [8, 23–26](S2 Fig).
Using ITC, we first characterized the binding between Pr55Gag protein (domain arrangements of Pr55Gag proteins used schematized in Fig 1A) and the packaging signal stem loop 3 RNA (Psi SL3 RNA), a region of the viral RNA involved in genomic packaging and known to bind to NC with high affinity [17, 27]. Representative ITC curves were shown in Fig 1B, while data in Fig 1C–1F reflect the means of all three experiments. The representative ITC curves in Fig 1B are the raw values prior to adjustments (such as fluctuations of detected temperature of the reaction; fluctuation of cabinet temperature of the ITC chamber for each point of the measurement; and the normalization of baseline temperature with the corresponding buffer controls). The ITC curves also acted as quality assurance with these ITC experiments, demonstrating the reliability of the data. We first benchmarked the thermodynamics properties of HIV assembly using Pr55Gag molecules that are capable of oligomerizing (Pr55Gag and Pr50GagΔp6) and Psi SL3 RNA that is the primary determinant for genomic RNA selection in the cytoplasm. Both Pr55Gag and Pr55GagΔp6 displayed a favourable binding enthalpy [ΔH ~ -17 kcal mol-1] as measured by the maximum heat energy release during ITC titrations of Psi SL3 RNA into the Pr55Gag protein (Fig 1B and 1C). The reaction is also thermodynamically favourable with ΔG<0. The detected energy release in these reactions is likely to be the sum of both Pr55Gag-RNA interaction and Pr55Gag-Pr55Gag oligomerization. To estimate the fraction of energy released that is contributed by Pr55Gag-RNA vs Pr55Gag-Pr55Gag oligomerization, we have used a mutant, Pr55Gag WM316-7AA (or Pr55Gag WM) [28] that is incapable of supporting the dimerization of Pr55Gag via the CA-hexamer dimerization domain. Therefore, mutant Pr55Gag WM is known to partially suppress the assembly of HIV via the dimerization defect of the CA hexamers [4]. As measured by the maximum heat energy release during ITC titrations of Psi SL3 RNA into the Gag protein, interaction of Psi SL3 RNA with Pr55Gag WM displayed a favourable binding enthalpy [ΔH ~ -10 kcal mol-1] (Fig 1B and 1C), but it was only 60% of the energy release of the wild type Pr55Gag-RNA interaction (Fig 1B and 1C). These data implied that the oligomerization of Pr55Gag during Pr55Gag-RNA interaction contributed to at least 40% of energy detected in the ITC experiment, and the Pr55Gag- Pr55Gag interactions increased its favourable binding enthalpy with Psi SL3 RNA. In support of this, interaction of Psi SL3 RNA with both processed NC proteins (p15NC-SP2-p6 and p7NC that lack the CA-CA interaction domains for oligomerization) also displayed less favourable binding enthalpies [ΔH ~ -5-6 kcal mol-1] compared to the full-length Gag proteins Pr55Gag.
The interaction of Pr55Gag and Pr50Δp6Gag with Psi SL3 RNA however had a more unfavourable entropic contribution compared to that of Pr55Gag WM (Fig 1D). As entropy is a measure of the disorder of the reaction [29], the unfavourable entropy likely results from loss of conformational freedom due to Pr55Gag-Pr55Gag oligomerization. Nevertheless, the larger accompanying favourable enthalpic contributions observed with Psi SL3 binding to Pr55Gag and Pr50Δp6Gag compensates for the unfavourable entropic component to produce an overall favourable Gibbs free energy (ΔG<0) of binding (Fig 1E), highlighting that the Pr55Gag-RNA (SL3) interaction is an energetically favourable process overall. On the other hand, reactions with processed NC proteins (p15NC-SP2-p6 and p7NC) were characterized by favourable entropic contributions, likely driven by displacement of ordered water molecules [30] as the result of NC—Psi RNA interaction (Fig 1D). The favourable entropic component of Psi SL3 binding to p15NC-SP2-p6 and p7NC, combined with the accompanying favourable enthalpic component contributed to an overall favourable Gibbs free energy (ΔG<0) of binding.
Interestingly, calculated binding affinities for Pr55Gag [Kd = 0.081 ± 0.025 μM], Pr50Δp6Gag [Kd = 0.098 ± 0.003 μM], Pr55Gag WM316-7AA [Kd = 0.088 ± 0.035 μM], p15NC-SP2-p6 [Kd = 0.057 ± 0.011 μM] and p7NC [Kd = 0.048 ± 0.023μM] did not differ significantly (Fig 1F). These data imply that the binding between Pr55Gag and SL3 RNA for genomic RNA selection during viral assembly is independent of the oligomerization state of Pr55Gag. Our ITC results demonstrate that the binding of Psi SL3 RNA to Pr55Gag is energetically favourable (ΔG<0). Bindings of Psi SL3 to p15NC-SP2—p6 and p7NC are results of both favourable entropic and enthalpic contributions, suggesting the involvement of both polar and hydrophobic interactions [30]. By definition, these interactions must occur between the Psi SL3 RNA and the nucleocapsid protein. Reactions of Psi SL3 with full length Pr55Gag that has an increased capacity to oligomerize were driven by a larger favourable enthalpy, which helps in overcoming the unfavourable entropy brought about by conformational restrictions due to oligomerization.
During the assembly of HIV, there are only 2 sets of SL3 sequences in the packaged HIV dimeric RNA genome. As a result, most of the 1500–2500 Pr55Gag molecules in the assembled virion must bind to other segments of the RNA genome. HIV Pr55Gag has been previously reported to transiently interact with different RNA sequence motifs during the viral assembly process [21]. It has been suggested that cytosolic Pr55Gag consisting of low-order Pr55Gag oligomers interact with Guanosine Uridine (GU)-containing RNA sequence motifs, while high-order Pr55Gag oligomers in the immature virion are more likely to associate with Adenosine (A)-containing RNA sequence motifs (schematized in Fig 2A). However, the role these non-packaging segments of the RNA genome play in viral assembly still remains unanswered. These GU-containing RNA sequences and A-containing RNA sequences are also interspersed throughout the HIV genome and in many cellular RNA sequences (S3 Fig).
To investigate whether the thermodynamic relationship with Pr55Gag can provide insight into the function of these sequences, we conducted ITC analyses of Gag interaction with the top 4 RNA sequence motifs that were previously identified to interact with HIV Gag either in the cytosol (GU-containing RNA motifs: 5’GAUGG3’ and 5’UGUGG3’) or within the immature virion (A-containing RNA motifs: 5’GAGAA3’ and 5’AAGGA3’) [21]. Binding of the 20 mer of the GU-containing RNA motif 4x 5’-GAUGG-3’ RNA to the oligomerization-impaired Pr55Gag WM (mimicking the low-order oligomeric form of cytosolic Gag) resulted in an enthalpy release [ΔH ~-20 kcal mol-1] (Figs 2B and 3A), which is comparable to that released in the binding to processed NC proteins (p15NC-SP2-p6 and p7NC) (Figs 2B and 3A). In contrast, higher enthalpy was released in the binding of the same 20mer GU-containing RNA motif 4x 5’-GAUGG-3’ RNA with the oligomeric forms of Gag (Pr55Gag and Pr50GagΔp6) with means of ΔH ~-45 kcal mol-1 and -35 kcal mol-1, respectively (Fig 3A) (or raw values of ΔH ~-35 kcal mol-1 and -30 kcal mol-1, respectively, in the un-adjusted representative ITC curves, Fig 2B), suggesting that oligomerization contributes at least in part to the favourable enthalpy release of Pr55Gag binding with GU-containing RNA. These data are consistent with those observed with Psi SL3 RNA based Pr55Gag-RNA ITC studies (Fig 1). In support of this, more enthlapy release was also observed in the interaction between the 20 mer of A-containing RNA motif 4x 5’-GAGAA-3’ RNA and the oligomeric forms of Gag (Pr55Gag and Pr50GagΔp6), with means of ΔH ~-80 kcal mol-1 (Fig 3A) (or raw values of ΔH ~-70 kcal mol-1 and -60 kcal mol-1, respectively, in the un-adjusted representative ITC curves, Fig 2B), over the interaction between the 20 mer of A-containing RNA motif 4x 5’-GAGAA-3’ RNA and the oligomeric-deficient forms of Gag (Pr55Gag WM)[ΔH ~ -60 kcal mol-1] (Fig 3A) (or raw value of -40 kcal mol-1 in the un-adjusted representative ITC curve, Fig 2B). Agreeably, ITC analyses repeated with the second set of GU-containing and A-containing RNA sequence motifs (4x 5’-UGUGG-3’ and 4x 5’-AAGGA-3’, respectively) produced a similar trend in binding enthalpies measured across the Gag constructs (Figs 2C and 3C). Our results showed that the favourable enthalpy and Gibbs Free energy release from oligomerization can potentially be a general feature of RNA-Pr55Gag protein interaction during viral assembly.
Conversely, across the Gag constructs tested independent of their size and oligomeric capacity, thermodynamic analysis also indicated that the interaction of the A-containing RNA motifs with Gag were characterized by about 2-times more favourable enthalpy compared to that of the GU-containing RNA motifs (Fig 3A and 3C). This result suggests that the type of RNA sequence interacting with the Gag protein also contributes to the favourable release of enthalpy. Additionally, the type of RNA sequence may also be a determinant of how tightly packed Gag can bind. An analysis of the stoichiometry of binding showed a trend that 1 A-containing RNA molecule would bind to 5 Gag molecules (RNA:Gag N~0.2) in comparison to 1 GU-containing RNA molecule would bind to 3 Gag molecules (RNA:Gag N~0.33) (Fig 4). However, it is important to acknowledge that the binding stoichiometry data are related to a single type of GU-containing RNA motif and a single type of A-containing RNA motif, and these motifs are artificially presented in 4 consecutive repeats within a synthetic RNA. The potential importance of different RNA-Gag binding stoichiometry must be validated using multiple different natural A-containing RNA motifs and natural GU-containing RNA motifs within the context of HIV RNA genome.
Furthermore, a larger unfavourable entropy was detected when A-containing RNA motifs were used in ITC experiments compared to GU-containing RNA motifs (Fig 3B and 3D), suggesting that these designated A-containing RNA motifs might in part better support tight packing of Pr55Gag-RNA and/or Pr55Gag-Pr55Gag interaction during the biological process, leading to greater loss of conformational freedom. Nevertheless, the larger favourable binding enthalpy associated with binding of A-containing RNA motifs to Pr55Gag overcomes the unfavourable entropy to drive the reaction forward. These differential thermodynamic relationships between Pr55Gag and A-containing RNA motifs in immature virus vs Pr55Gag and GU-containing RNA motifs in cytoplasm have provided insight on how viral particle formations are regulated. The Gibb’s free energy (ΔG) consistently remained at ~-10 kcal mol-1 (Fig 3E and 3F), indicating the process is an energetically favourable spontaneous event. We have also performed additional control experiments showing that less than 5% of materials can be pelleted from solution after our ITC experiments using Gag and RNA (S4A Fig). Furthermore, similar amount of pelletable materials (<5%) were collected in parallel ITC experiments when nucleic acid free buffer was injected into the ITC chamber containing recombinant Gag (S4B Fig). These data implied that no virus-like-particles were generated in these ITC reactions under our experimental conditions when ‘low’ concentrations of Pr55Gag and RNA were used (S4 Fig).
In addition to the more favourable binding enthalpy of A-containing RNA to various Gag constructs, calculated binding affinities (kd) from ITC A-containing and GU-containing RNA binding curves revealed that oligomeric capable Pr55Gag displayed ~3-times stronger binding affinity for the immature virus A-containing RNA motif over the cytosolic GU-containing RNA motif (Fig 5A and 5B). This occurrence was not observed when oligomerization-impaired forms of Gag (Pr55Gag WM) were used in parallel experiments (Fig 5A and 5B). Unexpectedly, Pr50Δp6Gag, which is capable of high-order Gag oligomerization, also did not display binding preference toward A-containing RNA motif (Fig 5A and 5B). It is possible that the p6 domain in Pr55Gag plays a part in facilitating the binding of A-containing RNA to oligomeric forms of Pr55Gag, implying a previously unknown role of p6 late domain in the packing of Pr55Gag-Pr55Gag and/or Pr55Gag-RNA interaction at the late stage of HIV assembly.
To independently assess whether cytosolic low-order Gag oligomer binding with A-containing RNA is a less energetically favourable process than the interaction between high-order Gag oligomer and A-containing RNA, three additional Gag oligomerization defective mutants were engineered for analysis. These Gag oligomierization-impaired mutants are: (1) Pr55Gag (CA Helix 6 mutations, TTSTLQ 239–44 AASALA), Pr55Gag [CA Helix 6] (Fig 6A left panel) [28, 31]; (2) Pr55Gag (CA Helix 10 mutation, D329A), Pr55Gag [CA Helix 10] (Fig 6A middle panel) [28, 31]; and (3) multiple sites Gag oligomerization mutant (designated Pr55Gag [CA All 4]) (Fig 6A right panel) that includes 4 sets of mutations at the: Pr55Gag (CA dimerization interface WM 316–7 AA), Pr55Gag (CA helix 6 TTSTLQ 239–44 AASALA), Pr55Gag (CA helix 10 D329A), and Pr55Gag CA major homology region (MHR K290A) [28]. These respective mutations (ie helix 6, helix 10 and MHR mutations) were chosen based on their reported inhibitory effects on viral assembly [28, 31], likely through their disruption of important intra-hexameric contacts within the immature CA hexamer [32] (Fig 6A). ITC analyses of these three Pr55Gag oligomerization impaired mutants with the 20mer of cytosol GU-containing RNA (4x 5’-GAUGG-3’) showed minimal to no detectable binding (Fig 6A). In contrast, the binding of these same mutants with the 20mer of immature HIV A-containing RNA (4x 5’-GAGAA-3’) have led to energy release in the range of [ΔH ~ -18-24 kcal mol-1] (Fig 6A). These data are consistent with that of the previously described oligomerization impaired mutant, Pr55Gag WM (Fig 2), showing that oligomerization of Pr55Gag can help to increase the level of energy being released during Pr55Gag-RNA interaction in HIV assembly.
Given that the C-terminus His-Tag is present in all of the recombinant Gag used in this study thus far, it is unlikely that the His-Tag would selectively interfere with the binding for one of these three groups of RNA (Psi RNA, GU-containing RNA, and A-containing RNA) with Gag. However, to directly rule out any potential selective bias that might be introduced by the C-terminus His-Tag, a Tobacco Etch Virus (TEV) protease cleavage site was engineered in between the C-terminus end of Gag sequences and the His Tag for Pr55Gag, Pr50GagΔp6, and Pr55Gag WM. The His-Tag was removed via TEV protease digestion, and the His-Tag free recombinant Gag proteins (Pr55Gag-TEV, Pr50GagΔp6-TEV, and Pr55Gag WM -TEV) were subsequently column purified for analysis. ITC analyses of Pr55Gag-TEV [Fig 6B left panel], Pr50GagΔp6-TEV [Fig 6B middle panel], and Pr55Gag WM-TEV [Fig 6B right panel] with either GU-containing RNA (4x 5’-GAUGG-3’) or A-containing RNA (4x 5’-GAGAA-3’) resulted in energy release ranging from ΔH ~ -8 kcal mol-1 to ~ -50kcal mol-1 (Fig 6B). More specifically, the ΔH are: (1) ~-10 kcal mol-1 for Pr55Gag-TEV-GU-containing RNA binding [Fig 6B left panel, dark blue]; (2) ~-10 kcal mol-1 for Pr50Gag Δp6-TEV-GU-containing RNA binding [Fig 6B middle panel, dark orange]; (3) ~-10 kcal mol-1 for Pr55Gag WM-TEV-GU-containing RNA binding [Fig 6B right panel, green]; (4) ~-45 kcal mol-1 for Pr55Gag-TEV-A-containing RNA binding [Fig 6B left panel, light blue]; (5) ~-25 kcal mol-1 for Pr50Gag Δp6-TEV-A-containing RNA binding [Fig 6B middle panel, light orange]; (3) ~-25 kcal mol-1 for Pr55Gag WM-TEV-A-containing RNA binding [Fig 6B right panel, dark yellow]. Analogous to the ITC results using the His-Tag containing Gag (Pr55Gag, Pr50GagΔp6, and Pr55Gag WM) (Fig 2), Gag binding with A-containing RNA (4x 5’-GAGAA-3’) were consistently shown to be a more energetically favourable reaction than parallel analyses that used GU-containing RNA (4x 5’-GAUGG-3’) as substrates (Fig 6B). Furthermore, the binding between Pr55Gag-TEV and A-containing RNA was a more thermodynamically favourable reaction than that between Pr50GagΔp6-TEV and A-containing RNA, highlighting the potential role of p6 in this process. Although the C-terminus His-Tag has no selective bias on the overall data interpretation, it is important to acknowledge that reduced levels of enthalpy (ΔH) were also detected with the Gag-RNA binding when the C-terminus His-Tag was removed from the recombinant Gag (Fig 2 vs Fig 6B), implying that the C-terminus His-Tag has consistently increased the amounts of enthalpy release during Gag-RNA interaction.
It is conceivable that the synthetic 20mer RNA with the 4 consecutive A-containing RNA repeats (4x 5’-GAGAA-3’) does not truly represent the binding between HIV Gag and RNA sequences during viral assembly. To directly examine the relationship between authentic HIV RNA sequences and HIV Pr55Gag protein during viral assembly, a 37mer RNA fragment representing the coding region of HIV reverse transcriptase (RNA positions 2671–2707) was used for ITC analyses. Using CLIP-sequencing, this RNA sequence (HIVNL4.3RNA2671-2707), consisting 3 independent A-containing RNA motifs, has been previously identified to be important for Pr55Gag binding in immature HIV [21]. The same 37mer RNA fragment with mutations of the AGAAA RNA motifs (HIVNL4.3RNA2671-2707 with AGAAA mutation) was used as a control. ITC analyses of Pr55Gag-TEV, Pr50GagΔp6-TEV, and Pr55Gag WM-TEV with the 37 mer RNA fragment HIVNL4.3RNA2671-2707 resulted in energy release ranging from ΔH of ~ -30 kcal mol-1 to ~ -45kcal mol-1 (Fig 6C), and the level of energy release with Pr55Gag-TEV was 50% and 28% more than when Pr50GagΔp6-TEV and Pr55Gag WM-TEV were used, respectively (Fig 6C). In contrast, mutations of these identified A-containing RNA motifs within these authentic HIV RNA sequences eliminated most to all of its bindings with HIV Gag proteins (Fig 6C). The distinct level of energy release between Pr55Gag-TEV and Pr50GagΔp6-TEV based ITC experiments further support a potential role of p6 in the Gag-RNA and/or Gag-Gag interaction at the late stage of HIV assembly.
Overall, our ITC analyses provide direct evidence that the interaction between HIV Gag and RNA for Pr55Gag oligomerization is an energetically (ΔG<0) favourable reaction, and it is associated with a favourable enthalpy (ΔH) and unfavourable entropy (ΔS). There is a general binding preference of HIV Gag toward A-containing RNA during viral assembly. Unexpectedly, our data have revealed that the p6 domain within Pr55Gag also has a role in this Pr55Gag-A-containing RNA binding preference during virion biogenesis, which has raised a new question on a novel contribution of the p6 domain in the process of Gag-RNA interaction during viral assembly.
Our thermodynamic analyses showed that Pr55Gag-RNA binding was energetically favourable in terms of free energy exchange (ΔG<0) and that both Pr55Gag-Pr55Gag interactions and the type of RNA sequence can contribute to the favourable change in enthalpy. Many studies have focused on the interaction of nucleic acids with the mature forms of the NC domain to probe the chaperone activity [16, 33] and nucleic acid binding properties of NC during reverse transcription [34, 35]. Our ITC studies represent the first thermodynamic characterization of the binding interaction between nucleic acids and full length Pr55Gag, allowing us to study the effects of Pr55Gag-Pr55Gag interaction and Pr55Gag-RNA interaction on energy release. In the context of Pr55Gag assembly, these ITC analyses reflect the net energy exchange when nucleic acids are injected into the Pr55Gag-Pr55Gag and Pr55Gag-RNA complexes. It is important to emphasize that some of the observed energy exchange detected via ITC would be the consequence of Pr55Gag-Pr55Gag oligomerization. Our analyses between Pr55Gag and SL3 RNA suggested that at least 40% of energy release in our ITC experiment is derived from Pr55Gag oligomerization. Our ITC study with RNA and full length Pr55Gag (that is capable of forming higher-order Pr55Gag oligomers) suggest that the additional release of energy from Pr55Gag-Pr55Gag interactions could drive the binding of Pr55Gag with nucleic acid and the oligomerization of Pr55Gag. This interpretation is consistent with recent in vitro membrane-bound Pr55Gag assembly data showing that the genomic RNA selection by Pr55Gag and the self-assembly of Pr55Gag are interdependent [36].
Likewise, our observation that the Pr55Gag binding with immature HIV A-containing RNA motifs has a greater ΔH than Pr55Gag binding with the GU-containing cytosolic Pr55Gag-interacting RNA motifs during HIV biogenesis. These data suggest the additional heat release (that is associated with a greater entropic penalty) could help drive the Pr55Gag-complex to interact preferentially with A-containing RNA sequences during assembly. Furthermore, our thermodynamic analyses indicated that the oligomeric capable Pr55Gag has 3-times greater affinity for immature virus A-containing RNA motifs compared to cytosolic GU-containing RNA sequences, but that oligomerization-impaired form of Gag (Pr55Gag WM) does not. Similar to Pr55Gag WM, other oligomerization-impaired forms of Gag (Pr55Gag [CA Helix6], Pr55Gag [CA Helix 10] and Pr55Gag [CA All 4]) are consistent to have a less favorable [ΔH] in comparison to the wild type Pr55Gag and A-containing RNA ITC analyses. The nucleic acid sequence dependence of the binding enthalpy is likely related to the process of HIV assembly and virion genesis. Early work by Campbell and Rein [37] has shown that recombinant Pr50GagΔp6 has distinct behavior with different nucleic acids during in vitro assembly. Recent work by Kutluay et al have reported that HIV Pr55Gag would have a different preference towards distinct RNA sequence during virion genesis [21], although the precise mechanistic details of this relationship between HIV Pr55Gag and specific RNA sequences during viral assembly will require further investigation.
These data allow us to propose a model for how viral assembly steps are regulated. The unfavorable entropic reaction with A-containing RNA motif and an increased binding stoichiometry between Gag and A-containing RNA sequences are also in agreement with a much tighter packing of Pr55Gag molecules in the immature virus stage over low-order oligomeric Pr55Gag in the cytoplasm. It is known that the HIV-RNA genome has a bias towards A-containing codons [38, 39]. While many of these A-rich sequences might be a consequence of G to A hyper mutation from APOBEC3G pressure [40, 41], HIV has found ways to utilise them to its own gain. Indeed, these A-rich viral sequences are already known to be important for RNA trafficking from nucleus to cytoplasm via the RRE-Rev relationship [42, 43], and also for supporting the synthesis of viral cDNA during reverse transcription [44]. Now, we have evidence to suggest the A-rich RNA codon (in the form of identified A-containing RNA motifs) [21] may also have a role in regulating the viral assembly process thermodynamically. It is important to note that HIV Pr55Gag consistently displayed more favourable enthalpy [ΔH] when it binds to A-containing RNA motifs over GU-containing RNA motifs, and it is remains to be determined why HIV Pr55Gag has a binding preference with GU-containing RNA motifs in the cytoplasm and the plasma membrane [21]. One potential explanation could be that the context of how these GU-containing RNA motifs are being presented in HIV genomes (or cellular mRNA) is also an important determinant for Pr55Gag-RNA binding during HIV assembly, and further investigations are needed to dissect this process. Moreover, it would be important to highlight that the preference of Pr55Gag toward A-containing RNA is in part associated with the p6 domain. This unexpected observation suggests a previous unknown role of p6 in Pr55Gag-RNA interaction, and how p6 (and potentially in conjunction with ESCRT protein) might achieve this mechanistically would require further evaluation.
Taken together, our findings show how the virus might derive the energy required to drive the Pr55Gag-Pr55Gag assembly and the mechanism by which HIV-1 assembles. We propose a model (Fig 5C) wherein many Pr55Gag molecules bind to GU-containing RNA in the cytosol forming low-order oligomers. The Pr55Gag-RNA complex then traffics to the plasma membrane and acts as a nucleation site for Pr55Gag assembly; with energy released from Pr55Gag-Pr55Gag and Pr55Gag-RNA interactions driving the formation of higher order Pr55Gag oligomerization complexes. The high-order Pr55Gag complex in turn interacts preferentially with A-containing viral RNA sequences with a lower Kd, and the process is further enhanced by the additional release of heat through interaction with the A-containing viral RNA sequences to assist in the packaging of the genome, thus driving the completion of the particle formation.
Recombinant Gag proteins (Pr55Gag, Pr50Δp6Gag, Pr55Gag WM 316–7 AA, Pr55Gag [CA Helix 6 TTSTLQ 239–44 AASALA], Pr55Gag [CA Helix 10 D329A], Pr55Gag [CA All 4]) and processed NC proteins (p15NC-SP2-p6, p7NC) were expressed with C-term His-tag and purified as previously described [45]. Large scale production and purification of Pr55Gag is described herein.
Defined medium (DM1) used for seed cultures contained per litre: KH2PO4, 13.3 g; (NH4)2HPO4, 4.0 g; citric acid, 1.7 g; glucose, 10 g; MgSO4.7H2O, 0.62 g; kanamycin, 50 mg; thiamine hydrochloride, 4.4 mg; and trace salts solution, 5 mL. Defined medium (DM2) used in the bioreactors contained per litre: KH2PO4, 10.6 g; (NH4)2HPO4, 4.0 g; citric acid, 1.7 g; glucose, 25 g; MgSO4.7H2O, 1.23 g; kanamycin, 50 mg; thiamine hydrochloride, 4.4 mg; and trace salts solution, 5 mL. The trace salts solution contained per litre: CuSO4.5H2O, 2.0 g; NaCI, 0.08 g; MnSO4.H2O, 3.0 g; Na2MoO4.2H2O, 0.2 g; boric acid, 0.02 g; CoCl2.6H2O, 0.5 g; ZnCl2, 7.0 g; FeSO4.7H2O, 22.0 g; CaSO4.2H2O, 0.5 g and H2SO4, 1 mL. As required, glucose, magnesium, trace salts, thiamine and kanamycin were aseptically added as concentrated stock solutions to media after sterilisation.
Primary seed cultures were prepared from single colonies taken from a fresh transformation plate, and grown in 10 mL of DM1 (in a 30 mL bottle). The cultures were incubated at 37°C shaking at 200 rpm for 23 h. A volume (0.5 mL) of the primary seed culture was used to inoculate 500mL of DM1 (in a 2 L Erlenmeyer flask). These secondary seed cultures were incubated at 37°C shaking at 200 rpm for 16 h.
Recombinant HIV Gag proteins were produced in 2 L stirred tank bioreactors connected to a Biostat B (Sartorius Stedim, Germany) control system. The initial volume of medium in the bioreactor was 1.6 L and glucose as used as the carbon source. A volume of the secondary seed culture was added to the bioreactor to attain an initial optical density (measured at 600 nm) of 0.25. Foaming was controlled via the automatic addition of 10% (v/v) polypropylene glycol 2025; 3 mL of the antifoam solution (Sigma, Antifoam 204) was added prior to inoculation. The pH set-point was 7.0 and controlled by automatic addition of either 10% (v/v) H3PO4 or 10% (v/v) NH3 solution. The dissolved oxygen set-point was 30% of saturation and a two-step cascade control was used to maintain the dissolved oxygen above the specified set-point. The agitator speed ranged from 500 rpm to 1200 rpm and airflow (supplemented with 5% pure O2) ranged from 0.3 L min-1 to 1.5 L min-1. The ratio of air to oxygen was manually changed as required. To assist with correct folding of the Gag proteins, the medium was supplemented with 50 μM ZnSO4 added 1.8 h after inoculation. A high cell density fed-batch process was used, with the feed solution comprised of 400 mL of 660 g L-1 glucose solution to which 40 mL of 1 M MgSO4.7H2O was added. The feed flow rate was 21 mL hr-1 and commenced once the initial glucose supply was exhausted (typically 8 to 9 hr after inoculation). Two hours after the fed-batch process was initiated the bioreactor temperature set-point was reduced to 18°C, with culture temperature dropping to 19°C within 30 min. After cooling, protein expression was induced via the addition of 1 mM isopropyl-β-D-thiogalactopyranoside (IPTG) and 0.02% (w/v) arabinose and the feed flow rate reduced to 4 mL hr-1. Cells were harvested by centrifugation (12000 g, 4°C, 10 min) 22 to 23 h after inoculation and cell pellets stored at -80°C.
Cell pellets (~200 g) were thawed and re-suspended in 1.0 L of ice cold lysis buffer (1 M NaCl, 50 mM TRIS-HCl pH 8.0, 5 mM MgCl2, 10 mM imidazole, 1 (v/v)% Tween-20, 10% (v/v) glycerol, 5 mM DTT) containing 50,000 units of DNase I, 5 mM benzamidine-HCl and 1 mM phenyl methyl sulfonyl fluoride (PMSF). The cell suspension was homogenized (EmulsiFlex-C5 homogenizer, Avestin) pre-chilled to 4°C, three times at 700 bar pressure. The lysate was clarified by centrifugation 12,000 g, 4°C, 30 min), and the supernatant filtered using a 0.45 μm Stericup filter (Millipore).
The supernatant was loaded at 5 mL min-1 onto a 30 mL HisTRAP fast flow IMAC column (consisting of 6 x 5 mL cartridges connected in series; GE Healthcare) using a MINIPULS peristaltic pump (Gilson). The column had previously been equilibrated with 5 column volumes (CV) binding buffer (1.0 M NaCl, 50 mM TRIS-HCl pH 8.0, 5 mM MgCl2, 10 mM imidazole, 1% (v/v) Tween-20, 10% (v/v) glycerol, 5 mM DTT). The column was washed with 10 CV wash buffer (1.0 M NaCl, 50 mM TRIS-HCl pH 8.0, 5 mM MgCl2, 25 mM imidazole, 1% (v/v) Tween-20, 10% (v/v) glycerol, 5 mM DTT) and bound proteins eluted with 5 CV elution buffer (1.0 M NaCl, 50 mM TRIS-HCl pH 8.0, 5 mM MgCl2, 250 mM imidazole, 1% (v/v) Tween-20, 10% (v/v) glycerol, 5mM DTT). Pr55Gag eluting from the IMAC column was concentrated to ~ 3.0 mg mL-1 using a centrifugal concentrator (Amicon Ultra-15, 10,000 molecular weight cut-off membrane; Millipore).
The concentrated protein (5x 10 mL) was fractionated by size exclusion chromatography (SEC) using a Superdex 200 26/60 column (GE Healthcare) previously equilibrated in SEC buffer (50 mM TRIS-HCl, 1.0 M NaCl, 5 mM DTT, pH 8.0) using an ÄKTApurifier chromatography workstation (GE Healthcare). Peak fractions (UV 280nm) containing Pr55Gag were collected, pooled and concentrated to 1–2 mg mL-1 as described above, and snap frozen in liquid nitrogen before storage at -80°C.
Purified protein containing the TEV sequence is digested with 1:25 (w/w) TEV protease (produced in-house) at 4°C for 14 hrs. Efficiency of the cleavage is assessed by sodium dodecyl sulfate—poly acrylamide gel electrophoresis (SDS-PAGE). A 1ml NI-NTA column pre-equilibrated with SEC buffer and the TEV digested protein is applied to the column and the flowthrough is collected and passed over the column two more times. The cleaved His6-Tag and uncut fusion protein is eluted from the column using 5 CV elution buffer (1.0 M NaCl, 50 mM TRIS-HCl pH 8.0, 5 mM MgCl2, 250 mM imidazole, 1% (v/v) Tween-20, 10% (v/v) glycerol, 5mM DTT).
Protein was buffer exchanged into 1x Na/K 10mM phosphate buffer (pH 7.4) with different NaCl concentrations and concentrated to 1 mg mL-1. Yeast tRNA (Sigma Aldrich) (10% (w/w) ratio of nucleic acid to protein) was added to the protein solution and incubated for 30 min at room temperature, followed by addition of paraformaldehyde (PFA) (final concentration of 0.2% w/v). After incubating the solution for further 30 min at room temperature, the crosslinking was stopped by addition of 50 μL of 3M TRIS pH 8.0. Samples were centrifuged 10 min at 10,000 g and supernatants were analyzed by size exclusion chromatography using a Superdex 200 10/30 column (GE Healthcare), previously equilibrated in phosphate-buffered saline (PBS) with respective NaCl concentrations. Peak fractions were concentrated to 2 mg mL-1 as previously described. Fractions were electrophorised on a NuPAGE Novex 3–8% TRIS-Acetate protein gel under denaturing conditions before being transferred onto nitrocellulose membranes for Western analysis.
Pr55Gag and Pr55GagΔp6 were concentrated to 2.0 mg mL-1 in 50 mM TRIS pH 8.0 containing 1.0 M NaCl and 10 mM dithiothreitol and mixed with TG30 at a nucleic acid to protein ration of 4% (w/w) prior to dialysis against 50 mM TRIS pH 8.0 containing 150 mM NaCl and 10 mM dithiothreitol overnight at 4°C. Particles were imaged using both negative stain transmission electron microscopy and cryo-electron microscopy.
For negative stain imaging, the stock solution was diluted 100-fold to give a single layer of well-separated particles in most fields of view. Carbon-coated grids were glow discharged in nitrogen prior to use to facilitate sample spreading. Aliquots of approximately 4 μL were pipetted onto each grid and allowed to settle for 30 s. Excess sample was drawn off with filter paper, and the remaining sample stained with a drop of 2% aqueous phosphotungstic acid. Again, excess liquid was drawn off with filter paper. Grids were air dried until required. Samples were examined using a Tecnai 12 Transmission Electron Microscope (FEI, Eindhoven) at an operating voltage of 120kV. Images were recorded using a Megaview III CCD camera and AnalySIS camera control software (Olympus.)
For cryo-electron microscopy, virus particles were prepared, processed and imaged as previously described [46].
For the nucleic acid mediated in vitro assembly, protein (2.5 mg mL-1) and a DNA 30-mer oligonucleotide with alternating TG motifs (TG30; Macrogen) at a 10% (w/w) ratio of nucleic acid to protein were mixed in 50 mM TRIS, 500 mM NaCl, pH8.0 prior to adding inositol hexaphosphate (IP6) (10 μM) and slowly decreasing the NaCl concentration by dialysis into 50 mM TRIS, 150 mM NaCl, pH8.0 buffer to initiate the assembly process. The buoyant density of the particles produced and their in vitro assembly efficiency was assessed by layering the assembly reaction mixture onto a 32.5 to 55% (w/w) linear sucrose gradient in TBS (150 mM NaCl, 50 mM Tris, pH 7.6). Assembled HIV-1 Gag samples were layered onto the gradient and centrifuged for 16 h at 110,000 g (SW41 rotor: Optima L-90k ultracentrifuge; Beckman). 750 μL fractions were collected from the top in separate 1.5 mL tubes after completion of the run and prepared for trichloroacetic acid (TCA) precipitation.
350 μL of 50% (w/v) TCA was added to each fraction and incubated at 4°C for 30 min. Tubes were centrifuged at 14,000 g for 10 min at 4°C and the pellet was resuspended in 200 μL of prechilled acetone and incubated for 5 min at 4°C. The centrifugation step was repeated and excess supernatant aspirated and tubes air dried. The dried pellet was resuspended in 20 μL of 4x NuPAGE reduced LDS Sample Buffer (Life Technologies) and heated for 5 min at 95°C. Samples were briefly centrifuged and 10 μL of each sample was loaded onto a NuPAGE 4–12% Bis-TRIS Midi gel and electrophoresed using NuPAGE MES SDS Running Buffer (Life Technologies) at 150 volts.
PAGE gels were transferred to a Nitrocellulose membrane (PerkinElmer) using XCell II Blot Module (30 volts for 60 mins), and the membrane blocked overnight in blotto (5% (w/v) skim milk powder in TBS + 0.05% (v/v) Tween-20) at 4°C. Membrane was washed 3x with wash buffer (TBS + 0.05% (v/v) Tween-20) for 5 min each and probed with a mouse anti-CA monoclonal antibody (Hybridoma Clone 183-H12-5C; NIH AIDS Reagent program) for 1 h at room temperature. Membrane was washed 3x with wash buffer, incubated with a secondary goat anti-mouse IRDye 800CW conjugate (LI-COR; diluted1:30,000 in blotto) for 1 hr at room temperature. Membrane was washed again 3x and analysed using an ODYSSEY CLx system (LI-COR).
Isothermal titration calorimetry (ITC) experiments were performed at 30°C using a Microcal Auto-ITC200 MicroCalorimeter (Malvern). For each ITC experiment, the cell contained soluble Gag protein (6–10 μM) in TBS with 1mM tris (2-carboxyethyl) phosphine (TCEP) and the syringe contained 15–50 μM of the DNA 30-mer oligonucleotide with alternating TG motifs (TG30) or RNA 20-mer oligonucleotides with 4 consecutive repeating GU-containing (4x 5’-GAUGG-3’; 4x 5’-UGUGG-3’) or A-containing (4x 5’-GAGAA-3’; 4x 5’-AAGGA-3’) sequence motifs. The SL3 DNA and the 20mer SL3 RNA (5’GGACUAGCGGAGGCUAGUCC3’) as well as the HIVNL4.3RNA2671-2707 (5’-CTTAGAAATAGGGCAGCATAGAACAAAAATAGAGGAA-3’) and the HIVNL4.3RNA2671-2707 (with AGAAA mutation) (5’-CTATCTTTTAGGGCAGCAATCTTCAAAAATAGTCCTT-3’) are based on NL4.3 HIV proviral RNA. DNA and RNA oligonucleotides were purchased from Macrogen and Integrated DNA Technologies, respectively, and dissolved in TBS with 1mM TCEP. The volume of the first injection for each ITC run was set to 0.4 μL over 0.8 s to minimize the experimental impact caused by dilution effects at the injection syringe tip; this initial injection was excluded from data analysis. The first injection was followed by 25 injections of 1.5 μL over 3 s or 38 injections of 1 μL over 2 s each with the interval between each injection set to 300 s. The reference power was set to 5 μcals-1. The syringe stirring speed was set to 750 rpm.
A baseline was drawn by linear extrapolation using the data points collected from control experiments and subtracted from the whole data set to correct for the heat of dilution. The total heat signal from each injection was determined as the area under the individual peaks and plotted against the [nucleic acid]/[Gag] molar ratio. The corrected data were analyzed to determine number of binding sites (n), and molar change in enthalpy of binding (ΔH) in terms of a single site model derived as follows:
The quantity r is defined as the moles of nucleic acid [D] bound per mole of protein [P] with an association constant (Ka):
r=Ka[D]1+Ka[D]
(Eq 1)
Solving Eq 1 for Ka leads to:
Ka=r(1+r)[D]
(Eq 2)
Since the total concentration of nucleic acid [D]T in the cell is known, it can be represented by Eq 3 wherein [P]T equals the total protein concentration in the cell and n the number of binding sites:
[D]T=[D]+nr[P]T
(Eq 3)
Since Eq 3 shows nr[P]T = [PD], the fraction of sites occupied by the nucleic acid, combining Eqs 2 and 3 leads to:
r2−r[[D]Tn[P]T+1nKa[P]T+1]+[D]Tn[P]T=0
(Eq 4)
Solving the quadratic for the fractional occupancy (r) gives:
r=12[([D]Tn[P]T+1nKa[P]T+1)−([D]Tn[P]T+1nKa[P]T+1)2−4[D]Tn[P]T]
(Eq 5)
The total heat content (Q) of the solution in the volume of the sample cell (Vo; determined relative to zero for the apo-species) at fractional saturation r is given by Eq 6, where ΔH represents the molar heat of nucleic acid binding:
Q=nr[P]TΔHV0
(Eq 6)
Substituting Eq 5 into Eq 6 gives:
Q=nr[P]TΔHV02[([D]Tn[P]T+1nKa[P]T+1)−([D]Tn[P]T+1nKa[P]T+1)2−4[D]Tn[P]T]
(Eq 7)
The total heat content, Q can be calculated as function of n, Ka, ΔH because [P]T, [D]T and Vo are known experimental parameters. The parameter Q defined in Eq 7 only applies to the known starting volume of protein solution in the sample cell (Vo). In order to correct for the displaced volume (Vi), the change in heat content Q(i) at the end of the ith injection is defined by Eq 8 to obtain the best fit for n, Ka, and ΔH by standard Marquardt methods until no further significant improvement in fit occurs with continued iteration.
The Gibbs free energy (ΔG0) was calculated from the fundamental equation of thermodynamics Eq 9:
ΔG°=ΔH−TΔS=−RTlnKa
(Eq 9)
All data fitting operations were performed with Origin V7.0 software (OriginLab, Northampton, MA).
Following ITC analysis, the solutions containing HIV protein and RNA complex was spun at 100,000 g for 1 hr (TLA 100.2 rotor; optima max ultracentrifuge; Beckman). The supernatant was removed and the pellet was resuspended in 50μl of TBS. Protein estimation was done using UV-Vis (A280) on both the supernatant and the pelletable materials (NanoDrop1000; Thermo Scientific) via Bradford Protein Assay [47].
|
10.1371/journal.ppat.1006783 | Beta HPV38 oncoproteins act with a hit-and-run mechanism in ultraviolet radiation-induced skin carcinogenesis in mice | Cutaneous beta human papillomavirus (HPV) types are suspected to be involved, together with ultraviolet (UV) radiation, in the development of non-melanoma skin cancer (NMSC). Studies in in vitro and in vivo experimental models have highlighted the transforming properties of beta HPV E6 and E7 oncoproteins. However, epidemiological findings indicate that beta HPV types may be required only at an initial stage of carcinogenesis, and may become dispensable after full establishment of NMSC. Here, we further investigate the potential role of beta HPVs in NMSC using a Cre-loxP-based transgenic (Tg) mouse model that expresses beta HPV38 E6 and E7 oncogenes in the basal layer of the skin epidermis and is highly susceptible to UV-induced carcinogenesis. Using whole-exome sequencing, we show that, in contrast to WT animals, when exposed to chronic UV irradiation K14 HPV38 E6/E7 Tg mice accumulate a large number of UV-induced DNA mutations, which increase proportionally with the severity of the skin lesions. The mutation pattern detected in the Tg skin lesions closely resembles that detected in human NMSC, with the highest mutation rate in p53 and Notch genes. Using the Cre-lox recombination system, we observed that deletion of the viral oncogenes after development of UV-induced skin lesions did not affect the tumour growth. Together, these findings support the concept that beta HPV types act only at an initial stage of carcinogenesis, by potentiating the deleterious effects of UV radiation.
| Many epidemiological and biological findings support the hypothesis that beta HPV types cooperate with UV radiation in the induction of NMSC, the most common form of human cancer. We have previously shown that K14 HPV38 E6/E7 Tg mice, when exposed to long-term UV radiation, developed NMSC, whereas WT animals subjected to identical treatments did not develop any type of skin lesions. Here, we show that the high skin cancer susceptibility of these Tg animals tightly correlates with their tendency to accumulate UV-induced mutations in genes that are frequently mutated in human NMSC. Importantly, deletion of the HPV38 E6 and E7 genes in existing skin lesions did not affect the further growth of the cancer cells. Together, these findings support the model that beta HPV infection is a co-factor in skin carcinogenesis, facilitating the accumulation of the UV-induced DNA mutations.
| Non-melanoma skin cancer (NMSC) is the most common cancer in adult Caucasian populations [1]. The cutaneous human papillomavirus (HPV) types belonging to genus beta are suspected, together with ultraviolet (UV) radiation, to be involved in NMSC [2,3]. The first two beta HPV types, 5 and 8, were isolated from skin lesions of patients with a disorder called epidermodysplasia verruciformis (EV). EV patients are highly susceptible to beta HPV infection in the skin and develop cutaneous squamous cell carcinoma (cSCC) at anatomical sites exposed to sunlight [4]. The fact that organ transplant recipients, due to their immunosuppressed status, have an elevated risk of beta HPV infection and development of cSSC provided evidence for the role of beta HPV types in skin carcinogenesis also in non-EV individuals [5,6]. Finally, many epidemiological studies support the link between these viruses and cSCC in the general population [2,3,7]. These studies showed that, compared with the general population, patients with a history of cSCC are more frequently positive for viral DNA in the skin and/or for antibodies against the major capsid protein L1.
Molecular analysis showed that not all cancer cells contain a copy of the beta HPV genome and that the copy number of the beta HPV genome is higher in pre-malignant actinic keratosis (AK), a precursor lesion of SCC, than in SCC [8]. Thus, these data suggest that beta HPV types may act at an initial stage of skin carcinogenesis and that after full transformation of the infected cells, viral DNA can be lost. This model is consistent with the fact that additional carcinogens are involved in skin carcinogenesis. Considering that UV radiation is the key risk factor for cSSC development [9–11], the most plausible hypothesis is that beta HPV types exacerbate the accumulation of a large number of UV-induced somatic mutations, facilitating cellular transformation. Subsequently, the expression of the viral oncogenes may become irrelevant for the maintenance of the malignant phenotype.
Several studies in human keratinocytes, the natural host of beta HPV types, showed that E6 and E7 from some beta HPV types target key pathways linked to DNA repair, apoptosis, and cellular transformation [3]. Several transgenic (Tg) models for beta HPV have been generated [12–16], some of which have highlighted the synergism between viral oncogene expression in the skin epithelium and UV radiation in promoting cSCC [3]. Tg mice expressing beta HPV38 E6 and E7 in the basal layer of the epidermis under the control of the cytokeratin K14 promoter (K14) did not spontaneously develop any lesions during their life span. Upon long-term exposure to UV radiation (30 weeks), they developed first skin lesions closely resembling human AK and subsequently cSCCs. In contrast, wild-type (WT) mice developed neither pre-malignant lesions nor cSCCs when exposed to the same dose of UV radiation [15]. However, it is still unknown whether the high susceptibility of the K14 HPV38 E6/E7 Tg animals to UV-induced skin carcinogenesis is linked to the accumulation of mutations facilitated by the viral oncoproteins, which may become dispensable after cSCC development. In this study, we addressed this open question on the synergism between UV radiation and beta HPV38 E6 and E7 oncoproteins using the Tg mouse model. We showed that viral oncoproteins act at an initial stage of UV-induced skin carcinogenesis, facilitating the accumulation of a large number of somatic mutations in crucial genes that are associated with cSCC development in humans. In addition, silencing of the expression of the viral genes in established skin lesions does not affect further tumour growth.
We have previously shown that HPV38 E6/E7 expression in mouse skin strongly increases susceptibility to UV-induced carcinogenesis [15]. To evaluate whether the development of skin lesions present in K14 HPV38 E6/E7 Tg mice of chronic UV irradiation correlated with the number of accumulated DNA mutations, we used whole-exome sequencing of WT and Tg samples.
For this analysis, we selected normal skin from WT mice not exposed or exposed to UV radiation for 30 weeks (n = 2) and histologically confirmed skin specimens from three independent K14 HPV38 E6/E7 Tg mice UV-irradiated for 30 weeks, i.e., (i) normal skin, (ii) pre-malignant skin lesions and (iii) cSCC. For the pre-malignant lesions, the histological analyses revealed that they have the classic features observed in humans of the precancerous condition of AK, including slight atypia, parakeratosis, and acanthosis (S1 Fig) [15]. Exome sequencing (Illumina Hi-Seq) of collected samples generated an average coverage of 141.71× ± 11.9 (mean ± standard deviation).
The genomic sequence of the WT mouse not exposed to UV radiation was used as a control sample in paired analysis. Only 10 mutations were detected in the skin of the UV-irradiated WT mouse. Similarly, less than 10 mutations were detected in the Tg mouse not exposed to UV irradiation. In both cases, all the mutations were in genes not directly linked to carcinogenesis (S1 Table).
In UV-irradiated Tg animals, the mutational load varied across our cohort of well-differentiated cSCC exomes, averaging 3541 somatic variants (range, 3261–4027) or 68.58 ± 7.64 variants per Mb. The exome of the pre-malignant samples had substantially fewer variants, with an average of 1337 somatic variants (range, 937–2026) or 23.14 ± 14.70 variants per Mb. The exome of the chronically UV-exposed normal skin of Tg mice harboured an average of 15 somatic variants (range, 11–20) or 0.29 ± 0.08 variants per Mb (S2 Table). Thus, the number of somatic mutations was proportional to the severity of the skin lesion; the average number in SCCs was approximately double that in the pre-malignant lesions (Fig 1A).
The vast majority of the somatic mutations detected in SCCs were C:G > T:A mutations, mutations that are also prevalent in the UV-induced mutational signature (Fig 1B and 1C). We applied the non-negative matrix factorization (NMF) method to extract the mutational signatures composed of 96 single base substitution (SBS) types considering the sequence context (one base upstream and one base downstream) (S2 Fig). The extracted signature was compared with known mutational signatures by the cosine similarity method [17,18]. The value of the similarity obtained for the new B signature is 0.86 for COSMIC signature 27 (UV signature) (S2 Fig), indicating the clear prevalence of the impact of UV radiation on the etiology of these cSCCs.
To assess the biological significance of the somatic mutations detected in the skin lesions of the K14 HPV38 E6/E7 Tg mice, we determined whether they were detected in the previously compiled lists of epi-driver and epi-modifier genes [19–23], as well as genes identified in the Cancer Gene Census [24]. As shown in Fig 2, three classes of genes were found to be recurrently mutated in pre-malignant and malignant skin lesions of K14 HPV38 E6/E7 Tg animals, suggesting a selective process for the enrichment of mutations in these groups of genes.
Pathway analyses confirmed that the mutations detected in mouse cSCC affect key pathways intimately linked to cellular transformation (S3 Table).
A comparison of somatic mutations detected in our experimental Tg mouse model and in human cSCC [25] revealed that a large number of epi-driver, epi-modifier, and Cancer Gene Census genes were recurrently mutated in murine and human cSCC (Fig 3A).
A recent study identified the top human genes mutated in cSCC [26]. Interestingly, most of these genes are also found to be mutated in the UV-induced skin lesions of the K14 HPV38 E6/E7 Tg animals (Fig 3B). In agreement with previous findings on human cSCC [25], Trp53 showed up as the most mutated gene in the murine Tg-derived cSCC (Figs 2A and 3B). Here, p53 mutations appear to be an early event in skin carcinogenesis, because they were detected in one sample of normal skin as well as in all pre-malignant lesions and cSCCs. In agreement with our data, it was reported that p53 mutations can be detected in keratinocytes of UV-exposed normal skin [27,28]. However, all mutations were identified in the p53 DNA-binding domain (S4 Table), supporting their key role in the process of carcinogenesis. Consistent with the fact that in keratinocytes the Notch signalling pathway promotes cell-cycle exit and differentiation [29,30], NOTCH1 and NOTCH2 have been found to be mutated in human cSCC [25]. In our Tg mouse model, mutated NOTCH1 and/or NOTCH2 were also detected in all three cSCCs, but never in pre-malignant lesions (Fig 3B).
Our previous data showed that HPV 38 E6 and E7 expression in human keratinocytes resulted in accumulation of TAp53, which is recruited to the internal promoter located in intron 3 of p53 gene, with resulting transcriptional activation of ΔNp73α [31,32]. Fig 4 shows that also in the mouse skin, expression of the viral genes leads to increased ΔNp73α transcription. In contrast, in histologically confirmed pre-malignant and SCC lesions, p53 mutation correlates with a strong decrease in ΔNp73α mRNA levels (Fig 4).
In conclusion, our findings show that the expression of HPV38 E6 and E7 oncogenes in mouse skin increases susceptibility to UV-induced cSCC by facilitating the accumulation of somatic mutations that have been clearly associated with skin cancer development in humans.
Many studies support the role of beta HPV types, together with UV radiation, in the development of skin SCC [2,3]. However, in contrast to the mucosal high-risk HPV types such HPV16 that are required in all steps of cervical carcinogenesis, beta HPV types appear to have a role in the initial steps of carcinogenesis. To test this hypothesis, we constructed our K14 HPV38 E6/E7 Tg mice as a conditional expression model with two loxP elements, located immediately upstream and downstream of the viral genes [15]. Originally, we crossed the K14 HPV38 E6/E7 Tg mice with K14 Cre-ERT2 Tg animals overexpressing the Cre recombinase gene fused to a triple-mutant form of the human estrogen receptor that gains access to the nuclear compartment only after exposure to 4-hydroxytamoxifen (TMX) but not to the natural ligand 17β-estradiol, in order to silence E6/E7 expression by Cre-mediated deletion of the floxed viral genes at different times of the chronic UV irradiation, i.e., different stages of SCC development. Although the expression of the viral genes could be efficiently silenced upon administration of TMX to 5-week-old K14 Cre-ERT2 HPV38 E6/E7 compound mice, in the compound mice a strong decrease in viral gene expression was observed during the 30 weeks of UV irradiation in the absence of TXM treatment (S3 Fig). The loss of HPV38 E6 and E7 genes in long-term experiments was most likely due to a basal, non-specific Cre recombinase activity in the nucleus of mouse skin keratinocytes. None of the K14 Cre-ERT2 HPV38 E6/E7 Tg compound lines developed cSCC after 30 weeks of UV irradiation, further highlighting the importance of the viral proteins in UV-induced carcinogenesis.
Therefore, we developed a different strategy to evaluate the requirement of HPV38 E6 and E7 genes for cancer maintenance (Fig 5A). K14 HPV38 E6/E7 Tg mice were exposed to long-term UV irradiation, and after the appearance of well-defined skin lesions, after about 22–25 weeks of irradiation, two different DNA vectors were delivered by electroporation into the abnormal tissues. Because of the small size of the electroporated skin lesions, we could not perform any biopsy; therefore, we did not have any histological information about whether they correspond to pre-malignant or malignant lesions. Results obtained in several independent experiments showed that the lesions that occurred after 22–25 weeks of UV irradiation correspond to pre-malignant lesions or an early stage of cSCC [15,16]. Both vectors contain a scaffold/matrix attachment region (S/MAR) that keeps the plasmid in an episomal state, avoiding any integration-mediated toxicity, and ensures robust and persistent gene expression [33]. The vector codes for luciferase and Cre recombinase genes (Cre-Luc) separated by the P2A cleavage site, whereas the control vector expresses only a luciferase gene (Luc). Luciferase was used to monitor the efficiency of transfection by non-invasive in vivo imaging, and Cre was used to induce the excision of the viral genes. A total of 23 lesions on 14 mice were transfected either with the Luc vector (n = 9) or with the Cre-Luc vector (n = 14). When possible, the same mouse was injected with both vectors, each on a different lesion. Three representative mice are shown in Fig 5B. Luciferase activity was detected in the animals’ skin in each of the electroporated areas independently of the vector type.
After electroporation, the animals were irradiated until the end of the 30-week UV irradiation protocol and closely monitored for several weeks to evaluate the progression of the skin lesions. No significant difference in tumour growth was observed in animals transfected with the Luc or Cre-Luc vectors (Fig 6A). Histological analyses confirmed that 100% percent of the Luc-injected lesions and 93% of the Cre-Luc injected lesions (13 out of 14) evolved into invasive cSCC; a morphological examination revealed no major differences between the two groups of tumours (Fig 6B). Detection of the viral RNA transcripts by RNA-RNA in situ hybridization confirmed that electroporation of skin lesions with the Cre-Luc vector, but not with the Luc vector, resulted in the loss of E6/E7 expression in large islands of cancer tissue (Fig 6B).
In conclusion, our findings show that after the accumulation of UV-induced DNA mutations and the development of skin lesions, the expression of the HPV38 E6/E7 genes is dispensable for the maintenance of the malignant phenotype of skin cancer cells.
Although the HPV family includes more than 200 types, to date only the mucosal high-risk (HR) HPV types have been clearly associated with human carcinogenesis. These viruses are the etiological agents of cervical cancers as well as a subset of other genital and oropharyngeal cancers [34]. Beta HPV types have been proposed to be associated with cSCC. They were initially linked to cSCC in EV patients, but now many epidemiological and biological studies support the role of beta HPV types in skin carcinogenesis also in non-EV individuals [3].
We have previously shown in a Tg mouse model that expression of beta HPV38 E6 and E7 in the skin strongly increases the risk of cSCC development upon UV irradiation [15]. Here, we showed that the higher susceptibility of K14 HPV38 E6/E7 Tg mice to UV-induced skin carcinogenesis tightly correlates with the accumulation of a high number of mutations in the keratinocyte genome. Remarkably, exposure of WT animals to the same doses of UV radiation did not lead to accumulation of DNA mutations and development of cSCC. These data suggest that the HPV38 oncoproteins can negatively affect the DNA repair machinery and/or immune pathways that lead to the elimination of damaged cells. We have recently shown that K14 HPV38 E6/E7 Tg mice are hampered in the production of interleukin 18 (IL-18) during their exposure to UV radiation [16]. Upon UV irradiation and activation of the inflammasome, keratinocytes secrete high levels of cytokines from the IL-1 family, including IL-18, thus inducing a broad spectrum of processes, such as infiltration and activation of inflammatory leukocytes, immunosuppression, DNA repair, and apoptosis [35–38]. Thus, it is likely that the high susceptibility to UV-induced DNA mutations and skin carcinogenesis of K14 HPV38 E6/E7 Tg mice may be linked to the negative impact of HPV38 on IL-18 production.
Analysis of the mutational profile revealed that a large number of genes encoding for epi-drivers or epi-modifiers and proteins known to be associated with carcinogenesis (Cancer Gene Census) harbour missense or nonsense mutations. Most importantly, the gene mutation profile found in murine cSCC shows remarkable similarities to the mutational profile found in human cSCC. In particular, mutations in p53 appear to be an early event in murine and human skin carcinogenesis. We have previously shown that beta HPV38 E7 alters the p53/73 network by inducing accumulation of p53/p73 antagonist ΔNp73α [31,32]. In human keratinocytes expressing beta HPV38 E6 and E7, ΔNp73α forms a transcriptional inhibitory complex, which binds a subset of p53-regulated promoters, preventing their activation in the presence of cellular stress [39]. Because the major role of p53 is to safeguard genome integrity, the high cancer susceptibility of K14 HPV38 E6/E7 Tg mice along with the high numbers of accumulated UV-induced DNA mutations can be explained, at least in part, by the properties of the beta HPV oncoproteins. However, once p53, and likely other cellular genes, are irreversibly inactivated by DNA mutations induced by UV radiation, the progression and maintenance of the skin carcinogenic process could become independent of the expression of viral genes. In agreement with this view, ΔNp73α mRNA levels decrease strongly in UV-induced skin lesions of K14 HPV38 E6/E7 Tg animals after accumulation of p53 mutations. In addition, we observed that the deletion of the HPV38 E6 and E7 genes does not affect further growth of the tumour. In contrast, in K14 Cre-ERT2 HPV38 E6/E7 Tg the loss of the viral genes at early stages of the irradiation protocol prevents the development of UV-induced skin lesions, underlining the key function of HPV38 E6 and E7 in UV-mediated carcinogenesis.
These findings in the K14 HPV38 E6/E7 Tg mouse model are in agreement with the studies on human skin lesions, supporting an early role of beta HPV types in skin carcinogenesis. Indeed, the copy numbers of the beta HPV genome appear to be higher in the pre-malignant lesion, AK, than in cSCC [8]. In addition, not all cancer cells contain a copy of a beta HPV genome [8]. Thus, the mechanisms of carcinogenesis induced by beta HPV types appear to be substantially different from those of the mucosal HR HPV types. In the case of the mucosal HR HPV types, the viral oncoproteins are the major drivers of cancer development (e.g. in the cervix) that, in addition, are required throughout the entire carcinogenic process (Fig 7). In contrast, UV-induced damage is the main carcinogen of cSCC. Here, however, beta HPV oncoproteins can facilitate the accumulation of UV-induced DNA damage but they are dispensable after full development of a malignant lesion (Fig 7).
Why do different HPV types display different biological properties? Cutaneous and mucosal HPV types infect cells at distinct anatomical sites exposed to different environmental stresses. Thus, it is not surprising that they have evolved with divergent biological properties. All HPV types rely on the DNA replication machinery of the host cell. Therefore, they must have developed several mechanisms to maintain the infected cell in a proliferative state to guarantee efficient viral genome replication. Exposure of skin keratinocytes to UV radiation leads to accumulation of DNA damage, which in turn induces cell-cycle arrest or apoptosis to allow repair or elimination, respectively, of the damaged cell. The cutaneous HPV types appear to be able to circumvent this adverse effect of UV radiation on keratinocyte proliferation, promoting the accumulation of damaged cells in the skin and, consequently, carcinogenesis.
Our previous findings showed that different HPV38 E6/E7 expression levels in independent Tg lines influence the rate of SCC development [15]. Thus, it plausible to hypothesize that also in humans, the viral gene expression levels may have an impact on UV-induced skin carcinogenesis. Limited data are available on beta E6 and E7 gene expression in normal skin and pre-malignant and malignant skin lesions (reviewed in [2,3]). There is no information on the different spliced forms of beta HPV genes and how they could determine a different efficiency in protein synthesis. Thus, additional studies are required in humans to corroborate the findings obtained in the Tg mouse model on the hit-and-run mechanism of HPV38 in UV-induced carcinogenesis.
In conclusion, our findings in a Tg mouse model highlight a novel mechanism of infection-associated carcinogenesis, in which the virus is not the driving force but synergizes with UV radiation in promoting cSCC.
The transgenic animal model FVB/NTgN(38E6E7)187DKFZ (https://mito.dkfz.de/mito/Animal%20line/10954) has been previously described [15]. UVB irradiation was performed under sevoflurane anaesthesia, and every effort was made to minimize suffering.
The animal facility of the German Cancer Research Center has been officially approved by responsible authority (Regional Council of Karlsruhe, Schlossplatz 4–6, 76131 Karlsruhe, Germany), official approval file number 35–9185.64. Housing conditions are thus in accordance with the German Animal Welfare Act (TierSchG) and EU Directive 425 2010/63/EU. Regular inspections of the facility are conducted by the Veterinary Authority of Heidelberg (Bergheimer Str. 69, 69115 Heidelberg, Germany). All experiments were in accordance with the institutional guidelines (designated veterinarian according to article 25 of Directive 2010/63/EU and Animal-Welfare Body according to article 27 of Directive 2010/63/EU) and were officially approved by Regional Council of Karlsruhe (File No 35–9185.81/G-64/13 and 35–9185.81/G-200/15).
To generate the Luc and the Luc-Cre vectors, the pS/MARt-GFP DNA vector was first digested with the restriction enzymes NheI and BglII to linearize the vector and eliminate the transgene GFP. The InFusion system provided by Clonetech was used to introduce the luciferase gene alone or in combination with the Cre recombinase gene to generate the vector pS/MARt-Luc or the vector pS/MARt-Luc-P2A-Cre, respectively.
UVB irradiation was performed with a Bio-Spectra system (Vilber Lourmat, Marne La Vallee, France) at a wavelength of 312 nm as previously described [15]. Briefly, animals were anesthetized with 3% Sevorane (Abbott, Wiesbaden, Germany) in an inhalation anesthetizer (Provet, Lyssach, Switzerland) and placed in a covered compartment with an upper square opening (3×2 cm) at a distance of 40 cm from the UVB lamp.
To study UV-induced carcinogenesis, 7-week-old female FVB/N WT or K14 HPV38 E6/E7 Tg animals were shaved on the dorsal skin with electric clippers and irradiated 3 times a week for 10 weeks with increasing doses of UVB, starting from 120 mJ/cm2 to a final dose of 450 mJ/cm2, with a constant weekly increase to allow skin thickening. For the following 20 weeks, mice were irradiated 3 times a week with 450 mJ/cm2. The UV irradiation protocol was based on the data described in [40] and to mimic the situation in humans. For instance, the maximum dose of the UV irradiation protocol, 450 mJ/cm2, corresponds to 50 minutes of solar exposure in July in Paris. The tumour incidence (tumour bearers/group) was recorded weekly. Tumours were identified first macroscopically and by histological diagnosis. After 30 weeks, or earlier if the tumour reached the ethically allowed maximal size, the animals were sacrificed and H&E-stained sections of dorsal skin were used for histological diagnosis.
To study the effect of the loss of the viral genes on skin cancer development, 7-week-old K14 HPV38 E6/E7 Tg mice (n = 14) were shaved on the dorsal skin and treated for 30 weeks with increasing doses of UVB as previously described [15]. As soon as skin lesions (maximum diameter 2.6 mm) became evident, 46 μg of pS/MARt-Luc or 50 μg of pS/MARt-Luc-P2A-Cre dissolved in isotonic saline solution was injected directly into the lesions. To facilitate the uptake of the injected DNA, an electric field was applied to the area of the injection site using a Tweezertrodes connected to a BTX ECM 630 generator (Harvard Apparatus, Holliston, MA, USA). A first high-voltage electric pulse (1400 V/cm, 100 μs, 2 times), to induce temporary gaps in the keratinocytes cell membrane, was followed by a low-voltage electric field (140 V/cm, 400 ms, 2 times), to facilitate the migration of the DNA into the cells. At 72 h after the DNA injection, the mice were injected intraperitoneally with 150 mg/kg of luciferin in sterile water, and the luciferase activity was then assessed using an IVIS Lumina III imaging system (Perkin Elmer, Rodgau, Germany). When possible, a single mouse received both plasmids at the same time, each on a different lesion. The UV irradiation continued until week 30, according to the protocol [15]. The lesions were then closely monitored and the animals were sacrificed in accordance with an ethical protocol to avoid animal suffering. Skin lesions were collected for histological examination and detection for HPV38 E6/E7 RNA by in situ hybridization.
Total RNA was isolated from dorsal skin of WT (n = 4) or K14 HPV38 E6/E7 Tg animals (n = 5) as well as histologically confirmed pre-malignant (pre-m) and SCC from three independent mice. cDNA was synthesized from 1 μg of total RNA using M-MLV reverse transcriptase (Invitrogen, Darmstadt, Germany), and a mix of random hexamers were used as primers. Quantitative reverse transcription PCR (RT-qPCR) was performed in a 20 μl mixture containing 1 μl of 1:10 diluted cDNA and Mesa green quantitative PCR (qPCR) Master Mix (Eurogentec, Angers, France) with specific mouse ΔNp73α primers (5′-GCCAAAAGGGTCATCATC-3′ and 5′-TGCCAGTGAGCTTCCCGTTC-3′) or mouse GAPDH primers to amplify a housekeeping gene as internal control (5′-GTGACCCCATGAGACACCTC-3′ and 5′-GTATGTCCAGGTGGCCGAC–3′), using an Applied Biosystems 7300 machine (Applied Biosystems, Darmstadt, Germany). The fluorescence threshold value was calculated using the SDS analysis software from Applied Biosystems.
Once the tumours reached the maximum ethically allowed size, the mice were killed and the lesions isolated. Half of the lesion was embedded in OCT medium and slowly cooled down to −80°C. Sense and antisense riboprobes were generated from linearized plasmid DNA containing full-length HPV38E6E7 cDNA using the Digoxigenin RNA labelling Mix from Roche. RNA-RNA in situ hybridization was performed as previously described[41]. In brief, serial 5 μm cryo-sections were mounted on Superfrost Plus slides (Thermo Scientific), fixed in 4% paraformaldehyde in 2× SSPE, digested with proteinase K (0.5 μg/ml), and pre-hybridized at 42°C for 2–4 h. Hybridization was performed overnight at 42°C in 50% formamide, 2× SSPE, 10% dextran sulfate, 10 mM Tris-HCl pH 7.5, 1× Denhardt’s solution, 500 μg/ml tRNA, 100 μg/ml herring sperm DNA, 0.1% SDS, and 10 μg/ml DIG-labelled riboprobe. After hybridization, slides were washed once in 50% formamide, 2× SSPE; 0.1% SDS for 30 min at 50°C, treated with RNaseA (50 μg/ml in 2× SSC, 0.1% SDS), and washed again in 50% formamide, 0.5× SSPE, 0.1% SDS for 30 min at 37°C. Hybridization signals were visualized using Biotin Tyramide (TSA Biotin System, PerkinElmer) according to the manufacturer’s protocol.
Tumour growth values of lesions injected with the pS/MARt-Luc or pS/MARt-Luc-P2A-Cre vector were compared with the two-sample t-test. The statistical analysis was performed with GraphPad Prism (version 6, GraphPad Software Inc., La Jolla, CA, USA).
The quality of the raw reads was estimated with FastQC software (version 0.11.5, http://www.bioinformatics.babraham.ac.uk/projects/fastqc/). Reads were mapped to the GRCm38 Mouse reference genome (ftp://hgdownload.cse.ucsc.edu/goldenPath/mm10/) using Burrows-Wheeler Aligner (BWA, http://bio-bwa.sourceforge.net/) version 0.7.15 and producing a BAM file. The following GATK Best Practice Recommendations were applied to the BAM files to improve variant detection quality. Picard (version 2.4.1, https://broadinstitute.github.io/picard/) SortSAM was used to sort and index BAM files, and the AddOrReplaceReadGroups tool was used to replace all read groups with a single new read group. The duplicate reads were marked with the MarkDuplicates tool from Picard, and the newly produced BAM file was indexed with the BuildBamIndex tool. GATK (version 3.6.0, https://software.broadinstitute.org/gatk/download/) RealignerTargetCreator was used to determine the position concerned by local realignment, and IndelRealigner was used to perform local realignment around these sites. The GATK BaseRecalibrator tool was used to detect systematic errors in base quality scores. Dbsnp and dbindel (version 142) for the mm10 reference genome was downloaded from the Sanger website (ftp://ftp-mouse.sanger.ac.uk/REL-1505-SNPs_Indels/) and considered as input. Lastly, the index of the output BAM file was created with Picard BuildBamIndex, and GATK PrintReads was used to write out sequence read data.
The quality of the alignment was estimated with Qualimap (version 2.0.2, http://qualimap.bioinfo.cipf.es/). Then the variant calling was done with Mutect (version 1.1.7, http://archive.broadinstitute.org/cancer/cga/mutect), by using a skin sample from a WT mouse not exposed to UV as the “normal sample” for paired analysis. Only somatic mutations passing Mutect internal filters were considered for the analysis. The VCF files are annotated with Annovar by using the MutSpec Annot Tool in Galaxy [42]. Variants were then filtered based on SegDup databases from UCSC (version from 4 May 2014, http://hgdownload.cse.ucsc.edu/goldenPath/mm10/database/genomicSuperDups.txt.gz), as well as Tandem Repeat and Repeat Masker (version from 9 February 2012, http://hgdownload.soe.ucsc.edu/goldenPath/mm10/bigZips/). House-made scripts were then used to keep only SNPs that have a functional impact and fall in exonic or splicing regions. Non-negative matrix factorization mutational signatures were inferred with MutSpec-NMF tools, as previously reported.
The pathway analysis was performed using the EnrichR web application (http://amp.pharm.mssm.edu/Enrichr/; citations*2). The input gene list was made by merging the mutations detected in the pre-malignant lesions (n = 3) or cSCCs (n = 3) of the K14 HPV38 E6/E7 Tg animals. The analysis included only genes harbouring mutations that are likely to alter the biological properties of the encoded products, i.e., 3111 genes in the pre-malignant lesions and 6372 genes in the cSCCs. The gene lists were then loaded into the EnrichR software, and the result from the KEGG database (version 2016) was considered. Only pathways with a significant adjusted p-value are shown in S1 Table. The list of pathways is ranked by combined score (combined score is computed by taking the log of the p-value from the Fisher exact test and multiplying it by the z-score of the deviation from the expected rank).
The list of epigenetic driver and modifier genes was constructed on the basis of genes reported in different publications [19–23]. The Cancer Gene Census list was downloaded from the COSMIC website (12 November 2016, http://cancer.sanger.ac.uk/census) and is based on a previous publication [24].
The comparison of the mouse data with the human data [25,26] was done with Bioconductor (release 3.4, https://www.bioconductor.org/) in R (version 3.3.2, “Sincere Pumpkin Patch”). The module BioMart[43,44], version 2.3 enables the conversion of nearly 87.86% of human gene names from the Chitsazzadeh et al. publication [26] to their corresponding mouse gene names.
|
10.1371/journal.pcbi.1002509 | Dynamic Prestress in a Globular Protein | A protein at equilibrium is commonly thought of as a fully relaxed structure, with the intra-molecular interactions showing fluctuations around their energy minimum. In contrast, here we find direct evidence for a protein as a molecular tensegrity structure, comprising a balance of tensed and compressed interactions, a concept that has been put forward for macroscopic structures. We quantified the distribution of inter-residue prestress in ubiquitin and immunoglobulin from all-atom molecular dynamics simulations. The network of highly fluctuating yet significant inter-residue forces in proteins is a consequence of the intrinsic frustration of a protein when sampling its rugged energy landscape. In beta sheets, this balance of forces is found to compress the intra-strand hydrogen bonds. We estimate that the observed magnitude of this pre-compression is enough to induce significant changes in the hydrogen bond lifetimes; thus, prestress, which can be as high as a few 100 pN, can be considered a key factor in determining the unfolding kinetics and pathway of proteins under force. Strong pre-tension in certain salt bridges on the other hand is connected to the thermodynamic stability of ubiquitin. Effective force profiles between some side-chains reveal the signature of multiple, distinct conformational states, and such static disorder could be one factor explaining the growing body of experiments revealing non-exponential unfolding kinetics of proteins. The design of prestress distributions in engineering proteins promises to be a new tool for tailoring the mechanical properties of made-to-order nanomaterials.
| A tensegrity structure is one composed of members that are permanently under either tension or compression, and the balance of these tensile and compressive forces provides the structure with its mechanical stability. Macroscale tensegrity structures, which include Buckminster Fuller's geodesic domes, achieve exceptional structural integrity with a minimal use of resources. The question we address in this work is whether nature makes use of molecular-scale tensegrity in the design of proteins. Using Molecular Dynamics simulations of the protein ubiquitin, we measure the network of pairwise forces connecting the amino acid residues and show that this network does indeed have the character of a tensegrity structure. Furthermore, we find that the arrangement of tensile and compressive forces is such that hydrogen bonds in the protein's beta sheet, which are crucial for bearing mechanical loads, are compressed. This pre-compression is enough to significantly lengthen the lifetime of a bond under a given force, and thus should be an important factor in determining the protein's mechanical strength. The rational design of molecular prestress networks promises to be a new avenue for the engineering of proteins with made-to-order mechanical properties, for applications in medicine, materials and nanotechnology.
| The principle of ‘minimal frustration’ [1], [2] underlies the thermodynamic picture of protein folding. According to this picture, proteins negotiate a rough, funnel-shaped energy landscape during the folding process, and eventually settle in a state that, as much as possible, satisfies the energetic constraints arising from the multitude of interatomic covalent, electrostatic and van der Waals interactions. Although frustration is minimised in the native state, it is not completely eradicated. Even in the simplest crystals, the equilibrium state is one that minimises the energy of the structure as a whole, not every atom-atom interaction individually; global constraints prevent every pairwise interactions from being perfectly satisfied. This is even more the case for proteins, in which the topological contraints of the backbone peptide bonds further restrict the freedom of individual atoms to individually satisfy every interaction.
Such local frustration in a protein must give rise to residual mechanical forces – thus, proteins are in some sense prestressed materials. D. Ingber has proposed that proteins and other biological structures should be understood in light of the architectural concept of tensegrity [3], [4], popularised by Buckminster Fuller, describing structures the mechanical stability of which arises purely from a balance between pre-tensed and pre-compressed members. The concept of biomolecular tensegrity has come under focus very recently in the work of Shih, Ingber, and co-workers, who have designed and synthesised prestressed DNA structures [5]; it has also been invoked in a novel method for interpreting free energy profiles inferred from the forced unfolding of single biomolecules [6]. In contrast to this tensegrity picture, classic coarse-grained models of proteins, which have been used extensively to study protein folding and dynamics, typically neglect prestress. Both G-style models [7]–[9] and elastic network models [10], [11] define the equilibrium separation of every residue pair to be precisely the separation measured in the native state, and thus every residue-residue interaction is individually relaxed; as such, the native state is defined to contain no residual force. Thus, especially in research areas that rely on these coarse-grained approaches, the consequences of prestress for folding and dynamics have not been well explored.
It has been demonstrated recently [12] that, in graphene sheets, the prestress of bonds around defects at grain boundaries is the key determining factor for toughness of the sheets. This result highlights the fact that the existence of prestress can qualitatively change the mechanical properties of a structure, and raises the question of to what extent such effects are utilised by nature to tune the mechanical stability of proteins. Since concentrations of internal force in a molecule could be used to drive conformational changes if released, via thermal fluctuations or due to interactions with other molecules, the spatial distribution of prestress in a protein may also provide important clues for understanding the mechanisms for protein-protein interactions [13], [14], protein-DNA interactions [15], and allostery; indeed, the existence of ‘tensed’ and ‘relaxed’ states in allosteric proteins has been a central concept in models of allosteric transitions since the classic early models of inter-domain allostery in hemoglobin [16]–[18]. Elastic stress also plays a central role in the more recent allosteric model of Savir and Tlusty [19]. Using short lengths of pre-tensed double-stranded DNA to stretch individual molecules, it is now possible to directly observe the role played by elastic stress in the allosteric control of protein enzymes [20], [21] and ribozyme [22].
But how is such a global elastic stress built up within a protein scaffold? We used all-atom Molecular Dynamics (MD) simulations to quantify the importance of prestress in the native state of ubiquitin. To this end, we adapted an earlier technique for measuring force distributions in mechanically perturbed proteins [23], [24] to allow the calculation of effective pairwise residue-residue force profiles. This procedure is a direct force measurement, unlike other methods based on inferring pairwise forces from fluctuations [25], [26], and does not require any a priori assumptions about the form of the force profile. From the effective force profiles we extracted average forces for each residue pair, thereby constructing a prestress network for the protein. We found that high residual forces exist throughout the protein, and are particularly associated with hydrogen bonds and salt bridges. The magnitude of these forces is shown to be enough to significantly influence the protein's mechanical properties, most notably its unfolding pathway. We also discover that, for some side-chains, prestress is dynamic – inter-residue mechancial coupling switches between a number of distinct regimes depending on side-chain conformations.
From 100 ns of MD trajectories, we calcuated effective force profiles for every pair of amino acid residues in ubiquitin (Fig. 1a), as described in the Methods section. The average forces inferred from these profiles are plotted in Fig. 1b superimposed on the 3D structure of the protein. For clarity, the same force network is also represented in Fig. 1c as a circular graph, with each vertex corresponding to a residue. Covalently bonded residue pairs are neglected [see the Supplementary Material (Text S1 and Figs. S6 and S7) for details on covalent bond forces]. Red (blue) edges represent attractive (repulsive) forces, and edge thicknesses correspond to the magnitude of the forces, which range between −490 pN (attractive) and +407 pN (repulsive). In the context of cell biology these are high forces – for comparison, the forces generated by the kinesin walk have been measured to be on the order of 2 pN [27]. An animation showing the projection of this network on the three-dimensional structure of the protein is provided in the Supplementary Material (Video S1).
The most obvious large-scale structures in the network are the relatively ordered bands of both tensile and compressive forces that connect neighboring beta strands: specifically, the two parallel pairs of beta strands 1/2 and 3/5, and the anti-parallel beta strand pair 1/5. In contrast, isolated cases of strongly tensed (red) residue pairs are also observed, such as Lys27-Asp52 and Lys11-Glu34, which do not correlate with neighboring residues. These high tensile forces occur only between residues with charged side-chains; as discussed in more detail below, they correspond to tensed salt bridges.
To get a clearer picture of the prestress pattern associated with the main-chain interactions in beta sheets, the force network accounting for only inter–main-chain interactions is shown in Fig. 2a and Video S2 [here ‘main-chain’ refers to the N, C, C, O, and H atoms making up the backbone]. Inter–main-chain interactions are found to be predominantly attractive, with a few strongly repulsive pairs. To better understand this phenomenon we examine in more detail the residue-wise force distribution in beta strands 1 and 5. These strands are of special relevance to the mechanical stability of ubiquitin, since they form a ‘force clamp’ that provides the primary resistance against rupturing of the protein by stretching from the N and C termini [28]. In Fig. 2b the residue-wise average main-chain forces within the beta force clamp are illustrated. Forces between neighboring, covalently-bonded residues are not shown, and will be discussed separately. There are five hydrogen bonds between these beta strands, formed by residues Gln2 and Glu64, Phe4 and Ser65, Phe4 and Leu67, Lys6 and Leu67, and Lys6 and Leu69; and it is evident that these pairs are precisely those for which the average pairwise force is repulsive (blue).
Apart from the hydrogen-bonded pairs, every other residue pair in beta strand pair 1/5 experiences an average attractive force (red lines in Fig. 2a); they are all pre-tensed. This gives the beta sheet an overall appearance reminiscent of a tensegrity structure, the mechanical stability of which is determined by a balance between tensed and compressed structural members [3]. The origin of the pre-compression of the hydrogen bonds can be understood via this tensegrity analogy: the ‘tensed’ attractive interactions between the two beta strands act to pull the strands closer together than they would otherwise like, compressing the hydrogen bonds until the tensile and compressive forces balance. The same pattern, of hydrogen bonds compressed by other attractive cross-strand interactions, also holds for the other beta strand pairs in the protein, both parallel and anti-parallel; see Fig. S1 for the force distributions in the anti-parallel beta strand pairs 1/2 and 3/5. The underlying atomic forces that give rise to the attractive and repulsive residue-residue forces are illustrated in Fig. S2.
To investigate how the combination of atomic forces gives rise to an effective force profile for each residue pair, we plot the distribution of residue-wise force versus separation of the C atoms. Fig. 2c shows the result of this procedure for the hydrogen-bonded residue pair Gln2-Glu64. Each point in the figure corresponds to a single frame of the trajectory. The scatter of the data points is large, due to fluctuations in the conformations and relative orientations of the two residues. The average fit (blue curve) represents an ‘effective’ pairwise force profile averaging over these fluctuations [29]. Around the mean separation, the effective force profile is approximately linear, and thus has the character of a compressed Hookean spring. But the curve is clearly non-linear at larger separations, approaching the rupture distance of the bond. The overall shape is reminiscent of a Morse-type potential traditionally used to approximate chemical bonds. Similar profiles are obtained for the other hydrogen bonds in the sheet.
Fig. 2d shows the effective force profile for one of the ‘tensed’ non-hydrogen-bonded pairs (Ile3-Ser65). The magnitude of the attractive force is found to reduce with separation. Such behaviour cannot be approximated by a physical Hookean spring, since the local effective spring constant is negative; it is instead more like a Morse-type potential where the interacting pair only samples the tail of the potential, never even approaching the equilibrium separation. Thus the analogy with macroscopic tensegrity structures is only superficial: it is not accurate to think of the tensed residue pairs as prestressed cables, which would exhibit Hookean behavior. Due to the partially non-Hookean springs in the network of ubiquitin, the prestress can be expected to have an impact on both the elastic behavior of the protein (if any) as well as the inelastic behavior including rupture.
Although the alpha helix does not play a direct role in determining the mechanical stability of the protein, it is interesting to look at the pattern of prestress in the helix and see whether pre-compression of hydrogen bonds is a general phenomenon or one restricted to beta sheets. Fig. S3 shows the main-chain-only residue-wise forces within the helix. Similar to the beta sheets, the helix exhibits a tensegrity-like pattern of balancing compressive and tensile forces. However, in this case the hydrogen bonded residue pairs are under tension, in contrast to the compressed beta-sheet hydrogen bonds. We conclude that pre-compression of hydrogen bonds is not a property intrinsic to all hydrogen bonds, but rather a context-dependent phenomenon: prestress in a given bond is determined by the interactions between other residues in its immediate neighborhood, and the local molecular geometry. This points to the fascinating possibility that the distribution of prestress in a protein can be engineered by intelligent modifications to the amino acid sequence, providing a new tool for designing proteins with made-to-order mechanical properties [30].
Any applied external force must work against the inherent compression imposed by the protein onto the rupturing bonds. We propose that the hydrogen bond compression influences the unfolding force and pathway of the force clamp between beta strands 1 and 5. Fig. 3 is a plot of the average force for each of the five bonds in this clamp. Of the two bonds at the edge of the sheet, pair Gln2-Glu64 () is significantly more compressed than Lys6-Leu69 (). Arguing from Bell's theory of the rupture of individual bonds under force [31], it can be shown that pre-compression of a bond should increase its average lifetime. Based on the kinetic theory of thermally-activated rupture in metals [32], Bell wrote down the following expression for the lifetime of a single bond subjected to an external force :(1)where is the inverse of the atomic oscillation frequency ( s), is the height of the energy barrier separating the bound and unbound states, and is a measure of the distance between the bound and transition states. If the bond is also subjected to a compressive ‘prestress’ force , we then have(2)Eq. 2 can be used to estimate the contribution of pre-compression to the lifetimes of the two end hydrogen bonds. For Gln2-Glu64, we have at room temperature, assuming Å; the characteristic lifetime of the Gln2-Glu64 bond is enhanced by a factor of 400, with respect to a non-compressed hydrogen bond. The analogous calculation for the Lys6-Leu69 bond gives , suggesting that hydrogen bond compression extends the lifetime of ubiquitin under a stretching force significantly, by approximately two orders of magnitude. We note that the elastic energy stored in such a prestressed hydrogen bond can be expected to be minor, as a force of 100 pN approximately corresponds to an energy of only approximately 1 J/mol.
The magnitude of compression of the hydrogen bonds is not uniform along the beta strand pair 1/5. We propose that differences in hydrogen bond compression influence the unfolding pathway for the beta force clamp. It is known from earlier MD simulation work [28], [33] that the Lys6-Leu69 hydrogen bond always ruptures first when the protein is unfolded by stretching the N- and C-termini. The stronger pre-compression of pair Gln2-Glu64 relative to pair Lys6-Leu69 should be a contributing factor in determining this unfolding pathway. The ratio of the lifetimes for the two edge hydrogen bonds is . Thus, differences in pre-compression of hydrogen bonds of the magnitude we observe here are enough to more than double the relative lifetime of the more-compressed bond, all else being equal. This calculation is made under the assumption that the magnitude of pre-compression does not change as the protein is stretched, which is unlikely to be the case in reality; how the network of pre-tensile and pre-compressed forces evolves under an applied stretching force will be a topic for future study. Despite this simplification, our rough calculation serves to demonstrate that prestress is an important factor in determining a protein's mechanical stability, and should be taken into account along with other factors such as the orientation of the bonds relative to the pulling direction and the shielding from water by hydrophobic side-chains [33].
Apart from intra-main-chain interactions, we found that side-chain–side-chain interactions also exhibit prestress. For clarity, the inter-side-chain forces are separated into those for side-chains comprising the hydrophobic core of the protein (Fig. 4a) and for side-chains facing outwards into the solvent (Fig. 4b). The two are also shown together, projected on the protein structure, in Video S3. The inward-facing hydrophobic side-chains, with few exceptions, repel each other. None of their atoms are highly charged, and thus the inter-residue forces are dominated by steric repulsion. This is consistent with the hydrophobic core being compressed by tension in the ‘skin’ of the protein comprising the main-chain and outer side-chains, as well as by entropic forces related to the hydrophobic effect.
In contrast to the core side-chains, the forces between outward-facing side-chains are found to exhibit a mix of both compressive and tensile prestress. The strongest attractive forces (red in the figure) all correspond to salt bridges between charged side-chains (lysine and arginine are positively charged, aspartic acid and glutamic acid negatively charged). The pair with the highest tensile prestress is Lys11-Glu34, which comprises a salt bridge connecting the C-end of the alpha helix with the N-end of beta strand 2. This particular salt bridge has been shown experimentally to contribute significantly to the thermodynamic stability of ubiquitin [34]. Because of the relatively large distance between the two residues, their side-chains are forced to fully extend to satisfy the electrostatic attraction, giving rise to the observed prestress of the residue-residue force: the electrostatic attraction is counterbalanced by entropic stretching of the side-chains. It is generally true that the residue pairs with the strongest tensile prestress (eg. Asp21-Lys29, Lys27-Asp52, and Asp39-Arg74) are salt bridges between spatially separated residues. Conversely, salt bridges between nearby residues, such as Glu51-Arg54, can be satisfied without stretching the side-chains and accordingly the inter-residue forces show no significant prestress. We find evidence that some of the pre-tensed salt bridges generate significant torsion in the backbone, and this torque can be removed by mutating one of the salt bridge partners to ‘break’ the salt bridge (see Supplementary Text S1 for more details). Thus, side-chain prestress should be an important factor in stabilising the protein's native conformation.
Unlike the main-chain-only prestress network, for which each residue has significant interactions with at most two others, some nodes in the side-chain network are connected to as many as four or five others, widely separated in sequence-space. In the context of network theory, these residues may be thought of as ‘hubs’ of the network; perturbing these residues may be expected to lead to a wide-spread redistribution of force in the prestress network. In fact, simulations in which two of the most obvious hub residues, Asp52 and Arg72, were separately mutated to glycines exhibited no statistically significant changes to the prestress network beyond the local neighborhood of the mutated residue. This suggests that, at least with respect to perturbations of these specific residues, redundancy in the mechanical network imbues the pre-stress distribution with a certain amount of rigidity, and that intentional engineering of a protein's prestress network may require a more sophisticated mutation strategy beyond simply perturbing individual network hubs.
The connections between the hub residues Lys27, Asp52 and Arg72 form a clear triangle in Fig. 4b, most notably featuring a strong tensile prestress between the salt-bridged residues Lys27 and Asp52. A clue to how the high connectivity of these hubs arises comes from the effective force profile for Asp52 and Arg72 (Fig. 5a). Unlike the main-chain hydrogen bond profiles, this distribution seems to show at least three separate overlapping force profiles. This suggests that the side-chains involved are visiting a number of distinct conformational states over the course of the simulation. We indeed find evidence of very complex dynamics for Arg72 and its neighbors, which alternately involves hydrogen bonds to Asp52 and other competing residues, involving their sidechains, backbone, or both (Fig. 5a, right). It is now possible to detect the dynamics of arginine side-chains from NMR [35], so it should be feasible to directly validate our predictions of Arg72's propensities for binding to its neighbors.
Such switching between discrete states is also observed for hydrophobic residues. Fig. 5b shows the effective force profile for the residue pair Leu8-Val70. These two residues are functionally important, since they comprise a hydrophobic binding patch that is crucial for the binding of Lys48-C-linked polyubiquitin to the proteasome [36]. The force-distance distribution seems to show two distinct force curves, one with an equilibrium separation around 5.5 Å, and another around 6.5 Å. The existence of two states is confirmed from examining representative states of the trajectories (Fig. 5b, right), as well as by analysing the distribution of the angle between the two side-chains as a function of residue-residue separation over the length of the simulation (Fig. S4). As for the Asp52-Arg72 pair discussed above, the overlapping effective force curves here reveal that these different orientational states for the side-chains correspond to different inter-residue mechanical coupling regimes. It is conceivable that the switching between these states has an influence on the local balance of tension and compression, and thus on the protein's mechanical stability. This degeneracy in mechanical stability may contribute to the signature of static disorder detected in ubiquitin's rupture kinetics as measured by recent AFM experiments [37], [38]. To what extent the local sidechain disorder influences the mechanical response might depend in nature on the type of polyubiquitin linkage, which is a topic for future research.
We have shown that forces in the native ensemble of ubiquitin, measured from all-atom MD simulations, generate a tensegrity-like pattern of prestress at the residue-residue level. This includes pre-compression of the hydrogen bonds connecting beta strands, and conversely pre-tensing of alpha helix hydrogen bonds. The differences between the pre-compression of individual beta strand hydrogen bonds are sufficient to significantly modify the kinetics of hydrogen bond breakage under force, and thus should be an important factor in determining the protein kinetic stability and unfolding pathway under mechanical perturbation. Salt bridges known to be important for ubiquitin's thermodynamic stability are found to be strongly pre-tensed, and the effective force profiles for side-chain–side-chain interactions reveal a connection between side-chain dynamics and inter-residue mechanical coupling. We propose that the observed dynamic equilibrium of multiple side-chain states contributes to the complex rupture kinetics observed in AFM experiments, since each discrete side-chain state corresponds to a different well in the rough global energy landscape. A correlation is found between tensed salt bridges and twisted peptide bonds in the protein backbone, which suggests that tension in stretched side-chains, transmitted as torque to the backbone, might play a role in determining the conformation of the protein's native state. Finally, we find the tensegrity network remarkably robust with regard to mutations at network hubs.
It remains to be shown whether the observations reported here apply generally to all proteins, or are specific to ubiquitin. A preliminary study of the titin immunoglobulin [I27] domain (PDB code 1WAA [23]) also found compression of hydrogen bonds in beta sheets, and tension in salt bridges, suggesting that these are general properties (Fig. S5). An early atomic force microscope study of the mechanical stability of I27 mutants [39] showed that the point mutations Val11Pro, Val13Pro and Val15Pro reduced the protein's rupture force, as expected due to proline's inability to form inter-strand hydrogen bonds; conversely, and unexpectedly, the mutant Tyr9Pro was found to be more stable than the wild type. Our I27 prestress network (Fig. S5) gives an intriguing clue as to the origin of this effect. Tyr9 is seen to be involved in a number of repulsive force pairs, with sequentially distant partners - not the case for Val11, Val13 and Val15. It may be that the mutation Tyr9Pro, by removing these frustrating repulsive forces, allows neighboring residues do adopt a more favorable conformation and thereby stabilise the protein. A detailed study of how such mechanically important point mutations involve changes to the prestress network will be a focus of future work, as will a survey of a wide range of protein structural types, necessary to better appreciate to what extent prestress is a ubiquitous aspect of protein structure. Futhermore, we have found that the prestress network is dynamic, due to the influence of side-chain dynamics on residue-residue forces, but more work needs to be done to quantify the relationship between applied force, side-chain states and protein function. Another question is whether the effective force profiles measured here can be used as a basis for prestressed coarse-grained protein models, and in what ways the predictions of such a model would differ from traditional elastic network models, which by definition lack any prestress. Our study opens the road to re-engineer molecular tensegrity structures, to eventually allow the rational tuning of mechanical or allosteric response.
We used the Gromacs 4.0.5 package [40] to perform equilibrium all-atom simulations of ubiquitin, starting from the x-ray structure with PDB accession code 1UBI [41]. This structure is illustrated in Fig. 1a. The protein was solvated with TIP4P water [42] in a periodic cubic box of 6.5 nm per side. 16 pairs of sodium and calcium ions were added to give an effective salt concentration of 0.15 M. The OPLS all-atom forcefield [43] was chosen to describe interatomic energies. The system was subjected to a steepest-descent energy minimization, followed by a 1 ns solvent equilibration with position restraints on the heavy atoms of the protein. Then a further 1 ns equilibration run was performed with no position restraints. From the second half of the resulting trajectory, five snapshots were chosen to be the starting conformations for five independent production runs, each of which was carried out for 20 ns, giving a total of 100 ns of simulation time. All runs were performed in the NpT ensemble, with a Nosé-Hoover thermostat [44], [45] set to 300 K and Parrinello-Rahman barostat [46] at 1 atm, using a time-step of 2 fs. Electrostatic interactions were calculated using the particle mesh Ewald algorithm [47]. The same procedure was also carried out for the single-residue mutants Asp52Gly and Arg72Gly, initial structures of which were generated using PyMOL [48].
For each of the production runs, all pairwise atomic forces within the protein were output with a frequency of 1 ps using the modified FDA version of Gromacs 4.0.5 [23]. These pairwise atomic forces were then converted to residue-wise forces by summing in a vector-wise fashion, for each frame of the trajectory, all atomic forces between each pair of residues, and then projecting this total force on the vector connecting the C atoms of the two residues at that instant of the simulation. We note that due to the projection, any forces orthogonal to this connecting vectors, i.e. torques, are neglected. Their contribution to a protein's pre-stress will be subject of future investigations. The magnitudes of the residue-residue forces were then averaged for each residue pair over the full 100 ns of the simulation to give the average prestress distribution of the protein. Note that this procedure differs from earlier applications of FDA, in which residue-wise forces were calculated simply by summing the scalar magnitudes of the mean atomic pairwise forces. The protein-water and protein-ion forces were neglected. Effective force profiles for each pair of residues were obtained by selecting 10000 evenly-spaced frames from the total trajectory, and plotting the residue-wise force for each frame against the separation of the residues' C atoms. For studying the specific atomic contributions to inter-residue forces in more detail, the average atom-atom force distribution was also calculated, simply by averaging the total force between each pair of atoms in the protein over the 100 ns of simulation time. We refer to the network of forces in the protein also as ‘prestress’, in aid of establishing an analogy to previous work on the link between prestress and protein function and allostery, even though a normalization of forces by area has not been carried out, and ‘preforce’ would be the more accurate terminology. The standard error of the mean for time-averaged forces from the five independent trajectories was typically in the range of 10 pN, which is less than 10% of typical forces in hydrogen bonds, suggesting sufficient convergence. Protein visualisations were carried out with VMD [49] and PyMOL [48].
|
10.1371/journal.pntd.0004433 | Surveillance of Canine Rabies in the Central African Republic: Impact on Human Health and Molecular Epidemiology | Although rabies represents an important public health threat, it is still a neglected disease in Asia and Africa where it causes tens of thousands of deaths annually despite available human and animal vaccines. In the Central African Republic (CAR), an endemic country for rabies, this disease remains poorly investigated.
To evaluate the extent of the threat that rabies poses in the CAR, we analyzed data for 2012 from the National Reference Laboratory for Rabies, where laboratory confirmation was performed by immunofluorescence and PCR for both animal and human suspected cases, and data from the only anti-rabies dispensary of the country and only place where post-exposure prophylaxis (PEP) is available. Both are located in Bangui, the capital of the CAR. For positive samples, a portion of the N gene was amplified and sequenced to determine the molecular epidemiology of circulating strains.
In 2012, 966 exposed persons visited the anti-rabies dispensary and 632 received a post-exposure rabies vaccination. More than 90% of the exposed persons were from Bangui and its suburbs and almost 60% of them were under 15-years of age. No rabies-related human death was confirmed. Of the 82 samples from suspected rabid dogs tested, 69 were confirmed positive. Most of the rabid dogs were owned although unvaccinated. There was a strong spatiotemporal correlation within Bangui and within the country between reported human exposures and detection of rabid dogs (P<0.001). Phylogenetic analysis indicated that three variants belonging to Africa I and II lineages actively circulated in 2012.
These data indicate that canine rabies was endemic in the CAR in 2012 and had a detrimental impact on human health as shown by the hundreds of exposed persons who received PEP. Implementation of effective public health interventions including mass dog vaccination and improvement of the surveillance and the access to PEP are urgently needed in this country.
| Rabies is a widespread fatal, but preventable, viral disease transmitted from animals to humans. It has been estimated that tens of thousands of people die of rabies annually, mainly in developing countries where rabies is still a neglected disease. In the Central African Republic (CAR), a poor country located at the heart of Africa, rabies is endemic, but its burden remains poorly investigated. Here, we reported a comprehensive analysis of data for 2012 from the Institut Pasteur in Bangui, the capital of the CAR. In 2012, 966 persons reported exposure to suspicious animals, mainly dogs, and 632 received post-exposure rabies vaccination. Most of these people were from Bangui area and were under 15-years of age. Meanwhile, 82 samples from suspected rabid dogs were tested and 69 were confirmed positive. Most of the rabid dogs were owned although unvaccinated. Positive samples were sequenced and we found that three different variants actively circulated in 2012. Theses variants clustered with other viruses found in surroundings countries. Our data suggested that canine rabies was endemic in the CAR with a detrimental impact on human health. We conclude that mass vaccination of domestic dogs and the improvement of the surveillance and the access to post-exposure vaccination are urgently needed to control rabies in the CAR.
| Although this fatal disease is preventable since 1884 when Louis Pasteur developed the first vaccine strategy, rabies is still a neglected zoonosis in developing countries where it poses a significant threat to human public health [1]. More than 55,000 people die of rabies every year mostly in Asia and Africa [2]. Rabies virus (RABV) belongs to the genus Lyssavirus and family Rhabdoviridae. Although all species of mammals are susceptible to rabies virus infection, only a few species are important as reservoirs of infection [3]. The most common route of rabies transmission to humans is the bite of rabid domestic dogs [4]. The WHO considers that canine rabies potentially threatens over three billion people in Asia and Africa [5]. In humans, early clinical features of rabies are nonspecific prodromal symptoms and local neurological symptoms. After an incubation period of variable duration, RABV infects the central nervous system and the clinical presentation of rabies evolves into either encephalitic (furious) or paralytic (dumb) forms [6]. Rabies is almost always fatal once symptoms appear. However, rabies is a 100% vaccine-preventable disease and vaccination can be used in two situations: to protect those who are at risk of exposure to rabies, i.e. pre-exposure prophylaxis (PrEP), and to prevent the development of clinical rabies after exposure has occurred, i.e. post exposure prophylaxis (PEP). Moreover, mass vaccination of domesticated and wild animals is also possible [7]. With a basic reproductive ratio of less than two, canine rabies is an ideal candidate for worldwide elimination [8]. Consequently, canine rabies elimination is the key towards ultimate reduction of the disease burden in humans, as illustrated in Europe and North America, and mass vaccination of dogs is the most cost-effective way to achieve it [1, 5].
In many African countries, dog vaccination programs to date have been inadequate and failed to reduce the incidence of canine rabies [9]. However, it has been shown that effective vaccination campaigns reaching a sufficient percentage of the canine population to potentially eliminate disease and prevent future outbreaks are feasible at a cost that is economically and logistically sustainable in developing countries [9, 10]. For instance, in a study conducted in Tanzania, vaccination of 60–70% of dogs has been sufficient to control dog rabies in the studied area and to significantly reduce demand for human post-exposure rabies treatment [11]. In another study using a model parameterized with routine data on rabid dog and exposed human cases from N’Djaména in Chad, Zinsstag et al. have predicted that a single parenteral rabies mass vaccination of 70% of the dog population would be the most beneficial and cost-effective intervention and would be sufficient to interrupt transmission of rabies to humans for at least six years [12]. This study has also predicted that dog vaccination campaigns combined with human PEP would be more cost-effective compared to human PEP alone beyond a time frame of seven years.
In Africa, previous molecular epidemiological studies have shown that at least four clades are circulating [13–15]. In Central Africa, RABV strains belong to the Africa I and Africa II clades [13, 16]. The Africa I clade is very similar to current Eurasian RABV lineages and is usually grouped into a larger Cosmopolitan clade [14, 17]. The Africa II clade includes RABV strains that circulate in dogs in several Central and Western African countries [16].
The Central African Republic (CAR) is a landlocked country in Central Africa. Its capital is Bangui. It is one of the poorest countries in the world and it has been affected by political instability and internal conflicts for several decades. In the CAR, rabies has been a notifiable disease since April 2009. The surveillance mainly consists of observation of suspicious animals and brain sample collection by the national veterinary services, and laboratory confirmation by the National Reference Laboratory for Rabies. The national veterinary services operate under the National Agency for Livestock Development (ANDE) through an extended network of veterinarians and livestock technicians distributed in 122 sentinel sites across the country. The National Reference Laboratory for Rabies is located at the Institut Pasteur in Bangui. Between August 2006 and December 2008, 86 animal samples (82 of dog origin) of 101 tested positive for rabies [18]. During the same period, seven human cases were recorded and, biologically confirmation was obtained for three of them [18]. Molecular epidemiological studies have shown the co-circulation in the CAR of strains that belong to Africa I and Africa II clades [13, 16]. Oscillations in the numbers of rabid dogs have been observed in Bangui with periods of absence or low circulation and then some increases every five years (manuscript in preparation). This pattern might be comparable to the oscillations observed by Hampson et al. in Southern and Eastern Africa [19].
Public strategies for preventing rabies have been very limited in the CAR. Animal vaccination is only at the dog owner's initiative but it is expensive for most people and rarely done. The only serious attempt to control the disease was episodic mass euthanasia of stray dogs in Bangui. In the whole CAR, there is only one anti-rabies dispensary. This dispensary is located at the Institut Pasteur in Bangui and the available post-exposure rabies prophylaxis consists only in vaccination after exposure and is freely available. Administration of immunoglobulin is exceptionally available through some non-governmental organizations (NGOs) for some very severe exposures. The persons exposed to suspicious rabid animals are usually referred to the anti-rabies dispensary by the veterinary services, as the first place visited by these persons is often a veterinary clinic. The assessment whether the post-exposure vaccination is needed is usually made by a livestock technician or veterinarian based on the type and circumstances of the exposure. When possible, the veterinary services put the animal under observation and depending on the evolution of the animal health status, they can eventually advise the anti-rabies dispensary to stop the PEP. It is important to note that many persons who have been referred to the anti-rabies dispensary by the veterinary services will never be seen by a medical facility and will then remain unvaccinated.
The aim of the present study is to evaluate the extent of the threat that canine rabies represents in the CAR and to determine the dynamics of its causative agent by using data from the National Reference Laboratory for Rabies and the anti-rabies dispensary for the year 2012.
The CAR had an estimated population of 4,487,000 inhabitants in 2011 [20]. The country is divided into 16 administrative prefectures. Its capital and most populous city is Bangui with an estimated population of 740,000 inhabitants in 2011 [20]. Bangui is divided into eight urban districts and subdivided into 205 neighborhoods (or quartiers). Two conurbations surround Bangui: the cities of Bimbo (the country's second-largest city) and Bégoua.
Post-mortem samples of brains from dogs with suspicious behaviors were routinely collected by the veterinarians and the livestock technicians of the ANDE through passive surveillance. These samples were provided to the National Reference Laboratory for Rabies at the Institut Pasteur in Bangui for routine rabies diagnosis by direct fluorescent antibody test (FAT) and polymerase chain reaction (PCR). Each sample was accompanied by a form with the following information about the suspicious dog: sex, location where the dog was found dead, captured and/or killed, name and address of its owner when identified, circumstances of the capture and/or reason why it was killed, and results of the rabies diagnostic tests. Anonymized data collected between January 1st and December 31st, 2012 were retrospectively reviewed.
Owners of dogs suspected of being rabid and persons exposed to these animals were referred to the anti-rabies dispensary located at the Institut Pasteur in Bangui to receive the post-exposure treatment upon a decision by a veterinarian or a livestock technician of the ANDE based on the type (mainly bites, scratches, and licks) and circumstances of the exposure (for instance, attack by the dog without any reason). This treatment consists of local treatment of the wound through the cleansing and disinfection of the wound (a tetanus shot is also often provided), followed by a rabies vaccination regimen (Verorab, a purified vero cell rabies vaccine made by Sanofi Pasteur, Lyon France) when prescribed. The schedule used was the WHO-approved abbreviated multisite schedule or 2-1-1 regimen [21]. According to this schedule, two doses were given at day 0 (one in the right arm and one dose in the left arm), one dose on day 7 and one dose on day 21. This schedule induces an early antibody response and is considered effective when post-exposure treatment does not include administration of rabies immunoglobulin [21]. Anonymized data on the rabies exposed humans (age, gender, and origin), exposure (number and type of wound, depth and sites) and delay to consult collected between January 1st and December 31st, 2012 were retrospectively reviewed. In addition, data on the rabies vaccination status of the dogs involved, when known, was also collected.
This was a non-research national public health surveillance activity approved by the Ministry of Public Health, Population and the Fight against AIDS of the CAR. Approval by institutional review board or written informed consent was not required. Data concerning the exposed humans and/or suspicious dog owners were anonymized before analysis.
At the National Reference Laboratory for Rabies, direct FAT was routinely done using an anti-RABV nucleocapsid fluorescent conjugate (Bio-Rad, USA). For diagnostic purposes, PCR was also routinely performed using primers that target a portion of the RNA-dependent RNA polymerase-coding region. Extraction of viral RNA from the original fresh brain samples was done using QIAamp Viral RNA Mini Kit (Qiagen) according to the manufacturer’s instructions.
To investigate the genetic diversity of RABV circulating in the CAR, we amplified and sequenced a portion of the N gene of 606 nucleotides in length from rabies-positive specimens [22]. The gene of the nucleoprotein has been extensively used for molecular phylogenetic studies because of its relatively conserved variation among reservoir-associated variants and geographic lineages [23]. The date of sampling and location (city or neighborhood if the city is Bangui) of the owners of rabid dogs were available for the majority of these sequences. Sequences were then compared with reference sequences from GenBank, and phylogenetic relationships and geographic distribution were determined. Phylogenetic analysis was conducted in MEGA5 [24].
All statistical analysis was performed using STATA version 11.1 (StataCorp, TX). Categorical variables were compared by Chi-square or Fisher exact test according to the headcounts, continuous variables where compared with Student t-test or Kruskal-Wallis test when appropriate (two-sided, significance assigned at P<0.05). To analyze the spatiotemporal association between reported human exposures and detection of rabid dogs, each geographic unit (for Bangui and its suburbs: the eight urban districts and the cities of Bimbo and Bégoua, and for the rest of the country: the prefectures) by epidemiological week (i.e. weeks of the year numbered sequentially from one to 52 –week one corresponding to the first complete week of the year) was considered as a spatiotemporal unit. We performed Spearman rank correlation tests to examine the relationship between the number of reported human exposures (according to the date of exposure) and rabid dogs (according to the date of death) by spatiotemporal unit. As rabies is a communicable disease, the number of case within a same geographic area is likely to be correlated over time. We used a generalized estimating equations (GEE) approach with a Poisson distribution to confirm the significant association between the number of exposed humans (dependent variable) and the number of reported dogs (independent variable) while taking into account the autocorrelation of data over time. GEE, extension of the generalized linear model, is a population-averaged approach that accounts for the correlation between observations by introducing a working correlation matrix and by using robust variance estimators. The model used has been fully described elsewhere [25]. The correlation matrix can be arbitrarily parameterized, and we choose here a first order autoregressive structure to model the correlation of weekly number of cases in each location. This structure is indicated for time series data when two measurements that are right next to each other in time are pretty correlated, but that as measurements get farther and farther apart they are less correlated. Finally, we took into account the overdispersion of the data by adding an overdispersion scaling parameter in the model [26].
After only four positive samples by direct FAT in 2011 of seven suspicious samples, the CAR has experienced an important recrudescence of canine rabies in 2012. Of the 83 samples from suspected rabid animals received by the National Reference Laboratory for Rabies, 82 were from dogs and 69 were tested positive with direct FAT (Fig 1 and S1 Table). The remaining sample was from primate and tested negative. Of these 69 samples, 67 were positive with PCR for diagnostic purposes. Characteristics of the 82 suspected rabid dogs that contributed to the samples panel is summarized in the Table 1. Most of the dogs were male (67.3% of the dogs for which the sex was known) with a known owner (68.3%), originated from Bangui (79.3%) and were found aggressive then killed by the owner or by the veterinary services (65.9%).
In 2012, a total of 966 persons visited the anti-rabies dispensary for rabid animals exposure mainly after being prescribed to do it by a veterinarian or a livestock technician of the ANDE (Fig 1 and S2 Table). Of these, 631 received the post-exposure vaccines course (regimen 2-1-1). The characteristics of these persons are summarized in Table 2. Briefly, 53.8% were male and 57.7% of the exposed persons were children under 15 years (median age = 13 years with inter-quartile range of 8–27 years) if documented. The main type of exposure was bite by suspicious rabid dogs (91.7%) with multiple (56.0%) but superficial (54.2%) wounds if documented. The vast majority of exposed persons originated from Bangui itself or its suburbs i.e. Bimbo and Bégoua (93.2%). Most of the exposed persons visited the anti-rabies dispensary within a week after exposure (74.8%) but 31 persons came only a month or more after exposure. The time delay to visit the anti-rabies dispensary since exposure was significantly associated with the location of the exposed persons, with a mean delay of 6.2 days for the exposed persons from Bangui, 9.9 days for the persons from Bimbo and Bégoua, and 18.1 days for the persons from the other prefectures (p-value < 0.001). Exposed males were older and more likely to have multiple wounds. Children were significantly more likely to be wounded on the face or the trunk but less on the lower limb, while adults were significantly more likely to be bitten on the lower limb (p-values < 0.001) (Table 2). In addition, persons exposed were asked about the rabies vaccination status of the dogs involved in the exposure. Fifty-eight persons declared to have been exposed to a dog vaccinated but only five of them were able to give details as brand name of the vaccine and/or date of vaccination.
The Fig 2 summarizes the data on rabies surveillance in the CAR in 2012. Thirteen dogs were diagnosed as having rabies of 16 samples received from outside of Bangui and its suburbs. They originated from six different prefectures: Lobaye, Ombella M’Poko, Mambéré Kadéï, Ouham Pendé, Kémo and Ouaka (Fig 2A). Most of the persons who were exposed outside of Bangui and its suburbs came from the following prefectures: Ombella M’Poko (17 exposures), Ouaka (12 exposures), and Ouham and Mambéré Kadéï (10 exposures both) (Fig 2B). Most of the canine rabies cases that occurred outside of Bangui and its suburbs were confirmed between July and December (12 cases of 13). Most of the human exposures that occurred outside of Bangui and its suburbs were reported between June and December with 52 exposures of 64 (Fig 2C). In Bangui and its suburbs, most of the canine rabies cases were from the 4th (with 13 rabid dogs), the 3rd and the 6th districts of Bangui (with seven reported rabid dogs both) (Fig 2D). Most of the persons who were exposed in Bangui and its suburbs were from the 4th district (216 exposures), the 5th district (159 exposures), and the 3rd and the 8th districts (129 and 123 exposures, respectively) (Fig 2E). Most of the canine rabies cases that occurred in Bangui and its suburbs were confirmed between June and October (43 cases of 55) while most of the human exposures were reported during the months of June (86 cases), July (95), August (118) and September (138 cases) (Fig 2F). In the most affected districts, the peak happened in June for the 3rd district (20 exposures), in August for the 5th and the 8th districts (33 and 16 exposures respectively), and in September for the 4th district (56 exposures).
These data suggest that the CAR was hit by an epidemic of canine rabies with a peak that happened during the second half of the year 2012 (Figs 1 and 2). We found a mean number of 0.7 (standard deviation 1.4) persons exposed when no rabid dog was reported the same week in the same area, against 4.3 (standard deviation 5.4) persons exposed when at least one rabid dog was reported in the same week in the same area. We found a significant correlation between the number of rabid dogs and the number of persons exposed (Spearman Rho 0.28, P <0.001) by spatiotemporal unit (Fig 3). The correlation was also significant with the number of persons who eventually received PEP (Spearman Rho 0.31, p-values < 0.001). The associations were confirmed when using a model taking into account the temporal autocorrelation (β = 0.80 and p-value < 0.001 for the number of persons at risk; and β = 1.09, p-value < 0.001 for the number of persons having received post-exposure prophylaxis).
From the 67 PCR positive rabid dog samples, 62 sequences of the N gene were obtained [GenBank: KF34677 to KF734738]. These sequences were grouped into three variants based on their similarity. The phylogenetic analysis showed that these variants clustered within two clades: Africa I and II (Fig 4). The Africa I virus isolates described in this study (seven samples) are close to viruses isolated in the CAR in 2000, and in 2003 to 2007 (identity ≥ 99%) indicating that regular circulation of this variant, though a minority in 2012, over the time since 2000. Among the eight main groups of Africa II RABV, the strains isolated in 2012 from the CAR fell within groups c (32 samples) and e (23 samples) [16]. The Africa II group c isolates are close to viruses isolated from Chad, Ivory Coast, Mauritania, Mali Burkina Faso and Senegal (identity around 96–97%). The Africa II group e viruses are close to viruses isolated in Chad, Niger and Nigeria (identity ≥ 98%). In Bangui, all these three variants were found but the Africa I viruses were a small minority with only three samples. Outside of Bangui, these three variants were found having a specific geographic distribution. The Africa I and Africa II isolates were only found in the west of the CAR while the Africa II group e viruses were only found north-east of Bangui (Fig 4).
The main findings of our study highlight that RABV circulated actively in the CAR in 2012 in the absence of any public health intervention to stop its transmission. They also highlight that canine rabies had an important impact on the human health as shown by the hundreds of exposed persons who received PEP. Our phylogenetic analysis has indicated that strains characterized several years before were still circulating together with new strains that were recently introduced as suggested by Nakouné et al [18].
Even if our results concerning the human exposures, rabid dogs and their distribution in the country are expected to be highly under-estimated and biased toward Bangui and its surroundings, they provide an overall picture of the severe situation of the CAR for rabies. These results are in accordance with results from other studies. Bangui has carried an important disease burden. This could partly be because Bangui concentrates an important part of the country population and is where the node of the main trunk roads lies. The most affected areas in Bangui are the most populated districts except the 8th district but its proximity to other very affected districts may explain that this district was also heavily affected. Using an approach similar to Tenzin et al., we have estimated the annual incidence of exposure to rabid dogs in Bangui at 109 per 100,000 population [27]. We found that children were over-represented compared to the adults amongst the potential human exposures to rabies. This has been observed in other African settings [28]. However, without reliable data concerning the population age structure in the area, this is difficult to ascertain if children are at greater risk because the population is likely to be young. In accordance with Kayali et al., most of the rabid dogs found in the CAR were free-roaming and not vaccinated against rabies although they were owned [29]. It is surprising but not uncommon that no confirmed human rabies cases were reported despite the number of rabid dogs reported in 2012, but this may be explained by a likely under-reporting of human rabies cases due to possible overlap of symptoms with those of other infections that could lead to misdiagnosis [30, 31]. It is also possible that human rabies deaths in Bangui have remained unnoticed due to limited ability to investigate relatively rare causes of death. In a recent study in N’Djamena, Chad the annual incidence of human rabies deaths was of 0.7 deaths/100,000 inhabitants [32]. It is also likely that more rabies deaths occurred in rural areas than in urban areas [2]. It is important to note however that several suspicious cases have been reported.
Our results showed that Africa I and Africa II group c isolates seemed to be limited to the west of the country while the Africa II group e viruses were only found in the north-east of Bangui. Together with results from the phylogenetic analyses, these data suggest that Africa I and Africa II group c viruses circulated along the trunk roads between Cameroon and Bangui city while the Africa II group e viruses had followed the north-south axis between Chad and Bangui city. These two axes are the two main roads to reach Bangui from neighboring countries. This is in accordance with the fact that rabies is an example of disease introduction and spread that have resulted from the human-mediated movement of animals even if the influence of these movements on the diffusion of RABV is often difficult to quantify [33, 34]. Several studies have also shown correlation between human-dependent transportation routes and the distribution of virus [35, 36]. The simultaneous circulation of several clades within the same geographical area might indeed indicate a gene flow due to human–mediated transport of dogs. This simultaneous circulation has been observed in other countries–for instance, in China and Thailand [23, 35, 37]. Control of dog movements associated with humans is then essential for rabies control and dog mass vaccination performance.
Our study has several limitations. As previously mentioned, our results are likely to be highly under-estimated and biased toward Bangui and its surroundings. Indeed, our data reflect primarily the situation of Bangui and its near surroundings, but little was known about the rabies situation in other urban or rural areas of the CAR. This is due to a very limited access to any anti-rabies dispensary for people living outside of Bangui and its surroundings. This is also due to the difficulty to ship to Bangui the brain samples collected from suspicious animal in the other urban or rural areas of the CAR. Indeed, the absence of anti-rabies dispensaries outside of Bangui, and the distance, poor road conditions and insecurity prevents any effective surveillance and access to PEP for exposed persons in most parts of the country. In addition, exposures in endemic areas are often ignored or deemed as minimal and it is likely that only a fraction of exposed persons has visited the anti-rabies dispensary even in Bangui and its surroundings like in other African countries [28, 38]. Moreover, the dog at the origin of the exposure was not identified for most of the human exposure cases and the circumstances of the exposures were often not well documented. Another limitation of our study is that bias of awareness and auto-correlation were likely to affect our spatiotemporal correlation analyses. Indeed, a recent work in Africa has demonstrated that the intensity of rabies-control efforts seems to depend on the level of perceived prevalence [19]. Despite these limitations, dog-bite injuries have been shown to provide a valuable and accessible source of data for surveillance in countries where case incidence data are difficult to obtain [11]. Indeed numbers of human rabies deaths and post-exposure treatments correlate closely with numbers of confirmed animal cases even though medical and veterinary authorities operate independently [19].
According to our results, there is an urgent need for implementing an effective national strategy plan to control rabies in the CAR. Reducing dog density through mass euthanasia is ineffective to control rabies [39]. We recommend that this plan should be based on dog mass vaccination. This is the most efficient intervention to move towards the rabies elimination and is more cost-effective than exclusive implementation of human PEP [12, 36]. Although there is evidence that some wild canid populations can support rabies cycles in Africa, most outbreaks in wild canids are triggered by epidemics in domestic dogs rather than the converse [40–42]. In Bangui, only 379 dogs were vaccinated in 2012 (data from the national veterinary services) while the dogs population is likely to be over several tens of thousands of dogs if compared with other capital cities in the region [43]. Effective vaccination campaigns need to reach a sufficient percentage of the population to eliminate disease and prevent future outbreaks, which for rabies is predicted to be around 70% [9, 11]. In addition of achieving sufficient vaccination coverage, regular follow–up campaigns are essential for achieving elimination, especially when achieving high coverage is problematic [36, 44]. An analysis conducted in 2002 and 2006 in N’Djaména, Chad has estimated that to achieve a minimum of 70% of vaccination of the owned animals, the maximum amount that could be charged would be approximately 400–700 CFA francs (1 USD ≈ 483 FCFA) [45]. Today, the charge for dog vaccination at the government-run veterinary clinic in Bangui is 3,500 CFA francs. This is clearly too expensive for average people and the vaccine needs to be provided for free by the government or other funding bodies. Moreover, the four doses of the human vaccine cost 50,000 CFA francs in Bangui but are freely available at the Institut Pasteur in Bangui to every exposed person. It is then likely that implementation of mass dog vaccination campaigns associated with continuation of providing PEP would be more cost-effective than post-exposure vaccination of exposed humans alone in a few years as predicted by Zinsstag et al. [12].
In parallel of implementing dog mass vaccination, the national strategy plan to control rabies should also include some interventions to guarantee a better surveillance and a nationwide improved access to PEP for human exposures to rabies [46]. Indeed, limitations of our study show that there is an urgent need for reinforcement of the national surveillance system, as surveillance is the first step toward an effective control. This could be achieved by decentralizing the data collection capacities i.e. the diagnostic and anti-rabies health centers, in order to cover the entire country, thus taking a better advantage of the extended network of veterinarians and livestock technicians across the country (122 sentinel sites for veterinary surveillance). This should be accompanied by the implementation of an integrated data reporting system. In addition, the surveillance can also be improved by regular communication campaigns to raise public awareness about this deadly disease. A special emphasis would be placed on the necessity to report to the veterinary services all suspicious dogs, on the importance for the dog owners of having his/her dog vaccinated and the urgency of going to the nearest anti-rabies dispensary in case of exposure to suspicious rabid dog.
An early and appropriate access to PEP should also be guaranteed to every exposed person when needed [46]. Currently, there are too many weaknesses in the management of suspected rabies exposures. For instance, there is no follow-up of persons exposed to suspicious rabid dog who did not come to the anti-rabies dispensary but who have been seen by the veterinary services. A closer cooperation between physicians and veterinarians is needed here [32, 47, 48]. It is also important to strengthen the links between the national rabies laboratory and the anti-rabies dispensary or even merge them into the same entity to enable the identification and proper management of potentially exposed persons for all confirmed rabid dogs. A better use of PEP for the exposed persons is necessary in order to optimize the use of limited resources. This could be achieved through a better training on WHO guidelines and staffing of the dispensary. In case of the involved animal is declared healthy, a good coordination should occur to ensure that the exposed persons and the anti-rabies dispensary are timely informed in order to discontinue the PEP. This information should be clearly and appropriately recorded. In order to reduce PEP cost (currently carried by the Institut Pasteur in Bangui), intradermal regimens that require considerably less vaccine than the intramuscular regimens represent a particularly appropriate method when resources are limited [6–8]. We also recommend to extend the administrations of rabies immunoglobulin to all cases of exposure of category III (i.e. single or multiple transdermal bites or scratches, licks on broken skin, contamination of mucous membrane with saliva from licks) as recommended by WHO [6].
Finally, this national strategy plan will necessitate a strong political commitment driven by a proper awareness on the burden of the disease. Several advocacy activities were carried out towards the Ministry of Health as well as the WHO, UNICEF and some NGOs to support the setup and the sustainable operation and supply of one anti-rabies dispensary in each of the seven health regions of the CAR. However, none were successful. As WHO pointed out: “the under-reporting of rabies also prevents mobilization of resources from the international community for the elimination of human dog-mediated rabies” [5]. In the CAR, human rabies is a notifiable disease since April 2009, but this decision has not been appropriately transmitted to the health facilities, and is therefore not applied. Only a few suspected human rabies cases are reported every year. In this situation, this is difficult to attract attention from Public Health stakeholders or policy makers on this problem. However, in the light of these results, we hope that a national strategy plan based on our recommendations will be set up in a near future and run with enough political willingness to effectively control this neglected disease that mostly affects the poor, and is a glaring example of global inequalities in health care.
Taken together, these data indicate that canine rabies was endemic in the CAR in 2012 and had an important impact on the human health, as shown by the hundreds of exposed persons who have received PEP. The vaccine coverage of domestic dogs against rabies was very low. Implementation of effective public health interventions including mass dog vaccination but also the improvement of the surveillance and the access to PEP is urgently needed to control rabies in the CAR.
|
10.1371/journal.pcbi.1000790 | Analysis and Computational Dissection of Molecular Signature Multiplicity | Molecular signatures are computational or mathematical models created to diagnose disease and other phenotypes and to predict clinical outcomes and response to treatment. It is widely recognized that molecular signatures constitute one of the most important translational and basic science developments enabled by recent high-throughput molecular assays. A perplexing phenomenon that characterizes high-throughput data analysis is the ubiquitous multiplicity of molecular signatures. Multiplicity is a special form of data analysis instability in which different analysis methods used on the same data, or different samples from the same population lead to different but apparently maximally predictive signatures. This phenomenon has far-reaching implications for biological discovery and development of next generation patient diagnostics and personalized treatments. Currently the causes and interpretation of signature multiplicity are unknown, and several, often contradictory, conjectures have been made to explain it. We present a formal characterization of signature multiplicity and a new efficient algorithm that offers theoretical guarantees for extracting the set of maximally predictive and non-redundant signatures independent of distribution. The new algorithm identifies exactly the set of optimal signatures in controlled experiments and yields signatures with significantly better predictivity and reproducibility than previous algorithms in human microarray gene expression datasets. Our results shed light on the causes of signature multiplicity, provide computational tools for studying it empirically and introduce a framework for in silico bioequivalence of this important new class of diagnostic and personalized medicine modalities.
| One of the promises of personalized medicine is to use molecular information to better diagnose, manage, and treat disease. This promise is enabled through the use of molecular signatures that are computational models to predict a phenotype of interest from high-throughput assay data. Many molecular signatures have been developed to date, and some passed regulatory approval and are currently used in clinical practice. However, researchers have noted that it is possible to develop many different and equivalently accurate molecular signatures for the same phenotype and population. This phenomenon of signature multiplicity has far-reaching implications for biological discovery and development of next generation patient diagnostics and personalized treatments. Currently the causes and interpretation of signature multiplicity are unknown, and several, often contradictory, conjectures have been made to explain it. Our results shed light on the causes of signature multiplicity and provide a method for extracting all equivalently accurate signatures from high-throughput data.
| A molecular signature is a computational or mathematical model that predicts a phenotype of interest (e.g., diagnosis or outcome of treatment in human patients or biological models of disease) from microarray gene expression or other high-throughput assay data inputs [1], [2]. Multiplicity is a special form of data analysis instability in which different analysis methods used on the same data, or different samples from the same population lead to different but apparently maximally predictive signatures [3], [4]. This phenomenon has far-reaching implications for biological discovery and development of next generation patient diagnostics and personalized treatments. Multiplicity in the best case implies that generation of biological hypotheses (e.g., discovery of potential drug targets) is very hard even when signatures are maximally predictive of the phenotype since thousands of completely different signatures are equally consistent with the data. In the worst case this phenomenon entails that the produced signatures are not statistically generalizable to new cases, and thus not reliable enough for translation to clinical practice.
Some authors motivated by classical statistical considerations, attribute signature multiplicity solely to the small sample size of typical microarray gene expression studies [5] and have conjectured that it leads to non-reproducible predictivity when the signatures are applied in independent data [6]. Related to the above it has been suggested that building reproducible signatures requires thousands of observations [7]. Other authors have proposed that the phenomenon of signature multiplicity is a byproduct of the complex regulatory connectivity of the underlying biological system leading to existence of highly predictively redundant biomarker sets [8]. The specifics of what types of connectivity or regulatory relationships may lead to multiplicity have not been concretely identified however. Another possible explanation of signature multiplicity is implicit in previously described artifacts of data pre-processing. For example, normalization may inflate correlations between genes, making some of them interchangeable for prediction of the phenotype [9]–[11].
Critical to the ability to study the phenomenon empirically is the availability of computational methods capable of extracting multiple signatures from the data. Several methods have been introduced with this intent. The available methods encompass four algorithm families. The first family is resampling-based signature extraction. It operates by repeated application of a signature extraction algorithm to resampled data (e.g., via bootstrapping) [6], [12], [13]. This family of methods is based on the assumption that multiplicity is strictly a small sample phenomenon. The second family is iterative removal, that is repeating signature extraction after removing from the data all genes that have been found in the previously discovered molecular signatures [14]. This approach is agnostic as to what causes multiplicity and is heuristic since it does not propose a theory of causes of multiplicity. The third family is stochastic gene selection techniques [15], [16]. The underlying premise of the method of [15] is that in a specific class of distributions every maximally predictive and non-redundant signature will be output by a randomized algorithm with non-zero probability (thus all such signatures will be output when the algorithm is applied an infinite number of times). Similarly, the method of [16] will output all signatures discoverable by a genetic algorithm when it is allowed to evolve an infinite number of populations. The fourth family is brute force exhaustive search [17]. This approach is agnostic as to what causes multiplicity, and requires time that is exponential to the total number of genes, thus it is computationally infeasible for signatures with more than 2–3 genes (as almost all maximally predictive signatures are in practice).
The above methods, useful first attempts as they may be, are either heuristic or computationally intractable, are based on currently unvalidated conjectures about what causes multiplicity, and output incomplete sets of signatures with currently unknown generalizability. The practical benefits of an algorithm that could systematically extract the set of truly maximally predictive and non-redundant signatures include: (i) a deeper understanding of the signature multiplicity phenomenon and how it affects reproducibility of signatures; (ii) improving discovery of the underlying biological mechanisms by not missing genes that are implicated mechanistically in the disease processes; and (iii) catalyzing regulatory approval by establishing in-silico equivalence to previously validated signatures in a manner similar to bioequivalence of drugs.
To achieve these goals we provide a theoretical framework based on Markov boundary induction that enables probabilistic modeling of multiple signatures and formally connects it with the causal graph (i.e., pathways) of the data generating process [18]–[21] even when these pathways are not known a priori. We introduce a provably correct algorithm (termed TIE*) that outputs the set of maximally predictive and non-redundant signatures independent of the data distribution. We present experiments with real and resimulated microarray gene expression datasets as well as with artificial simulated data that verify the theoretical properties of TIE* and showcase its advantages over previous methods in practical settings. In particular, it is shown that TIE* having excellent sample and computational efficiency not only extracts many more maximally predictive and non-redundant signatures than all previous methods, but also that TIE* signatures are reproducible in independent datasets whereas signatures produced by previous methods are often not reproducible or have lower predictivity. The theoretical and experimental results obtained in the present study also suggest that some of the previous hypotheses about the causes and implications of signature multiplicity have to be radically reevaluated.
To simplify analysis, and without loss of generality, instead of considering all possible signatures derivable from a given dataset (via a potentially infinite variety of classifier algorithms) we only consider the signatures that have maximal predictivity for the phenotypic response variable relative to the genes (variables) contained in each signature. In other words, we exclude from consideration signatures that do not utilize all predictive information about the phenotypic response variable contained in their genes. This allows us to study signature classes by reference only to the genes contained in each class. Specifically, for a gene set X there can be an infinite number of classifiers that achieve maximal predictivity for the phenotype relative to the information contained in X. Thus, when we say “signature X” we refer to one of these predictively equivalent classifiers. This reduction is justified, for example, whenever the classifiers used can learn the minimum error decision function given sufficient sample (for a given set of genes X, the minimal error decision function minimizes the error of predicting the phenotypic variable T given X over all possible decision functions). Most practical classifiers employed in this domain as well as classifiers used in our experiments (SVMs) satisfy the above requirement either on theoretical [22], [23] and/or empirical grounds [24].
Given the above reduction of signatures to equivalence classes, the focus of this work is in extracting signatures that satisfy two desirable optimality properties: (a) maximally predictive of the phenotype (informally this means that they can form the inputs to a predictor of the phenotype which for the given dataset and population cannot be improved by any other classifier-gene subset combination), and at the same time (b) do not contain predictively redundant genes (i.e., genes that can be removed from the signature without adversely affecting the signature predictivity). Every suboptimal signature (i.e., one that does not satisfy these two properties) can be discarded from consideration when studying multiplicity.
As is proved in Text S1, two signatures X and Y of the phenotypic response variable T are maximally predictive and non-redundant if and only if X and Y are Markov boundaries of T. A Markov boundary M of T is a set of variables that (i) renders all other variables outside M and T independent of T conditioned on M (i.e., M is a Markov blanket of T) and (ii) no proper subset of M is a Markov blanket of T [18]. This definition also implies causal interpretability of M under distributional assumptions [18]–[21]. It was shown previously that the so-called intersection property of probability distributions is a sufficient condition for uniqueness of Markov boundaries [18], therefore it is also a sufficient condition for uniqueness of optimal molecular signatures. However, the extent to which signature multiplicity is present in distributions that violate the intersection property is not known.
Figure 1 shows by means of an illustrative example implications of signature multiplicity. It describes a class of Bayesian networks that share the same pathway structure (with three gene variables A, B, C and a phenotypic response variable T) and constraints on their joint probability distributions. Each member of this class is derived by parameterizing the joint probability distribution subject to the constraints. An example of a parameterized Bayesian network is provided in Figure S1. The following hold in all Bayesian networks that belong to this example class:
The above example is concerned with the large sample case. In practice, one deals with small samples where statistical inferences have to be made about large sample predictivity and redundancy. This creates an additional source of error and concomitant multiplicity. An example of this is given in Text S2.
Figure 2 presents the high-level operation of the TIE* algorithm that uses Markov boundary induction to identify the set of maximally predictive and non-redundant signatures. Text S3 provides an example trace, proof of correctness, and implementation details of the algorithm. In step 1, TIE* uses a base Markov boundary induction algorithm that identifies a single molecular signature M of the phenotype with maximal predictivity and no redundancy (Markov boundary). The same base algorithm is applied repeatedly to versions of the original dataset in which some subset of variables G has been removed (step 4). If a new signature Mnew has the same predictivity for the phenotype as M, then it is a Markov boundary and it is output (step 5). Steps 3–5 are repeated until no subset G can be generated in step 3.
The base Markov boundary induction algorithm must be suitable for the distribution at hand. Thus, TIE* is a generative algorithm that is instantiated differently for different distributions. In the experiments reported in this paper, we use as the base algorithm HITON-PC (Figure S2), which is an instance of a very broad class of Markov boundary inducers termed Generalized Local Learning [25], [26]. This choice of the base algorithm is motivated by its empirical performance in microarray gene expression and other high-throughput data as well as its theoretical properties [25]–[28].
TIE* is guaranteed to be correct in the large sample under its stated assumptions. In the small sample some signatures that are not maximally predictive and/or redundant will be statistically indistinguishable from the maximally predictive and non-redundant ones. This indistinguishability occurs at two different levels: one is estimation of predictivity and testing for statistical significance of differences in predictivity among signatures. The second level is the performance of tests of conditional independence (or functional equivalents such as Bayesian scoring) with small samples inside the base algorithm which incurs errors of type I and II. As the sample size grows, the algorithm will output only truly maximally predictive and non-redundant signatures.
We present several experiments testing the new algorithm and comparing it against 8 previously described multiple signature extraction methods. The methods comprise of four resampling-based algorithms, one iterative removal method, and three stochastic gene selection methods (details in Text S4). Brute force exhaustive search and genetic algorithms were not applied due to their computational intractability.
Before applying TIE* to real data, we test its behavior in controlled (i.e., simulated and resimulated data) experiments where generative models are known and in the case of simulated data all maximally predictive and non-redundant signatures are known as well (details about data generation are provided in Text S5). This allows us to test whether the algorithm behaves according to theoretical expectations, whether it is robust to moderate sample sizes, and whether it is sensitive to high dimensionality. This also provides clues about the behavior of TIE* and the baseline comparison algorithms in our experiments with real human microarray data.
To test reproducibility of molecular signatures, we adopt an experimental design where one microarray dataset (“discovery dataset”) is used for identification of signatures and estimation of their predictivity by holdout validation [29], and another independent dataset (“validation dataset”) originating either from a different laboratory or from a different microarray platform is used for validation of predictivity of the signatures. No overlap of subjects between discovery and validation dataset analyses occurs in this design. The criteria for dataset admissibility and exact protocol for quality assurance and processing of pairs of datasets is described in Text S6. In total, 6 pairs of gene expression microarray datasets covering both human cancer diagnosis and outcome prediction were used (listed in Table S1).
Operationally we define maximal predictivity for each dataset as follows: we apply all tested methods for extraction of multiple signatures to a dataset; then for each method we compute average predictivity of the phenotype (over all identified signatures by this method) measured by area under ROC curve (AUC); finally we compute the maximum value of the above average predictivity estimates and refer to it as “maximal predictivity”.
Statistical comparisons of predictivity between methods in the same dataset are accomplished by Wilcoxon rank sum test with α = 0.05 [30]. This is a two-sided test of the null hypothesis that two samples come from distributions with equal medians. When we use this test, the first sample contains AUC estimates of all signatures identified by one multiple signature extraction method; and the second sample contains AUC estimates of all signatures identified by another method.
Tables S2 and S3 present the results of experiments with TIE* and baseline comparison algorithms. The following are observed: (i) TIE* perfectly identifies all 72 maximally predictive and non-redundant signatures that exist in the distribution using datasets with either 30 or 1,000 variables; (ii) iterative removal identifies only 1 signature because all other signatures have a common variable and thus cannot be detected by this method; (iii) KIAMB fails to identify any optimal signature due to its sample inefficiency, and because of the same reason its signatures have poor classification performance; (iv) resampling-based methods either miss many optimal signatures and/or output many redundant variables in the signatures.
We applied TIE* to resimulated gene expression data with sample sizes: 300, 450,…, 1500, 2250, 3000,…, 30000. A signature is operationally considered as non-reducible if it is not properly included in any other signature output by this method (i.e., it is a proxy of having no redundant genes). For example, if a method outputs 3 signatures with the following genes: {A, B, C}, {A, B, X}, and {A, B}, only signature {A, B} is non-reducible. The number of maximally predictive signatures (as confirmed in independent data by holdout validation) and the number of maximally predictive and non-reducible signatures output by the algorithm for each sample size in resimulated data is shown in Figure S3. As sample size increases, the number of output maximally predictive signatures drops but then remains constant in the range 160–644 (or 53–279 for non-reducible signatures) for datasets with ≥4,500 samples. This is consistent with the existence of at least two sources of multiplicity: one is small sample size and the other is multiplicity intrinsic to the nature of gene-gene and gene-phenotype relations. As sample size grows, the first source vanishes and only the second one remains. Since the resimulated data distribution closely mimics the real-life distribution (Text S5), this experiment supports the hypothesized existence of multiple maximally predictive and non-redundant signatures in very large samples (>10,000) contrary to the theoretical model of [5].
Table S4 shows that TIE* achieves maximal predictivity in 5 out of 6 human microarray validation datasets. Non-TIE* methods achieve maximal predictivity in 0 to 2 datasets depending on the method. In the dataset where TIE* has predictivity that is statistically distinguishable from the empirical maximal one (Lung Cancer Subtype Classification), the magnitude of this difference is <0.009 AUC on average over all discovered signatures; thus this particular deviation from maximal predictivity may be considered negligible for most practical purposes.
A detailed example of application of multiple signature extraction methods to the Leukemia 5 Yr. Prognosis task is provided in Figure 3. The figure shows predictivity estimated in the discovery dataset (using an unbiased error estimator and protocol) against predictivity verified in the validation dataset for each signature. As can be seen, TIE* signatures have superior predictivity and lower variance compared to the signatures output by other methods. Similar behavior can be observed in other tasks as well.
Figure 4 plots predictivity estimated in the discovery dataset (using an unbiased error estimator and protocol) against predictivity verified in the validation dataset for all methods averaged over all datasets and all discovered signatures. Recall that validation datasets originate from different laboratories and/or using different microarray platforms than discovery datasets. The horizontal distance of each method to the diagonal measures the magnitude of overfitting defined as the difference (ε1-ε2), where ε1 = expected performance in the validation data obtained by holdout validation in the discovery dataset, and ε2 = observed validation dataset performance. TIE* rests slightly right of the diagonal denoting no overfitting, or equivalently perfect statistical reproducibility on average. However all other methods exhibit varying degrees of non-reproducibility. Depending on method the average magnitude of overfitting varies from 0.02 to 0.03 AUC.
Analysis of the signatures output by TIE* reveals that they share many genes in common. Table S5 shows the number of common genes in 50%, 60%, …, 100% of output signatures for each dataset. Genes differ in the percentage of signatures they participate in. A heuristic that genes that belong to a larger fraction of signatures are localized closer to the pathway(s) affecting and being affected by the phenotypic response variable may be useful in exploratory studies, however this does not hold in all distributions [31].
The properties of the data-generative process affect computational feasibility of the signature discovery. In the worst case, it is computationally infeasible to discover even one of all optimal signatures with all known sound algorithms (i.e., algorithms that under specific conditions provably guarantee to provide the desired output; for the purposes of the present paper, to find a signature that is optimal in the population). However, there exist several sound algorithms for extracting an optimal signature that run in low-order polynomial time in real high-throughput data (e.g., HITON-PC). Even if the computational cost of discovery of one signature was constant, the number of all optimal signatures can grow exponentially large in the number of genes measured (for an example see Text S7). Thus the computational cost of dissecting signature multiplicity ranges from low-order polynomial (tractable) to super-exponential (infeasible) depending on the distribution. The worst-case characteristics are a property of the distribution analyzed and not the algorithm employed. One can thus only hope that real-life high-throughput data distributions are not representative of the worst-case theoretical ones. In addition, algorithms are needed that exploit the structure of the generative process to discover multiple signatures efficiently when the distribution allows it. Our experiments support that real-life data does not behave as the worst-case expected theoretical scenarios because TIE* terminated within at most several hours in each of the 6 microarray datasets that contains more than 10,000 oligonucleotide probes (using a Matlab implementation on a workstation with a single Intel Xeon 2.4 GHz processor and 4Gb of RAM). One can postulate various reasons for the tractability such as: (a) that biological pathways are sufficiently sparse thus not allowing for an exponential number of optimal signatures; (b) that to the extent that multiplicity denotes biologically redundant function, there is an “economy” of such redundant mechanisms, and (c) that a very large number of optimal signatures requires constraints on the network topology that are inconsistent with the structure of many biologically functional pathways.
The results of the present study refute or suggest that modifications are needed to several widespread positions about causes of signature multiplicity. The example model pathway in Figure 1 demonstrates that signature reproducibility neither precludes multiplicity nor requires sample sizes with thousands of subjects. It also shows that multiplicity of signatures does not require dense connectivity of the underlying pathways. Similarly, it shows that noisy measurements or normalization are not necessary conditions for signature multiplicity. The resimulation experimental data suggest that networks modeled on real gene expression data can exhibit signature multiplicity even in large sample sizes and that in this type of data, multiplicity is produced by a combination of small sample size-related variance and intrinsic multiplicity in the underlying network. The results with real human microarray datasets show that multiple signatures output by TIE* are reproducible and maximally predictive even though they are derived from small sample, noisy, and heavily-processed data.
Our results are consistent with the hypothesis that signature multiplicity in real-life datasets is created by a combination of several factors that include the following: First, the intrinsic information redundancy (due to gene-gene and gene-phenotype relations) in the complex regulatory network of the underlying biological system. Second, the variability in the output of gene selection and classifier algorithms especially in small sample sizes. Third, the small sample statistical indistinguishability of signatures that have different large sample predictivity and/or redundancy characteristics (example is given in Text S2). Fourth, the presence of hidden/unobserved variables (example is given in Text S8). Fifth, correlated measurement noise components that introduce a bias in gene expression profiles (e.g., noise that is localized in regions of microarray chips) [32]. Sixth, RNA amplification techniques that systematically distort measurements of transcript ratios (e.g., double-round T7-based amplification protocol) [33]. Seventh, cellular aggregation and sampling from mixtures of distributions that affect inference of conditional independence relations that are needed to establish model equivalence according to our framework for multiplicity [34]. Eighth, normalization and other data pre-processing methods that artificially increase correlations among genes (e.g., multivariate normalization in microarrays) [9]–[11]. Finally ninth, the engineered redundancy in the assay technology platforms (e.g., multiple probes for the same gene). In datasets produced by dissimilar underlying biological mechanisms, assayed with different platforms and pre-processed and modeled with a variety of algorithms, the relative contributions of the above factors to multiplicity will vary. As a result, methods that rely on a specific cause of multiplicity or combination of causes will not output all maximally predictive and non-redundant signatures in all types of high-throughput data.
With regard to non-TIE* baseline comparison algorithms, we note that resampling-based methods that use bootstrap samples to extract signatures may stop producing multiple signatures in large sample sizes. This is expected because resampling methods are designed to address directly only the small sample multiplicity and not the intrinsic multiplicity which persists in large samples. Iterative removal, on the other hand, by its design always fails to identify all maximally predictive and non-redundant signatures that have genes in common. KIAMB among the baseline algorithms has the strongest theoretical motivation because it can be shown to discover all Markov boundaries for a subset of distributions. However, a major limitation of KIAMB is that it has sample size requirements that range from at least linear to exponential to the number of genes in a signature (depending on the test of independence employed). This makes the algorithm not only computationally inefficient but also prone to statistical errors in small sample sizes. This leads to its substantial observed overfitting in the experiments with real data and its inability to find the maximally predictive and non-redundant signatures in simulated data. KIAMB, being a randomized search algorithm, also guarantees to output all optimal signatures that satisfy its distributional requirements, but only after infinite number of runs. The method by design will discover the same signatures over and over again further compounding its computational inefficiency.
Dealing with molecular signature multiplicity using a Markov boundary framework and the TIE* algorithm does not require a particular combination of factors causing signature multiplicity in order to be able to discover all maximally predictive and non-redundant signatures. Because of efficient heuristics, TIE* can extract the signature set very quickly when the connectivity is locally sparse, and the number of true optimal signatures is low-order polynomial or smaller in the number of variables. A very important factor for the performance of TIE* is the choice of the base algorithm to discover non-redundant and maximally predictive signatures in the distribution at hand. Latest developments in Markov boundary discovery provide such tools for high-throughput data. One of the key advantages of these methods is their ability to implicitly control for false discovery rate [25].
Our experiments used real data exclusively from human cancer gene expression microarrays because of pragmatic reasons: known identity of observed variables, number and size of datasets, and maturity of standardization protocols that allows for multiple independent dataset validation experiments. The methods introduced here are however directly applicable to any high-throughput data, and future research in this direction is warranted. As an example of applicability of TIE* to high-throughput data beyond gene expression microarrays, we applied the method to proteomics mass-spectrometry data where TIE* identified hundreds of signatures of ovarian cancer with AUC = 0.95−0.98 (details in Text S9).
|
10.1371/journal.pntd.0003144 | Direct Comparison of the Efficacy and Safety of Oral Treatments with Oleylphosphocholine (OlPC) and Miltefosine in a Mouse Model of L. major Cutaneous Leishmaniasis | Cutaneous leishmaniasis (CL) represents a range of skin diseases caused by infection with Leishmania parasites and associated with tissue inflammation and skin ulceration. CL is clinically widespread in both the Old and New World but lacks treatments that are well tolerated, effective and inexpensive. Oleylphosphocholine (OlPC) is a new orally bioavailable drug of the alkylphosphocholine family with potent antileishmanial activity against a broad range of Leishmania species/strains.
The potential of OlPC against Old World CL was evaluated in a mouse model of Leishmania (L.) major infection in BALB/c mice. Initial dose-response experiments showed that an oral daily dose of 40 mg/kg of OlPC was needed to impact time to cure and lesion sizes. This dose was then used to directly compare the efficacy of OlPC to the efficacy of the antileishmanial drugs miltefosine (40 mg/kg/day), fluconazole (160 mg/kg/day) and amphotericin B (25 mg/kg/day). OlPC, miltefosine and fluconazole were given orally for 21 days while amphotericin B was administered intraperitoneally for 10 days. Ulcer sizes and animal weights were followed up on a weekly basis and parasitemia was determined by means of a real-time in vivo imaging system which detects luminescence emitted from luciferase-expressing infecting L. major parasites. Amphotericin B and OlPC showed excellent efficacy against L. major lesions in terms of reduction of parasitic loads and by inducing complete healing of established lesions. In contrast, treatment with miltefosine did not significantly affect parasitemia and lesion sizes, while fluconazole was completely ineffective at the dose regimen tested.
Given the data showing the outstanding efficacy and tolerability of OlPC, our results suggest that OlPC is a promising new drug candidate to improve and simplify current clinical management of L. major CL.
| Cutaneous leishmaniasis (CL) is a vector-borne parasitic disease transmitted to humans by sandflies and characterized by local ulcerative skin lesions. The disease is linked to poverty in the Middle-East, North and East Africa, South-Central Asia and South America, with 0.7 to 1.2 million new annual cases estimated. In most endemic regions CL treatment relies on injections with pentavalent antimonials, old generation drugs with considerable side effects and long treatment regimens. CL is therefore a highly undertreated disease in need of easy-to-administer, orally bioavailable and well-tolerated agents with broad clinical activity. To date, the only oral drug with acceptable efficacy against leishmaniasis is miltefosine, an alkylphosphocholine with a narrow therapeutic window that limits its use. Given the existing clinical need for CL, we tested the efficacy of oleylphosphocholine (OlPC) in a validated mouse model of Old World (Leishmania major) CL. OlPC is a new orally bioavailable drug of the same family as miltefosine with potent and broad leishmanicidal activity. In direct comparison with miltefosine, our results indicate that OlPC induces higher parasite clearance and lesion healing with measurable improved tolerance. These promising observations warrant further research on OlPC as a new drug to improve clinical management of CL.
| Leishmaniasis describes a range of visceral and cutaneous disease forms caused by infection with protozoal parasites of the Leishmania genus, transmitted to humans by phlebotomine sandflies [1], [2]. Cutaneous leishmaniasis (CL) is characterized by primary localized skin infections that sometimes resolve without treatment, but can also evolve into disseminated, diffuse, or mucocutaneous lesions. In the Old World, CL is caused mainly by L. major, L. tropica and L. aethiopica, whereas in the New World L. braziliensis, L. panamensis, L. amazonensis, L. guyanensis and L. mexicana are the main causative agents [3]. Based on most recent estimates, about 0.7 to 1.2 million new CL cases occur annually [4]. Treatment of leishmaniasis in most endemic regions relies on multiple intralesional, intramuscular or intravenous injections of pentavalent antimonials, old generation drugs that cause considerable toxicity and have unacceptably long treatment schedules which undermine adherence to therapy and contribute to resistance development [2], [5]. Although in the past decade significant progress was made in the field of antileishmanial drug development with the approval of amphotericin B, paromomycin and miltefosine, considerable disadvantages remain [6]. In particular for CL, treatment regimens are poorly justified and have sub-optimal efficacy. Although local therapy can be used to treat certain forms of CL, procedures such as intralesional injection, cryo- or thermotherapy can be painful and may require local anesthesia [3]. Ointments or creams such as those containing paromomycin (WR279, 396) are more suitable for uncomplicated CL cases, and their efficacy for treatment of New World CL and complicated CL (multiple lesions) is still under study in well controlled clinical trials in Panama and Peru. Whether administered topically or systemically, treatment efficacy against CL is highly variable and depends both on the infecting Leishmania strain and on the geographic region [2]. As CL is not a life-threatening disease, treatment recommendation is based on a risk-benefit ratio for every case [2]. In view of the considerable drawbacks of current therapies, in particular the long treatment times and associated side effects, moderate clinical manifestations of CL are likely to be undertreated which increases the chance of patients developing debilitating scars or more severe forms of the disease [3]. Orally bioavailable and well-tolerated agents that are effective against a wide range of clinical CL manifestations are needed, especially against complicated CL. So far the only oral drug with acceptable efficacy against leishmaniasis is miltefosine, an alkylphosphocholine generally used in a long 28-day treatment regimen that associates with dose-limiting gastro-intestinal toxicity [3], [6], [7]. Miltefosine has been tested for CL treatment showing acceptable but variable clinical efficacy [8]. Despite these variable results miltefosine (brand name Impavido) was recently approved by the United States Food and Drug Administration.
Oleylphosphocholine (OlPC) is a new chemical entity belonging to the alkylphosphocholine family showing antileishmanial activity against a broad range of Old and New World Leishmania species/strains. While OlPC and miltefosine demonstrate comparable activity in vitro, OlPC revealed to be of higher efficacy in vivo when tested in a predictive hamster model of visceral leishmaniasis [9]. This study evaluates the value of OlPC for the treatment of Old World CL (OWCL) by testing it in laboratory models of L. major infected-mice. These models have undergone internal validation and are reproducible according to industry standards [10].
Female BALB/c mice weighing 20–25 grams were purchased from Charles River (Wilmington, MA). The animal protocol was approved by the Walter Reed Army Institute of Research (Silver Spring, MD) institutional animal ethics committee in accordance with national guidelines (protocol number 13-ET-26). Research was conducted in compliance with the Animal Welfare Act, other federal statutes, and regulations that relate to animals and experiments involving animals, and principles stated in the Guide for the Care and Use of Laboratory Animals [11]. The authors abide to the reductionist approach of using animal models in drug development.
Luciferase-labeled or standard L. major promastigotes were cultured in Schneider's medium (Lonza) supplemented with 20% heat-inactivated fetal bovine serum at 25°C. Animals were infected at the base of the tail with 1×107 stationary phase promastigotes. The ulcer areas were measured with a calibrated digital caliper once a week. The average diameter of each tail lesion was calculated as the mean of the horizontal and vertical diameters, and this value was used to calculate the ulcer size area in mm2. The following parameters were examined to determine toxicology inequity of the study drugs: cure or distress/death, body weight, general physical and coat appearance.
Crystalline OlPC was supplied by Dafra Pharma Research & Development (Turnhout, Belgium) while miltefosine was purchased from Panslavia Chemicals LLC and provided by the WRAIR depository (Rockville, USA). Fluconazole was purchased from Sigma-Aldrich (St-Louis, USA) and amphotericin B (Ambisome) from Astellas Pharma US Inc. (Northbrook, USA). Miltefosine and OlPC stock solutions were prepared in 1× PBS and stored at room temperature in the dark for a maximum of 7 days. Fluconazole was dissolved in HECT (in 0.5% (w/v) hydroxyethyl cellulose and 0.2% (0.5% HECT, v/v) Tween-80 in distilled water), then homogenized using a PRO Scientific Inc. Monroe, CT homogenizer. AmBisome was dissolved in double distilled sterile water.
Efficacy was assessed by comparing the suppression of lesion size after 28 days in the drug treated group to that in negative vehicle control as previously described [10]. Percent suppression is defined as {[(LS(-)C)−LS(drug)]/LS(-)C}×100, where LS(-)C = lesion size in negative control and LS(drug) = lesion size in drug group. The threshold for success is a percent suppression which is at least 50% of the positive control amphotericin B [10].
Luciferin (D-Luciferin potassium salt, Xenogen Corporation, Almeda, CA /Goldbio, St Louis, MO), the luciferase substrate, was intra-peritoneally injected into mice at a concentration of 200 mg/kg 18 minutes before bioluminescence analysis. Mice were anaesthetized with isoflurane (MWI veterinary Supply, Harrisburg, PA) and maintained in the imaging chamber for analysis. Emitted photons were collected by auto acquisition with a charge couple device (CCD) camera (IVIS Imaging System 100 Series) using the medium resolution (medium binning) mode. Analysis was performed after defining a region of interest (ROI) that delimited the surface of the affected area. Total photon emission from each infected tail base area was quantified with Living Image software (Xenogen Corporation, Almeda, CA), and results were expressed in photons/sec.
A first set of dose-response experiments was used to assess the capacity of OlPC to cure L. major cutaneous lesions in BALB/c mice when given orally. L. major promastigotes were injected at the tail base of the mice and local lesions were allowed to develop until they reached optimal lesion size of ∼50 mm2. Mice were then grouped (n = 5 per group) based on equivalent average lesion sizes and daily oral treatment with OlPC was initiated. Based on previous data generated in L. infantum infected hamsters 9, doses of 10, 20, and 40 mg/kg of OlPC were selected to be given for 5 or 10 consecutive days (total doses of 50, 100, 200 and 400 mg/kg) (Table 1). Ulcer sizes were measured from the first treatment day (Day 0) up to Day 28 post treatment start, and compared to those of vehicle treated animals. In this mouse treatment model, dosing of 10 and 20 mg/kg daily for 5 or 10 days had little to no impact on lesion growth, while the dose of 40 mg/kg was able to significantly reduce their sizes. For the 5-day and 10-day regimens, lesion sizes were reduced by 34.0% and 93.5%, respectively (Table 1). The 10-day regimen at 40 mg/kg was independently validated by intraperitoneal (IP) treatment. In this experiment the reduction of lesion sizes was also significant (66.8%, Table 1), although lower than what had been seen with oral treatment. No sign of drug toxicity (as defined in Materials and Methods) was observed in any of the treatment groups. Taken together, these data pointed that an oral daily dose of 40 mg/kg was needed for effective treatment of L. major lesions in BALB/c mice, although total disappearance of lesions was not observed with 10 days of treatment.
Building on the previous dataset, the efficacy of OlPC to cure L. major induced lesions in BALB/c mice was directly compared to those of the clinically used antileishmanial drugs miltefosine, fluconazole and amphotericin B. For this experiment, mice were infected with luciferase-labeled promastigotes and lesions were allowed to develop until they reached optimal lesion size of ∼50 mm2. Mice were then grouped (n = 6 per group) based on equivalent average lesion sizes. To allow direct comparison between treatments, OlPC (40 mg/kg/day), miltefosine (40 mg/kg/day) and fluconazole (160 mg/kg/day [10]) were used orally for 21 days alongside PBS-treated control animals (considering first treatment day as Day 0). As amphotericin B is not orally bioavailable, this drug was administered IP at 25 mg/kg/day for 10 days based on previous experience [10] and served as a positive control. Parasitemia (IVIS), ulcer sizes, and animal weights were followed in each group on a weekly basis.
As expected, treatment with the reference drug, Amphotericin B, led to a rapid reduction of the parasite loads (visible on Day 7) correlating with lesion size reduction as of Day 19. By Day 27, lesions had healed/cured (defined as 100% re-epithelialization – normal skin) and parasites could not be detected in the mice of this group (Figure 1A, 1B ◊). Although occurring more slowly, response to oral OlPC also led to gradual but complete clearance of parasitemia (seen on Day 12) followed by lesion regression/healing as of Day 27 (Figure 1A, 1B •). In the OlPC-treated group the lesions had completely re-epithelialized/healed by Day 34. In contrast, parasitemia in both miltefosine and fluconazole treated groups never significantly differed from those of the control group, and no lesion regression was observed (Figure 1A, 1B). On Day 34, the control group had an average lesion size of 118.9 ± SEM 32.1 mm2, the fluconazole-treated group 129.2 ± SEM 25 mm2 and the miltefosine-treated group 55 ± SEM 23.7 mm2.
A detailed analysis of the Day 19 post treatment start time point is presented on Figure 2, with pictures of the luminescent signal in individual mice (Figure 2A), individual group luminescence values (Figure 2B) and lesion sizes (Figure 2C). The IVIS analysis of OlPC-treated mice clearly shows that OlPC is highly effective at clearing parasites at the lesion site despite the fact that the lesions have not yet started to regress in size. Although the OlPC- and miltefosine- treated groups show similar average lesion sizes at this time point, the difference in the activities of both drugs is nevertheless unambiguous.
Mice remaining beyond Day 34 were closely monitored until the lesion grew to a size >200 mm2 or until recrudescence of the ulcers, which were considered as clinical end points. Mice of the control group were euthanized as of Day 35 due to excessive lesion sizes together with the fluconazole-treated mice, indicating that fluconazole was ineffective at the selected dose regimen. Miltefosine treatment appeared to slow down the progression of the ulcers up to Day 34 (Figure 1B) suggesting partial efficacy at 40 mg/kg/day×21 days. However the lesions showed enlargement as of Day 40 (end point 96.4 ± SEM 30.8 mm2). In contrast, both OlPC and amphotericin B-treated animals remained lesion free up to Day 54. On day 74, which was the final time point evaluated, both the amphotericin B and OlPC mice had relapsed, showing average ulcer sizes of 31.3 ± SEM 14.9 mm2 and 28.4 ± SEM 14.6 mm2, respectively (not shown). In conclusion, although treatment with oral OlPC had a slower action than IP amphotericin B, the overall capacity of both drugs to clear the infection in the studied model appeared to be similar and far superior to the one of oral miltefosine at equivalent dose.
A moderate weight loss was observed in all treatment groups during the 35-day follow-up period, which generally correlated well with disease progression in terms of lesion size. The average weight loss reached a maximum of 8.9% in the vehicle control group and 13.4% in the fluconazole-treated group on Day 34 (Table 2). The miltefosine-treated mice experienced higher weight loss during the treatment period, with an average of 20.6% weight loss on Day 13 (i.e. mid-treatment), indicating potential drug safety issues at the dose used. Amphotericin B- and OlPC- treated mice both experienced a ∼7% weight loss during treatment (peaking on Day 10 and Day 13, respectively), followed by overall weight gain compared to baseline by Day 34 post treatment start.
Animals euthanized or found dead during the first 35-day follow-up period are reported in Table 3. For two of the mice (1 in control group, 1 in OlPC group), any association with drug toxicity is formally excluded. As for the other found dead animals (3 in the miltefosine group, 2 in the OlPC group and 1 in the fluconazole group), the possibility of cumulative toxicity or complications due to daily gavage (or a combination or the two) could not be excluded. Of those, it is interesting to note that the 3 deaths in the miltefosine group occurred earlier (Day 12, 13 and 19) compared to the ones in the OlPC group (Day 22 and 23), and were associated with piloerection, a recognized sign of sickness in mice, and important weight losses (based on last weight measurement before death; Table 3). Gross necropsy in the two found dead animals of the OlPC-treated group (mouse # 581 and #575) revealed no specific pathological findings, and only mouse #575 underwent weight loss during treatment.
In conclusion, both OlPC and amphotericin B showed excellent efficacy against L. major lesions in mice by reducing parasitemia and inducing healing of established lesions. In contrast, treatment with miltefosine at the same dosing regimen as OlPC did not significantly affect parasitemia or induce lesion regression, while fluconazole was completely ineffective at the dose tested. OlPC also appeared better tolerated than miltefosine at equivalent dosing regimen.
As human leishmaniasis comprises several clinical syndromes caused by dozens of Leishmania species across the globe, it is unlikely that one drug or drug combination will be effective for all clinical forms of the disease [12]. Therefore, the development of new antileishmanial drugs is needed, preferably with a low side effect profile, oral bioavailability, efficacy in a short treatment regimen, and which can be manufactured at low-cost and adapted for use in rural areas [10], [13]. Currently the only orally bioavailable drug for leishmaniasis is miltefosine, an alkylphosphocholine with a narrow therapeutic window mainly due to its gastrointestinal toxicity. Vomiting and/or diarrhea have been reported in every clinical trial performed with miltefosine [8], and although clinical evidence has suggested efficacy against CL, there is a large variation in clinical response and, in particular for OWCL, more data is needed [8], [14], [15]. Its main limitations are treatment compliance and hence potential for selection of drug resistant parasites and teratogenicity (pregnancy must be avoided during treatment and during the following two months). In this study, the two alkylphosphocholines, miltefosine and oleylphosphocholine (OlPC), were compared side by side for efficacy and safety in a mouse model of OWCL. Of note, the daily dose used approximates the human equivalent dose at which miltefosine is generally used in clinical practice against CL, namely 2.5–3.3 mg/kg (corresponding to 30–40 mg/kg/day in mice) [3], but for 21 days instead of the recommended 28 day regimen.
Based on data accumulated so far in two independent rodent models of leishmaniasis, namely L. infantum visceral infection in Golden hamsters [9] and L. major cutaneous infection here in BALB/c mice, OlPC has greater in vivo efficacy and superior safety profile compared to miltefosine when compared at equivalent dose regimen. In addition, although no direct comparison with miltefosine was performed, the clinical efficacy of OlPC against L. infantum canine leishmaniasis (CanL) was also demonstrated in naturally infected dogs using a 14-day regimen of 4 mg/kg/day [16], a daily dose exceeding the maximum tolerated dose of miltefosine in that species (recommended miltefosine regimen in dogs: 2 mg/kg for 28 days). Therefore overall, OlPC is better tolerated than miltefosine and has better efficacy (i.e. wider therapeutic window). Since the antileishmanial activity of OlPC and miltefosine is similar in vitro [9], the difference in therapeutic windows between the two drugs could result from differences in oral bioavailability, tissue distribution, or in the affinity of the drugs for the parasites. The detailed PK/PD analysis of OlPC vs. miltefosine in animal models is interesting and deserves further attention. These studies will allow efficient translation of the knowledge accumulated in animal models in future clinical studies in humans aiming at comparing the clinical efficacy of OlPC and miltefosine.
The other two comparative drugs used in our study were fluconazole and amphotericin B. Regarding fluconazole, despite the fact that clinical efficacy in CL patients has been reported in the literature against L. major CL at 200 mg daily for six weeks in Saudi Arabia [17] and L. braziliensis CL with 8 mg/kg (highest dose tested; corresponding to about 100 mg/kg as an equivalent dose for mice) [18], this drug appeared to be ineffective in our study when given at 160 mg/kg×21 days. However it cannot be excluded that fluconazole would still be effective in the context of a longer treatment period in Balb/c mice. As for amphotericin B, this drug came out as the “best” overall (considering speed of recovery, average group weight loss and group mortality). However as this drug was given IP, it most likely had an earlier Tmax compared to the other drugs given orally. In addition, IP injections require different types of manipulations than oral gavage, which might influence overall group mortality, aside from the fact that the treatment was only for 10 days as opposed to 21 days for the other groups. Nevertheless, despite the differences in the routes of administration, OlPC achieved similar absolute efficacy than amphotericin B in terms of reduction of parasite loads and lesion remission. The fact that OlPC is orally bioavailable represents a huge practical advantage over amphotericin B considering similar efficacy.
Taken together, our study suggests that even though the optimal oral regimen of OlPC against CL still requires further study and optimization, this new alkylphosphocholine opens the possibility of future improvement of CL patient management in terms of having a well-tolerated oral treatment associated with good patient compliance. In this regard the full FDA/EMA-compliant toxicology and safety pharmacology analysis of oleylphophoscholine is being assembled to pave way to further human clinical development in CL/MCL patients. Having a new oral treatment available will also reduce treatment costs, a factor extremely important in remote settings where cold chain distribution and parenteral drug administration remain challenging.
|
10.1371/journal.pntd.0004549 | Transforming Clinical Data into Actionable Prognosis Models: Machine-Learning Framework and Field-Deployable App to Predict Outcome of Ebola Patients | Assessment of the response to the 2014–15 Ebola outbreak indicates the need for innovations in data collection, sharing, and use to improve case detection and treatment. Here we introduce a Machine Learning pipeline for Ebola Virus Disease (EVD) prognosis prediction, which packages the best models into a mobile app to be available in clinical care settings. The pipeline was trained on a public EVD clinical dataset, from 106 patients in Sierra Leone.
We used a new tool for exploratory analysis, Mirador, to identify the most informative clinical factors that correlate with EVD outcome. The small sample size and high prevalence of missing records were significant challenges. We applied multiple imputation and bootstrap sampling to address missing data and quantify overfitting. We trained several predictors over all combinations of covariates, which resulted in an ensemble of predictors, with and without viral load information, with an area under the receiver operator characteristic curve of 0.8 or more, after correcting for optimistic bias. We ranked the predictors by their F1-score, and those above a set threshold were compiled into a mobile app, Ebola CARE (Computational Assignment of Risk Estimates).
This method demonstrates how to address small sample sizes and missing data, while creating predictive models that can be readily deployed to assist treatment in future outbreaks of EVD and other infectious diseases. By generating an ensemble of predictors instead of relying on a single model, we are able to handle situations where patient data is partially available. The prognosis app can be updated as new data become available, and we made all the computational protocols fully documented and open-sourced to encourage timely data sharing, independent validation, and development of better prediction models in outbreak response.
| We introduce a machine-learning framework and field-deployable app to predict outcome of Ebola patients from their initial clinical symptoms. Recent work from other authors also points out to the clinical factors that can be used to better understand patient prognosis, but there is currently no predictive model that can be deployed in the field to assist health care workers. Mobile apps for clinical diagnosis and prognosis allow using more complex models than the scoring protocols that have been traditionally favored by clinicians, such as Apgar and MTS. Furthermore, the WHO Ebola Interim Assessment Panel has recently concluded that innovative tools for data collection, reporting, and monitoring are needed for better response in future outbreaks. However, incomplete clinical data will continue to be a serious problem until more robust and standardized data collection systems are in place. Our app demonstrates how systematic data collection could lead to actionable knowledge, which in turn would trigger more and better collection, further improving the prognosis models and the app, essentially creating a virtuous cycle.
| The 2014–15 EVD outbreak in West Africa has eclipsed in magnitude all combined past EVD outbreaks since the disease was first identified in 1976 [1]. As of February 17, 2016 (http://www.cdc.gov/vhf/ebola/outbreaks/2014-west-africa/case-counts.html), a total of 28,639 cases have been reported (15,251 laboratory-confirmed) and 11,316 total deaths. The outbreak constitutes one of the most serious worldwide health emergencies in modern times, with severe socioeconomic costs, particularly in the West African nations of Liberia, Sierra Leone, and Guinea. Although vaccine development is promising [2], the prospect of future outbreaks looms. The report of the WHO Ebola Interim Assessment Panel also points to several shortcomings in the initial response [3], noting that “better information was needed to understand best practices in clinical management” and that “innovations in data collection should be introduced, including geospatial mapping, mHealth communications, and platforms for self-monitoring and reporting”.
Given these circumstances, the development of accurate and accessible computational methods to track the progression of the outbreak and model various aspects of the disease is beneficial not only for the research community, but also for health care personnel in the field. In particular, prognosis prediction models based on the available patient information would be of great utility. Such predictive models can identify the clinical symptoms and laboratory results that should be tracked most closely during the onset of EVD, and give health care workers the ability to more accurately assess patient risk and therefore manage treatment more efficiently [4]. This data-driven prioritization could lead to higher recovery rates through stratified treatment [5], especially in resource-constrained areas, and would help doctors limit the evaluation of experimental EVD vaccines and treatments [6] with potentially harmful side effects only to highest-risk patients. These improvements in treatment, however, will only be achieved once larger datasets become available to overcome biases resulting from small samples.
Schieffelin et al. [7] presented the only publicly accessible, at the time of publication, clinical dataset from the West African EVD outbreak (available in various formats at http://fathom.info/mirador/ebola/datarelease) to enable clinical investigations. Although a large amount of very useful case and resource data has been made public throughout the outbreak (https://data.hdx.rwlabs.org/ebola), thanks to the efforts of numerous individuals and organizations, there is to our knowledge no other public source offering a similar level of clinical detail. The Schieffelin et al. dataset includes epidemiologic, clinical, and laboratory records of 106 patients treated at Kenema Government Hospital in Sierra Leone during the initial stages of the outbreak. The study also provides a simple heuristic to estimate mortality risk by defining an Ebola Prognostic Score (EPS), which predicts patient outcome based on symptom counts. EPS offers statistically significant differences between surviving and deceased patients with p < 0.001.
While data from other published clinical studies are not available, their summary results suggest that more advanced prognostic prediction models could be potentially useful to the field. Levine et al. [8] developed a diagnostics model using data from the Bong County Ebola Treatment Unit in Liberia, which predicts laboratory-confirmed EVD cases using six clinical variables. Yan et al. [9] carried out a multivariate analysis of 154 EVD patients from the Jui Government Hospital in Sierra Leone, and reported that age, fever, and viral load are independent predictors of mortality, while Zhang et al. [10] recently reported that age, chest pain, coma, confusion, and viral load are associated with EVD prognosis using a set of 63 laboratory-confirmed cases also from the Jui Government Hospital.
In this study, we employed the Schieffelin et al. EVD dataset to develop novel predictive models for patient prognosis, integrating a data-driven hypothesis making approach with a customizable Machine Learning (ML) pipeline, and incorporating rigorous imputation methods for missing data. We evaluated the predictors using a variety of performance metrics, identifying top predictors with and without viral load measurements, and packaged them into a mobile app for Android ad iOS devices (http://fathom.info/mirador/ebola/prognosis).
Our protocol exemplifies how data-driven computational methods can be useful in the context of an outbreak to extract predictive models from incomplete data, and to provide rapidly actionable knowledge to health workers in the field. Moreover, prognosis prediction software could complement ongoing efforts to develop rapid EVD diagnostics [11] and safe data-entry devices [12]. Given the availability of only one dataset from a single location, one Ebolavirus species (Zaire ebolavirus), and very specific time span and laboratory protocols, these models need to be interpreted in an exploratory sense and require further validation with independent clinical data from other EVD treatment sites [8] [9] [13] [14] [15]. We have made all of these resources publicly available and fully documented with the hope to encourage further methods development, independent validation, and greater data sharing in outbreak response.
Our analysis and modeling is based on the EVD clinical and laboratory data initially described by Schieffelin et al [7]. The Sierra Leone Ethics and Scientific Review Committee and the ethics committee at Harvard University have approved the study and public release of this clinical data, which has been de-identified to protect patient privacy. As indicated by Schieffelin, “these committees waived the requirement to obtain informed consent during the West African Ebola outbreak” and “all clinical samples and data were collected for routine patient care and for public health interventions.” The larger dataset comprises 213 suspected cases evaluated for Ebola virus infection at the Kenema Government Hospital (KGH) in Sierra Leone between May 25 and June 18, 2014. Outcome data was available for 87 of 106 Ebola-positive cases, giving a Case Fatality Rate (CFR) of 73% over the entire dataset. We considered 65 patients between 10 and 50 years of age. Within this group, not all individuals had complete clinical chart, metabolic panel, and virus load records available (Fig 1). Sign and symptom data were obtained at time of presentation on 34 patients that were admitted to KGH and had a clinical chart. Metabolic panels were performed on 47 patients with adequate sample volumes, with a Piccolo Blood Chemistry Analyzer and Comprehensive Metabolic Reagent Discs (Abaxis), following the manufacturer’s guidelines. Virus load was determined in 58 cases with adequate sample volumes using the Power SYBR Green RNA-to-CT 1-Step quantitative RT-PCR assay (Life Technologies) at Harvard University. Both metabolic panel and PCR data used to develop our models was collected during triaging of the patients upon admission, and follow-up data, although available for some patients, was not included in our analyses. We compiled this data into a single file in CSV format, and made it available in a public repository (http://dx.doi.org/10.5281/zenodo.14565), together with all original Excel spreadsheets and the cleaning and aggregation scripts (http://fathom.info/mirador/ebola/datarelease), as well as a Dataverse hosted on the Harvard Dataverse Network (http://dx.doi.org/10.7910/DVN/29296).
In a separate effort, we designed the tool Mirador (http://fathom.info/mirador/) to allow users to identify statistical associations in complex datasets using an interactive visualization interface. This visual analysis is guided by an underlying statistical module that ranks the associations using pairwise Mutual Information [16]. Mirador automatically computes a sample estimate of the Mutual Information between each pair of variables inspected by the user, and performs a bootstrap significance test [17] to determine if the variables are independent within a confidence level set through the interface. This calculation relies on an optimal bin-width algorithm [18], which finds the grid minimizing the Mean Integrated Squared Error between the estimates from the data and the underlying joint distributions. The user can then rigorously test the hypothesis of association suggested by Mirador using more specialized tools such as R or SPSS, and finally incorporate them into predictive models. We used the Maximal Information Coefficient (MIC) statistic developed by Reshef et al [19], calculated with the MINE program (http://www.exploredata.net/), to rank the associations found with Mirador.
Since only 21 patients in the dataset contain complete clinical, laboratory, and viral load information, we applied three Multiple Imputation (MI) programs to impute the missing values: Amelia II, which assumes the data follows a multivariate normal distribution and uses a bootstrapped expectation-maximization algorithm to impute the missing values [20]; MICE [21] Multivariate Imputation by Chained Equations, where missing values in each variable are iteratively imputed given the other variables in the data until convergence is attained; and Hmisc [22], which is also based on the chained equations method. All MI methods require that the missing entries satisfy the Missing Completely At Random (MCAR) condition in order to generate unbiased results. Specifically, MCAR means that the distribution of the missing entries is entirely random and does not depend neither on the observed nor the missing values. Furthermore, Amelia requires the observed data to follow a multivariate normal distribution. We used Little’s MCAR chi-square test [23] and Jamshidian and Jalal's test for Homoscedasticity, Multivariate Normality, and MCAR [24] to rigorously test for these conditions.
After testing for the MCAR condition, we run each MI program m times to generate m “completed” copies of the original dataset, which we aggregated into a single training set of larger size (S4 Fig). We performed a detailed comparison of the performance of the predictor when using values imputed by each of the three MI programs, which is described in the results.
The ML pipeline takes as inputs the source data and a list of covariates, and outputs a trained predictor that can be evaluated with several accuracy metrics. It includes the following classifiers: a single-layer Artificial Neural Network (ANN) [25] implemented from scratch, and Logistic Regression (LR), Decision Tree (DT), and Support Vector Machine (SVM) classifiers from scikit-learn [26]. Each classifier was trained on all possible combination of input covariates, from the subset of found with Mirador and MINE, to avoid issues with variable selection methods [27], and to generate an ensemble of predictors that could be applied to different combinations of available clinical data.
We applied multiple cross-validation in order to train the classifiers for each selection of covariates. We first split the records without missing values into two sets with identical CFR, then set one aside for model testing. We combined the second set with the remaining records that include missing values, and used this data as the input for the MI programs. Depending on the percentage of complete records reserved for testing and the number of MIs, we ended up with testing sets of 6–10 cases and training sets of 200–300 cases. This ensured having more than 10 samples per variable during predictor training, the accepted minimum in predictive modeling [28]. We generated 100 of such testing/training set pairs by randomly reshuffling complete records between test set and training set.
Each model was initially ranked by its mean F1-score, which is the weighted average of the precision and sensitivity. The mean and standard deviation were calculated over the 100 cross-validation iterations for each combination of input covariates. We then used the bootstrap method originally introduced by Harrell [29] to quantify the optimistic bias [30] in the area under the receiver operator curve (AUC or c-statistic). We generated 100 bootstrap samples with replacement for each model, and re-trained the model on these samples. We evaluated the AUC on the bootstrap sample and the original sample, and reported the mean of the AUCboot—AUCorig difference as the estimated optimism.
Finally, we carried out standard logistic regression with variable selection, with the goal of evaluating the effect of our MI protocol on other model selection algorithms, and comparing the resulting standard model with the top-ranking models from our pipeline. We used the built-in step() function in R to perform backward variable selection with the Akaike Information Criterion (AIC), the ROCR package to compute AUC, and the Boot package to estimate of the optimistic bias with bootstrap sampling.
The ML models generated by our pipeline are essentially Python scripts together with some parameter files. The Kivy framework (http://www.kivy.org) allowed us to package these scripts as mobile apps that can be deployed on tablets or smartphones through Google or Apple’s app stores. We created a prototype app including the models described in this paper, currently available as Ebola CARE (Computational Assignment of Risk Estimates), shown in Fig 2. We have only implemented the ANN classifier into the Ebola CARE app for the time being, because the scikit-learn classifiers could not be compiled to run on Android devices, which is a requirement for our prognosis app. Once installed, the app is entirely stand-alone, does not require Internet connectivity to run, and can be updated once better models are available.
We began by identifying the clinical and laboratory factors that provide the strongest association with EVD outcome. Earlier reports indicate that EVD mortality rates in this outbreak are found to be significantly different among children [31] and older adults [7], and this pattern holds in our data: CFR is higher than 90% for the 18 patients older than 50 years of age, and 75% for the 14 patients under 10 years of age; we therefore restricted our analyses to patients between 10 and 50 years of age. Within this age range, exploratory analysis with Mirador (http://fathom.info/mirador/), led us to identify 24 clinical and laboratory factors that show plausible association with EVD outcome: virus load (PCR), temperature (temp), aspartate aminotransferase (AST), Calcium (Ca), Alkaline Phosphatase (ALK), Chloride (Cl), Alanine Aminotransferase (ALT), Creatinine (CRE), Total Carbon Dioxide (tCO2), Albumin (Alb), Blood Urea Nitrogen (BUN), Total Protein (TP), weakness, vomit, edema, confusion, respiratory rate, back pain, dizziness, retrosternal pain, diarrhea, heart rate, diastolic pressure, and abdominal pain. Boxplots and histograms for all factors are depicted in S1 and S2 Figs, which also presents the P-values for the association between Outcome and each factor, for the Fisher exact and T-tests (for nominal and numerical factors, respectively).
We applied the Maximal Information Coefficient (MIC) statistic developed by Reshef et al. [19], calculated with the MINE program (http://www.exploredata.net/), to rank these 24 factors. We used the ranking to select two informative subsets of 10 variables each (shown in Fig 3), one with PCR and the other without, by picking the top 5 laboratory results and top 5 clinical chart variables. The PCR set comprises PCR, temp, AST, ALK, CRE, tCO2, heart rate, diarrhea, weakness, and vomit, while the non-PCR set includes temp, AST, ALK, CRE, tCO2, BUN, heart rate, diarrhea, weakness, and vomit. None of these variables are capable of predicting outcome accurately in isolation. The performance of the univariate LR classifier is highest with PCR as input, with an F1-score of 0.67, and below 0.5 for all other variables. This result is consistent with the recent report from Crowe et al. [32], which highlights the importance of viral load in the prognosis of EVD.
We evaluated the impact of the MI step on the predictors’ performance, and chose MI parameters accordingly. In all three MI modules, Amelia II, MICE and Hmisc, we can adjust the fraction of complete records to be included in the data to impute, as well as the number of imputed copies that are aggregated into a single training set. We considered all combinations of these two parameters, when allowing 20%, 35%, and 50% as the percentages of complete records used during imputation, and 1, 5 and 10 for the number of imputed copies. We examined the resulting 9 combinations of parameters across the 4 predictors, LR, ANN, DT, and SVM. Accuracy, as measured by mean F1-score, in the PCR case does not seem to depend on the number of imputed copies, percentage of completed records, and MI algorithm (S3A Fig). In contrast, both higher percentage of completed records and higher number of imputed copies do have a definite enhancing effect in the mean F1-score for the non-PCR case (S3B Fig), while the choice of MI algorithm does not seem to have a significant impact. Counter intuitively, the standard deviation of the F1-score in the PCR case increases with larger percentage of completed records. However, this trend can be explained as follows: the complete records not included in the training set are used to construct the testing set, therefore higher percentages of complete records used during MI result in smaller testing sets. The effect of a single false positive or negative is proportionally larger in smaller testing sets than in larger ones, which results in higher variation of the F1-score in the latter.
We then verified the validity of the MCAR condition in both the PCR and non-PCR sets, crucial to guarantee unbiased imputations, using Little’s chi-square test and Jamshidian and Jalal’s test. Since the all data used in our models was collected at presentation, there is lower risk of non-random missing patterns due to patient death and withdrawal. The tests for MCAR indeed confirm this: Little’s statistic takes a value of 45.28 with a P-value of 0.11 on the PCR set, while the non-PCR set gives a statistic value of 19.06 with a P-value of 0.32, meaning that in both cases there is no evidence in the data against the MCAR hypothesis at the 0.05 significance level. Furthermore, Jamshidian and Jalal’s test for Homoscedasticity, Multivariate Normality, and MCAR does not reject the multivariate normality or MCAR hypothesis at the 0.05 significance level for both the PCR and non-PCR sets, with P-values of 0.79 and 0.06, respectively. This last result in particular validates the use of the Amelia II package, which assumes that the data follows a normal distribution.
Based on these findings as well as on a published review from Horton et al. [33], which shows a marginal improvement with Amelia over the other MI methods, we chose Amelia II as the default MI method. One weakness of the Amelia II program is that combinations of variables that are highly collinear might cause the MI computation to fail to converge. We addressed this problem by re-running the MI using either MICE or Hmisc when Amelia is detected to fail converging more than 5 times. We generated out training sets with 50% of the complete records in the data to impute, and 5 imputed copies for aggregation into a single training set. The performance difference between 5 and 10 imputed copies did not seem large enough to justify the increased computing times.
Having developed and carefully evaluated our models, we demonstrate that we are able to predict EVD prognosis with a mean F1-score of 0.9 or higher, for EVD patients aged 10 to 50. We arrived at this by exhaustively generating two separate ensembles of predictors, one with PCR data and the other without.
The predictors including PCR data are plotted on a scatter plot of the mean F1-score vs standard deviation (Fig 4a) computed over 100 rounds of cross-validation for each predictor. The ensemble consists of 4 × (29–1) = 2044 predictors (LR, ANN, DT, SVM) that were trained on all combinations of the PCR set (9 variables), having PCR as a fixed input variable. The LR and ANN classifiers are the best performers over all the four prediction methods, with 156 models (71 ANN, 64 LR, 21 SVM) yielding an F1-score of 0.9 or higher.
Similarly, we generated 4 × (210–11) = 4052 predictors without PCR data (Fig 4c), which were trained on all combinations of the non-PCR set of variables (10 variables) with at least two elements. We obtained 45 models (18 ANN, 24 LR, 3 SVM) with a mean F1-score of 0.9 or higher. A number of the variables emerged as those most often included in the top-ranked models, both in the PCR and non-PCR cases respectively (Fig 4b and 4d). Notably, in addition to temperature, CRE, ALK, and tCO2 levels are consistently present in the predictors including PCR, while the lack of PCR data makes AST levels and the onset of diarrhea more relevant for accurate prognosis.
The optimistic bias of the AUC for the top predictors, both in the PCR and non-PCR cases, is below 0.01 for most of them, with a standard deviation of 0.03 (Fig 5a and 5b). This analysis indicates that even though our models are over-fitted for the current data, the magnitude of bias is minor. S1 and S2 Tables detail all the top-performing predictors and their optimism-corrected AUC scores in the PCR and non-PCR cases, respectively. S5 Fig shows aggregated ROC curves over all the models for each predictor, for the PCR and non-PCR cases. The aggregated AUCs are 0.96 (LR), 0.95 (ANN), 0.94 (SVM), and 0.84 (DT) in the PCR models, and 0.88 (LR, ANN), 0.86 (SVM), and 0.77 (LR) in the non-PCR models. The similar performance of our simple ANN predictor and scikit-learn’s LR classifier suggests that the dependency between the covariates and outcome can be modeled linearly, however larger datasets would enable us to train more complex ANNs with potentially better performance across different groups of patients.
The comparison with variable selection shows an effect of the MI protocol similar to that observed in the top-ranked models. The optimistic bias of the AUC for the selected PCR and non-PCR models consistently decreases to less than 0.01 as the number of imputed copies increases from 1 to 5 (Fig 5c and 5d). On the other hand, these models assign very small coefficients and odd ratios very close to 1 to the laboratory covariates (Tables 1 and 2). This suggests that most of the information in these models is captured by the clinical symptoms (temperature, diarrhea, vomit), although weakness consistently presents an odd ratio less than 1, contradicting the expected dependency with outcome. In general, the laboratory variables are the highest ranked according to MIC, and are also included in most of the top-ranked models, using either the LR or ANN classifiers. These results lead us think that the variable selection approach is discarding relevant information for outcome prediction, which we are able to capture in our ensemble of ML predictors.
The Ebola CARE app packages a total of 82 ANN models, selected from those with a mean F1-score above 0.9, but discarding the models with a standard deviation of 0, in order to avoid potentially overfitted models. This set incorporates 64 PCR and 18 non-PCR models, so the app can still be used when viral load information is not available. We entered into Ebola CARE all the patients who had complete data for at least one model in the app, and recorded the risk prediction as presented after inputting the symptoms. Predictions for a total of 34 patients were obtained in this way. For this subgroup of patients, the mortality rate was 79% (7 survived, 27 died), and the app only misclassified two, one in each outcome group. In other words, the precision and sensitivity were both 0.96. However, this number is likely overestimating the performance of the app, since some of these patients used in this test were also included in model training.
The data used in this study is hosted at a Dataverse in the Harvard Dataverse Network (http://dx.doi.org/10.7910/DVN/29296), the source code of Mirador and the ML pipeline is available on Github (https://github.com/mirador/mirador, https://github.com/broadinstitute/ebola-predictor), and the model files (all training and testing sets) are deposited on Zenodo (http://dx.doi.org/10.5281/zenodo.19831).
This work represents the first known application of ML techniques to EVD prognosis prediction. The results suggest that a small set of clinical symptoms and laboratory tests could be sufficient to accurately prognosticate EVD outcome, and that these symptoms and tests should be given particular attention by health care professionals. By aggregating all the high-performing models obtained in our exhaustive analysis, we can construct a composite algorithm that runs the best predictor depending on the available data. We have developed a simple app, Ebola CARE, which can be installed on mobile tablet or phone devices, and would complement rapid EVD diagnostic kits and data-entry devices.
Our Ebola CARE app is a proof-of-concept, only applicable to Ebola Zaire patients treated in similar conditions as those in KGH. New clinical data will enable us and other groups to independently validate the app, and to generate more generalizable models with higher statistical significance. Within the current constrains, the results also shed light on the most informative clinical predictors for adult patients -temperature, diarrhea, creatinine, alkaline phosphatase, aspartate aminotransferase, total carbon dioxide- and demonstrate that PCR provides critical additional information to quantify the seriousness of the Ebola virus infection and better estimate the risk of the patients. In general, these results are consistent with the findings from Schieffelin, Levine, Yan, and Zhang. Current discrepancies–for instance Zhang reports chest pain, coma, and confusion as significantly associated with EVD prognosis whereas we do not–could be attributed to the small sample sizes, missing data, and different clinical protocols at the various treatment sites.
The prevalence of missing data in the dataset used in this study, and the lack of other publicly available datasets, are fundamental challenges in predictive modeling. By combining MI with four distinct ML predictors, we offer a direct approach for dealing with the first challenge. The use of ANN and LR classifiers in combination with a MI enrichment methodology shows promise as a way to accurately predict outcome of EVD patients given their initial clinical symptoms and laboratory results.
New patient data is critical to validate and extend these results and protocols. Richer datasets incorporating more diverse samples from different locations will allow us and other researchers to train better ML classifiers and to incorporate population variability. The development of survival models could be another very important application of these techniques to assist not only in prognosis upon patient intake but also during treatment, as shown by Zhang. Our current data includes time courses that would be useful in this kind of models, but unfortunately only for a handful of patients. All these facts highlight the importance of immediate availability of clinical data in the context of epidemic outbreaks, so that accurate predictive tools can be quickly adopted in the field.
In summary, we have made our protocol and mobile app publicly available, fully documented (https://github.com/broadinstitute/ebola-predictor/wiki), and readily adaptable to facilitate and encourage open data sharing and further development. Our integration of Mirador, a tool for visual exploratory analysis of complex datasets, and an ML pipeline defines a complete framework for data-driven analysis of clinical records, which could enable researchers to quickly identify associations and build predictive models. Our app is similarly designed to be easily updated as new predictive models are developed with our pipeline, validated with better data, and packaged, to generate actionable diagnosis and help inform urgent clinical care in outbreak response.
|
10.1371/journal.pcbi.1003943 | Multitask Learning of Signaling and Regulatory Networks with Application to Studying Human Response to Flu | Reconstructing regulatory and signaling response networks is one of the major goals of systems biology. While several successful methods have been suggested for this task, some integrating large and diverse datasets, these methods have so far been applied to reconstruct a single response network at a time, even when studying and modeling related conditions. To improve network reconstruction we developed MT-SDREM, a multi-task learning method which jointly models networks for several related conditions. In MT-SDREM, parameters are jointly constrained across the networks while still allowing for condition-specific pathways and regulation. We formulate the multi-task learning problem and discuss methods for optimizing the joint target function. We applied MT-SDREM to reconstruct dynamic human response networks for three flu strains: H1N1, H5N1 and H3N2. Our multi-task learning method was able to identify known and novel factors and genes, improving upon prior methods that model each condition independently. The MT-SDREM networks were also better at identifying proteins whose removal affects viral load indicating that joint learning can still lead to accurate, condition-specific, networks. Supporting website with MT-SDREM implementation: http://sb.cs.cmu.edu/mtsdrem
| To understand why some flu strains are more virulent than others, researchers attempt to profile and model the molecular human response to these strains and identify similarities and differences between the resulting models. So far, the modeling and analysis part has been done independently for each strain and the results contrasted in a post-processing step. Here we present a new method, termed MT-SDREM, that simultaneously models the response to all strains allowing us to identify both, the core response elements that are shared among the strains, and factors that are uniquely activated or repressed by individual strains. We applied this method to study the human response to three flu strains: H1N1, H3N2 and H5N1. As we show, the method was able to correctly identify several common and known factors regulating immune response to such strains and also identified unique factors for each of the strains. The models reconstructed by the simultaneous analysis method improved upon those generated by methods that model each strain response separately. Our joint models can be used to identify strain specific treatments as well as treatments that are likely to be effective against all three strains.
| The relative ease of high-throughput data collection enables profiling a system of interest in many ways with complementary assays, at different times, and under various perturbations to compare and contrast the outcomes. The resulting computational challenge is to develop analysis strategies that maximally leverage these related experiments to improve our ability to reconstruct biologically accurate models.
Even when applied to study the same condition, different types of high-throughput data (e.g., functional genetic screens and gene expression) often times implicate largely disjoint groups of genes or proteins because each experiment highlights different facets of the biological processes and networks involved [1]. Consequently, there has been extensive research to develop techniques for integrating one or more types of condition-specific high-throughput data with general purpose physical interaction networks, such as protein-protein interactions (PPIs), to reconstruct signaling and regulatory networks [1]–[3] (see [4] for a review). These methods discern how the genes identified in complementary types of experiments relate to one another in a network context and propose new condition-specific regulators that are not directly observed to be relevant in the original data but form connections in the inferred networks.
Due to the dynamic nature of biological systems, especially those controlling stimulus response and development, it is critical to observe genome-wide changes over time [5]. As reviewed in [5], there are now computational approaches that exploit the unique structure in temporal datasets (e.g., time series gene expression) to model dynamic processes and reverse engineer regulatory networks [6], [7]. Recent algorithms integrate temporal data and PPI networks to improve signaling pathway prediction by capitalizing on the dynamic information [8], [9].
Despite advances in modeling the temporal dimension and different types of assays per condition, there has been considerably less progress made for datasets that contain multiple related perturbations or stimuli. Typically each condition is analyzed in isolation, and a post-processing comparison of the independent models is required to draw conclusions across conditions [9], [10]. Individual models of related conditions are required to appreciate the unique aspects of each, but building these models independently ignores that the observations may be generated from structurally similar networks. As an example, consider the case of host gene expression following virus infection. Although different viruses do not have identical effects on the host (hence the gene expression patterns are unique to each virus), they also commonly affect a similar core set of host proteins. These include Toll-like receptors (TLRs), which recognize a large number of RNA viruses and activate a downstream pathway that leads to common expression response [11], [12], and other elements of innate immune response pathways [13]. Similarly, in yeast several different types of stresses activate a large common set of genes (termed the environmental stress response genes [14]), and additional examples abound in other species.
When modeling such responses, one may be able to take advantage of these commonalties without sacrificing the ability to reconstruct individual models for each response. This type of machine learning is termed Multi-task learning [15] and usually applies to cases where one learns models for different problems that share information and/or parameters. A key advantage of such framework is the ability to utilize additional data from related conditions when reconstructing networks for a specific response. This is especially important when reconstructing biological response networks from high-throughput data because the number of parameters to fit is very large relative to the number of samples. In addition, extensive data from a well-characterized condition may be able to compensate for sparse data in a similar, less-understood condition.
Multi-task learning has been applied to other problems in the biological domain including classification [16], genome-wide association studies [17], [18], protein structure [19], and pairwise protein-protein interaction prediction [20], [21]. Multi-commodity flow [22] and iterative applications of a prize-collecting Steiner forest algorithm [23] have been used to simultaneously reconstruct related response or disease networks, but these methods do not employ multi-task learning. In addition, these previous approaches operate on static data and cannot account for the dynamic behaviors that are crucial for understanding many types of stimulus responses.
Here we present the Multi-Task Signaling and Dynamic Regulatory Events Miner (MT-SDREM), which uses multi-task learning to reconstruct response pathways and temporal regulatory networks. MT-SDREM is equipped to capitalize on the many dimensions in complex systems biology datasets by integrating different types of experimental data in each condition, explaining the time-dependent elements of a response (as observed in gene expression data), and constraining the inferred networks to be similar for related conditions or perturbations. Like its single-condition predecessor [8], MT-SDREM iterates between finding pathways that connect the upstream proteins that directly interact with an external stimulus (called source proteins) and the downstream transcription factors (TFs) that regulate the response and learning dynamic regulatory networks activated by these TFs. The learning process involves the simultaneous reconstruction of several such networks. While a different network is learned for each condition, the joint learning framework allows sharing and/or constraining parameters across the different networks which helps overcome the overfitting problem that is often an issue when reconstructing biological networks.
We demonstrate how MT-SDREM can be used to gain insights into a clinically-relevant problem: characterizing the human response to viral infection. In particular, we explore the differences between mild, seasonal strains of the influenza A virus, which are typically H1N1 or H3N2 strains [24], and lethal, pandemic strains such as the H1N1 1918 Spanish flu and highly pathogenic avian H5N1 strains. Influenza A strains are subtyped and named by their hemagglutinin (HA) and neuraminidase (NA) proteins. Although there are presently 18 known HA subtypes and 11 NA subtypes [25] only a fraction of these have have infected humans. Previous studies have characterized some of the differences between seasonal and pathogenic strains. Seasonal H1N1 and H3N2 and highly pathogenic H5N1 influenza strains infect macrophages at similar rates, but H3N2 and H5N1 causes apoptosis more rapidly than H1N1 [24]. H1N1 also lead to weaker induction of MAPK signaling pathways than the H3N2 and H5N1 strains [24]. Genomic comparisons of human and avian influenza strains identified 52 species-associated positions that could potentially enable an avian strain to cross over to humans if mutated [26]. Influenza strains also vary in the cells and tissues they infect [27], [28] with highly-virulent strains causing more widespread inflammation, including in the alveoli [27]. Highly pathogenic strains have been shown to induce a stronger inflammatory cytokine response than seasonal influenzas [28] and the host inflammatory response is often more deadly during infection than the pathogen itself [29]. However, much remains unknown about the host factors that are required for viral replication or to mount cellular defenses.
We study three strains of the influenza A virus — seasonal H1N1, seasonal H3N2, and highly pathogenic avian H5N1 — to explain how common host proteins react to the viral infection in a similar manner despite the differences in the temporal transcriptional programs that are activated. The MT-SDREM networks identified many known regulators of influenza response and also suggested putative novel regulators. Because the responses are jointly modeled using the multi-task setting, MT-SDREM is able to correctly recover TFs that are important drivers of the immune response that are missed when each viral strain is analyzed independently [10] and by previous methods for combining gene expression data across experiments. In addition, MT-SDREM networks are more predictive of host genes that are required for viral replication, a potentially clinically-relevant phenotype [29], than corresponding independent models or gene prioritization algorithms.
MT-SDREM simultaneously infers signaling and dynamic regulatory networks for multiple related conditions. It extends the SDREM tool [8], [10] which discovers signaling pathways by orienting edges in protein interaction networks. To demonstrate the performance of such multi-task network learning we looked at data from 3 different flu viruses: H1N1, H3N2, and H5N1.
For each of these viruses we obtained time series gene expression measurements of cells infected with the virus. For H1N1 the data is from [30] and contains 10 time points. For H5N1, we obtained data from [31] with 5 time points, and the H3N2 data from [32] had 6 time points. In addition, for each of these viruses we obtained a set of sources (host proteins interacting with the virus proteins) from mass spec experiments. Data for H1N1 is from [33] and literature [30], [34] and contains 200 human proteins that were experimentally determined to interact with H1N1 proteins. Data for H3N2 is from [33] and consists of 153 host proteins and source data for H5N1 is from [33] and literature [34]–[39] and consists of 41 sources.
To test the advantages of multi-task learning we compared MT-SDREM with previous methods that can be used to analyze expression and interaction data. Since we are not aware of prior methods that utilize multi-task learning in biological network reconstruction we first looked at the differences between applying MT-SDREM and applying SDREM separately to each of the three flu datasets. We have also compared MT-SDREM's results to a baseline joint ranking of differentially expressed (DE) genes from different experiments in a single analysis. This approach is similar to several previous studies that perform follow up analysis using such joint sets [52].
Since the 'ground truth' (complete underlying networks for each condition) is obviously unknown, we used three different types of complementary information for these comparisons. First, we examined the set of TFs identified by each of these methods and determined their relevance to the condition being studied. Next, we used the Gene Ontology (GO) to test the differences in the identified functional categories between the different analysis methods. Expression experiments and RNA interference (RNAi) screens have revealed a multitude of host pathways and processes that are involved in viral host response including MAPK signaling, apoptosis, trafficking, mRNA export, splicing, and proteolysis [30], [53], [54]. A statistical meta-analysis implicates nearly 3000 host genes [55] in these pathways. Although many processes as a whole are relevant to influenza response, not all genes participating in those processes necessarily are important. Therefore we focused our TF and GO evaluation on immune processes, which were shown to compose a critical component of the host response that kills infected cells, protects uninfected cells, combats viral components, and promotes inflammation [56] Finally, we used a set of RNAi experiments that were performed for H1N1 and H5N1 to test the ability of these different methods to identify key disease related proteins. In these experiments proteins are knocked down one at a time and the impact on viral load is measured. A protein affecting viral load is likely participating in the host response and so methods that can identify such proteins more accurately are in better agreement with the observed response. The RNAi data for H1N1 was obtained from [30], [53], [54], [57], [58] resulting in a total of 980 screen hits, 925 of which were present in our initial interaction network (which contained 16671 genes, Methods). 32 screen hits for H5N1 were obtained from [57], all of which are present in our interaction network.
In Table 1 we present the overall and pairwise overlap of the inferred TFs for the 3 conditions (extracted by same mechanism as in SDREM [8], [10]) for MT-SDREM and compare it to when SDREM is run independently on the 3 conditions (I-SDREM). Note that the pairwise intersections shown are in addition to the overall intersection between all of the 3 conditions.
The shared TFs identified by MT-SDREM among all 3 conditions that are missed by I-SDREM include several that are known to be immune response related. In particular, CEBPA is known to be responsible for regulating a large variety of cell functions including immune and inflammatory response [59]. MT-SDREM also identifies SMAD4 in all three conditions. SMAD family proteins are part of the TGF pathway as mentioned above. MT-SDREM also identifies RB1 which has been implicated in viral immune response [60], JUN which is part of the AP-1 TF complex, and PPARG an important TF regulating immune response mentioned above. In contrast, I-SDREM does not identify any TF in the intersection that MT-SDREM does not.
In addition, we also find several immune response related TFs in the pairwise overlaps for MT-SDREM that we do not see for I-SDREM. For the overlap between H1N1 and H3N2, MT-SDREM identifies IRF1/3/5 which are known to regulate interferons and thus important for immune response. For the overlap between H1N1 and H5N1, MT-SDREM finds the the STAT3 gene which is part of the JAK-STAT signaling pathway and ATF2, part of the AP-1 TF complex.
For the pairwise intersection of H1N2 and H3N2, I-SDREM identifies NR3C1 as a TF while MT-SDREM only selects it as an intermediate (signaling) protein. It also identifies another member of the SMAD family (SMAD3 whereas MT-SDREM identifies SMAD4). For H3N2 and H5N1 it identifies AHR whose activation inhibits inflammation [61] and RELA in the intersection of H1N1 and H5N1, which as part of the NF- complex.
We also compared MT-SDREM to the popular TF prediction tool oPossum [62]. Our primary goal when comparing MT-SDREM with oPossum is to highlight the fact that using network information in the multi-task learning framework is useful. The input to oPossum is a list of genes identified by the experiment(s) and using this list it attempts to find overrepresented TF-binding sites. To select a common gene list from all three experiments we ranked the genes for each condition according to their differential expression and then merged the 3 rankings using the Kemeny-Young method [63]. Similar to the number of genes used by MT-SDREM we used the top 3000 in the joint ranking as input to oPossum. In Table 2 we present the comparison. Note that since we used oPossum as the tool for the comparison of MT-SDREM with other methods for integrating data from several conditions, the results shown for Table 2 are different from the intersection results of Table 1. Here, for the MT-SDREM rankings we used the sum of % path flow going through each gene across the 3 networks to rank TFs (Methods). The oPossum TFs are ranked according to their Z-score.
While oPossum is able to identify a few relevant TFs, for most of the TFs identified by oPossum, we could not find significant roles in immune response regulation for them. In contrast, several of the shared MT-SDREM TFs that are not identified by oPossum are known to play major roles in immune response as discussed above. These include STAT1/3, JUN/ATF2, CEBPA/B which regulate a large number of immune response genes, RB1 which has been implicated in viral immune response networks [60], PPARG, and SMAD. MT-SDREM also uniquely identifies IRF1 which plays a major role in viral immune response by regulating interferons. oPossum was able to identify only two relevant TFs that were not found by MT-SDREM. These are ZEB1 which regulates the IL2 interleukin, part of the immune response system and AHR, part of the ANTR-AHR complex. See also Tables S13–S15 in S1 Text for condition-specific comparisons using oPossum.
We also tried to compare MT-SDREM with the Inferelator method [6] but following email discussions with the authors of that method determined that such comparison is not feasible since Inferelator requires expression data for a large number of conditions while we only had time series response for three types of infections.
To compare the GO enrichment of shared genes/proteins we examined the top 500 genes in the combined MT-SDREM network (ranked using the same sum of % of path flow going through genes across the 3 networks as we did for the oPossum comparison) with the top 500 genes from the combined ranking of the differentially expressed (DE) genes from each condition (combined using the Kemeny-Young method as explained before). We used FuncAssociate [64], [65] to compute standard GO enrichment for the genes. We found 3 categories, only 2 of which were immune response related for which the p-value for DE genes was but which were not present in the MT-SDREM list or if present, their p-value was . The categories are listed in Table 4. However, for the vice versa comparison, we found a large number of categories for which the MT-SDREM p-value was but which were either not enriched for in the DE genes list (most common outcome) or if present, their p-value was . A subset of the immune response related categories are listed in Table 3. Note that we find significant enrichment for a very varied set of immune response processes including T cell activation, cytokine production, activation of immune response, etc. as well as categories related to viral genome expression and positive regulation of viral process. The DE genes list is only enriched for negative regulation of viral process and viral genome replication. The complete set of the categories is in S45 Table.
To further compare methods that are based on joint expression analysis to those that are based on joint network learning we looked at the GO enrichment for the top 50 TFs identified by MT-SDREM and oPossum. The top 50 TFs for MT-SDREM are ranked using the joint ranking based on path flow for the 3 conditions as done for the GO comparison above. We used the TF Z-score provided by oPossum to rank TFs for oPossum. We again used FuncAssociate [64], [65] to compute standard GO enrichment for the TFs. We obtained only one immune-response related category (interleukin related) for which the p-value for the oPossum TF set was while being for MT-SDREM (presented in Table 6). However we obtained 270 categories in total for which the MT-SDREM p-value was but the p-value for oPossum was , a large number of which were immune response related. Due to space constraints, only a subset of these are presented in Table 5. These include 'postive regulation of innate immune response', 'viral process', and 'cytokine-mediated signaling pathway'. The complete list of categories is in S46 Table. See also Tables S14–S25 in S1 Text for several additional comparisons of MT-SDREM and other methods using GO enrichment data.
Using the screen hit data for H1N1 and H5N1 we compared the performance of MT-SDREM, I-SDREM and Endeavour [66], [67]. Endeavour is a gene prioritization algorithm which uses a set of seed genes (the sources) to rank genes based on several types of evidence including gene expression, interaction networks derived from various sources, text mining, sequence similarity, and functional annotations. It combines the individual rankings to create a global ranking for all genes. For the MT-SDREM and I-SDREM results we ranked proteins based on the total number of paths weighted by their score going through them. See Supplementary Methods in S1 Text for details. For Endeavour, we configured it to use only BioGRID and HPRD as data sources as those are the only sources we use to construct our PPI network. The expression data is not used by Endeavour. We gave the source proteins as the seed genes to Endeavour. We further compared these three methods with a baseline method that is condition-independent: ranking nodes by their weighted degree in the PPI network. The results are presented in Figure 3. For H1N1, the top 100 genes in the Endeavour ranking include only 20 screen hits (p-value is 4.9E-7). For I-SDREM the number increases to 35 (p-value 2.0E-19) whereas MT-SDREM obtains the highest number of protein in the overlap 39 (p-value 1.7E-23). The baseline comparison where we rank by degree has an overlap of 30 genes (p-value 9.4E-15). For H5N1, the top 100 genes for Endeavour and for ranking by degree include only 5 screen hits (p-value 1.2E-6) whereas both I-SDREM and MT-SDREM have an overlap of 9 screen hits (p-value 1.7E-13). See also S1 Text for comparison of RNAi screen hits using GSEA.
We also compared MT-SDREM, I-SDREM with GeneMania [68], [69] and concluded that MT-SDREM greatly improves upon the GeneMania results. See Supplementary Results in S1 Text for details.
We developed MT-SDREM a multi-task learning framework that simultaneously reconstructs signaling and dynamic regulatory networks across related conditions. Given the small number of condition-specific samples that are often available (i.e. time series expression data and host-pathogen interaction data) sharing parameters across related conditions allows the reconstruction of more accurate networks while still retaining the ability to explain condition-specific signaling and regulation.
We applied MT-SDREM to reconstruct networks for 3 related influenza A virus infections – H1N1, H3N2, and H5N1. The resulting signaling and regulatory networks were able to identify several known and novel regulators of immune and viral response. Many of these were shared between all condition including PPARG, FOS, ATF, and JUN. Similarly, we identify key signaling proteins, some shared by all conditions while others are unique to one or two of the conditions. Specifically, we identified the signaling protein SUMO1 as part of pathway from UBE2I for all 3 conditions. This agrees with recent findings that UBE2I interacts with SUMO1 to degrade influenza A's virus, NS1 which is present in all three strains [70]. We also identified the AKT1 gene, part of the PI3K/AKT pathway that is activated by NS1 in all conditions.
MT-SDREM is the first method to jointly reconstruct such dynamic networks. Comparing MT-SDREM with methods that have been suggested to integrate gene expression data or with methods reconstruct such networks independently for each condition highlighted the advantages of multi-task network learning. MT-SDREM outperformed previous methods in identifying a set of TFs controlling immune response, a set of functionally relevant proteins and a set of proteins whose knockdown affects viral loads.
While MT-SDREM can successfully utilize experiments from similar conditions to reconstruct signaling and regulatory networks, there are still issues we would like to improve in future work. One direction we intend to explore is extending MT-SDREM to allow time based (as opposed to global) sharing of TFs across conditions so that splits representing the same time will be more likely to share TFs compared to other splits. We would also like to improve on the models by using additional types of data, including epigenetic data which can help improve the priors for TF binding at specific time points by making them a function of the epigenetic code.
MT-SDREM simultaneously investigates and infers regulatory networks and signaling pathways for several biologically related conditions. For this, it uses both condition-specific gene expression and interaction data and general interaction data. We first discuss the input data that the method utilizes and then present the modeling and learning frameworks.
We use to denote the set of conditions that are jointly modeled by MT-SDREM. Below we list the datasets used by MT-SDREM.
MT-SDREM extends the Signaling and Dynamic Regulatory Events Miner (SDREM) which has so far only been applied to reconstruct response networks for a single condition at a time [8]. Prior to discussing the multi-task learning procedures we first briefly discuss the SDREM method. SDREM is an iterative procedure that combines regulatory and signaling network reconstruction to model response pathways. For the regulatory part, SDREM uses time series gene expression data with protein-DNA interaction data to identify bifurcation events in a time series (places where the expression of previously co-expressed set of genes diverges – see Figure 2), and the transcription factors (TFs) controlling these split events. While some TFs are transcriptionally activated, others are only activated post-translationally via signaling networks. To explain these TFs, the second part of SDREM links sources (host proteins that directly interact with the virus/treatment) to the TFs determined to regulate the regulatory network. This part of SDREM uses protein-protein interaction (PPI) and protein modification data to infer such pathways – while imposing the constraint that the direction of PPI in the inferred pathways is consistent. These two parts (regulatory and signaling reconstruction) iterate a fixed number of times until the final network is obtained. See [8] for complete details.
Following the multi-task learning procedure we arrive at directed, weighted networks for each of the conditions being studied. To further select the key proteins from each of these networks we rank the proteins based on the "path flow" going through a node. The path flow through a node is defined as follows –where is the set of paths containing node .
To combine the rankings from each condition into a single ranking, we compute the total flow through all the nodeswhere is the set of genes and is the condition and then we computed the % flow through a node. To get the combined score for a gene across conditions, we sum up the condition-specific % flows to get where is the number of conditions. Then we rank the genes in descending order of the final score .
|
10.1371/journal.pgen.1004203 | Cleavage Factor I Links Transcription Termination to DNA Damage Response and Genome Integrity Maintenance in Saccharomyces cerevisiae | During transcription, the nascent pre-mRNA undergoes a series of processing steps before being exported to the cytoplasm. The 3′-end processing machinery involves different proteins, this function being crucial to cell growth and viability in eukaryotes. Here, we found that the rna14-1, rna15-1, and hrp1-5 alleles of the cleavage factor I (CFI) cause sensitivity to UV-light in the absence of global genome repair in Saccharomyces cerevisiae. Unexpectedly, CFI mutants were proficient in UV-lesion repair in a transcribed gene. DNA damage checkpoint activation and RNA polymerase II (RNAPII) degradation in response to UV were delayed in CFI-deficient cells, indicating that CFI participates in the DNA damage response (DDR). This is further sustained by the synthetic growth defects observed between rna14-1 and mutants of different repair pathways. Additionally, we found that rna14-1 suffers severe replication progression defects and that a functional G1/S checkpoint becomes essential in avoiding genetic instability in those cells. Thus, CFI function is required to maintain genome integrity and to prevent replication hindrance. These findings reveal a new function for CFI in the DDR and underscore the importance of coordinating transcription termination with replication in the maintenance of genomic stability.
| DNA damage occurs constantly in living cells and needs to be recognized and repaired to avoid mutations. DNA repair is particularly relevant for lesions occurring in actively transcribed DNA strands because the RNA polymerase cannot proceed through a damaged site. Stalled RNA polymerases and persisting DNA lesions can lead to genome instability or cell death. Specific mechanisms to repair obstructing DNA lesions are found from bacteria to higher eukaryotes, their malfunction leading to severe genetic syndromes in humans. Termination of transcription comprises cleavage and polyadenylation of the nascent transcript and displacement of the RNA polymerase from its DNA template. These processes, which are crucial for cell viability and growth in eukaryotes, require two major multi-subunit complexes in budding yeast. Here, we found that one of these complexes, Cleavage Factor I (CFI), participates in the cellular response to DNA damage. In addition, we found that CFI dysfunction leads to replication defects, conceivably mediated by stalled RNA polymerases, rendering cell cycle checkpoints mandatory to prevent genomic instability. Our findings emphasize the importance of coordinating transcription termination, DNA damage response and replication in the maintenance of genomic stability suggesting that CFI plays a fundamental function in the coupling of these processes.
| All cells are continuously exposed to DNA damaging agents, which can arise from exogenous sources or from endogenous metabolic processes. The DNA damage response (DDR) includes the activation of checkpoints and induction of DNA repair pathways. DNA lesions can generate structural distortions that interfere with basic cellular functions like transcription and replication. Such helix-distorting DNA lesions are generally handled by nucleotide excision repair (NER), which can be divided into global genome repair (GG-NER) and transcription-coupled repair (TC-NER) sub-pathways, depending on whether the DNA lesion is located anywhere in the genome or on the transcribed strand (TS) of an active gene, respectively. At transcribed genes, TC-NER acts when elongating RNA polymerase (RNAP) stalls at bulky DNA lesions such as UV-induced cyclobutane pyrimidine dimers (CPDs) (reviewed in [1], [2]). Transcription down-regulation and proteasome-mediated degradation of engaged RNAPII take place as part of the DDR to UV-induced damages [3], [4]. In humans, defects in TC-NER are responsible for two severe genetic disorders called Cockayne Syndrome (CS) and UV Sensitivity Syndrome (reviewed in [5], [6]). In S. cerevisiae, the major TC-NER factor is Rad26, the yeast homologue of CS protein B (CSB) [7]. However, residual TC-NER activity remains in the absence of Rad26, indicating that other factors are also involved in the process [7], [8]. Mutations in several transcription and messenger ribonucleoprotein (mRNP) biogenesis factors including the RNAPII subunit Rpb9, THO, THSC/TREX-2, Paf1, and Ccr4-NOT are partially defective in TC-NER in yeast [9]–[12].
During the past few years it has become clear that the different mRNA processing steps (including 5′-end capping, splicing, and 3′-end cleavage), mRNP export, and transcription are connected to each other (reviewed in [13]) and that surveillance mechanisms ensure that these processes occur in a coordinated manner (reviewed in [14]). THO and THSC/TREX-2 both work at the interface between transcription elongation, mRNP biogenesis and export and defects are characterized by a strong transcription-dependent hyperrecombination phenotype (reviewed in [15], [16]). THO might also act in the process of transcription termination, as in vitro assays suggest that THO mutants lead to polyadenylation defects [17]. Interestingly, other factors required for proficient TC-NER also function during transcription termination. The Paf1 transcription elongation factor contributes to the recruitment of 3′-end processing factors necessary for accurate transcription termination (reviewed in [18]). The Ccr4-NOT complex acts, among other gene expression functions, during transcription elongation and interacts with mRNP export factors (reviewed in [19]).
In the yeast Saccharomyces cerevisiae, the transcription termination machinery can be divided into three different sub-complexes: cleavage factor IA (CFIA), cleavage factor IB (CFIB), and cleavage and polyadenylation factor (CPF). CFIA is comprised of the Rna14, Rna15, Pcf11, and Clp1 proteins. CFIB consists of the RNA-binding protein Hrp1, which is tightly associated with CFIA. The CPF complex is a large complex that can be further classified into the cleavage factor II (CFII) made out of the Cft1, Yhh1, Pta1, Brr5, Ysh1, Cft2, and Ydh1 proteins; the polyadenylation factor I made of Fip1, Yth1, and Psf1; and other proteins including the Pap1 polymerase. In vitro reconstitution of the cleavage reaction demonstrated that it requires the joint action of CFIA, CFIB, and CFII [20], [21], while additional proteins such as the 5′-3′-exoribonuclease Rat1 are required for termination downstream of poly(A) sites in vivo and dismantling of RNAPII complexes in vitro [22]–[24]. In addition to their role in cleavage, many of the components of the cleavage machinery are required for transcription termination downstream of the poly(A) site and polyadenylation of the transcript (reviewed in [25], [26]). Notably, the CFIA rna14-1 and rna15-1 mutants suffer from transcription elongation defects and increase in transcription-dependent hyper-recombination [27], suggesting that the CFIA complex serves important functions in transcription beyond termination and 3′-end processing.
To assess the possible function of RNA 3′-processing and transcription termination on TC-NER, we analysed the impact of a number of mutations on the DDR and the repair of UV-induced lesions. We found that CFI mutants become sensitive to UV in the absence of GG-NER, but surprisingly are proficient for CPD repair. By contrast, DDR is compromised in those cells, as seen by RNAPII degradation and checkpoint activation analyses upon UV irradiation. In addition, we show that rna14-1 cells are impaired in cell cycle progression and rely on a functional G1/S checkpoint to prevent genomic instability and cell death. Our study reveals that CFI functions in DDR and is required for genomic integrity maintenance in yeast.
We first analysed the sensitivity of several transcription termination mutants to DNA damage in the absence of Rad7, a protein required for GG-NER in yeast. Growth of each double mutant was compared to the growth of rad7Δ after irradiation with UV light and in the presence of the UV-mimetic agent 4-nitroquinoline 1-oxide (4-NQO) (Figure 1A). The rna14-1 rad7Δ, rna15-1 rad7Δ, and hrp1-5 rad7Δ double mutants were significantly more affected by UV irradiation or 4-NQO than the respective single mutants, while the remaining assayed alleles (pcf11-2, rat1-1, and yhh1-3) were not. Notably, deletion of the RAD26 gene, which encodes the main TC-NER factor, further increased the sensitivity of rna14-1 rad7Δ and hrp1-5 rad7Δ mutants, indicating that the rna14-1 and hrp1-5 alleles are not epistatic to rad26Δ (Figure 1B).
Because UV sensitivity in the absence of GG-NER is a phenotype mostly associated with TC-NER deficiencies, we tested whether functional CFI was required for proficient TC-NER by monitoring the repair rates of the transcribed (TS) and non-transcribed (NTS) strands of the constitutively expressed RPB2 gene in rna14-1, rna15-1, and hrp1-5 cells (Figure 2, A and B). With the exception of the 60 min. time-point in rna14-1, which is seemingly lower than the wild type on the TS, no significant differences were observed between the repair rates of the mutants and the wild type in either RPB2 strand. Repair experiments were thus performed in rad7Δ and rna14-1 rad7Δ cells. As can be seen in Figure 2 (A and B), both strains show a similar low repair on the NTS and are repair-proficient on the TS. Together, our results indicate that the rna14-1, rna15-1, and hrp1-5 mutants are repair-proficient for CPDs. Because deficiencies in NER may cause an increase in recombinational repair and rna14-1 cells show moderate hyper-recombination [27], we assessed whether recombination increased upon UV irradiation in rna14-1, rad7Δ, and rna14-1 rad7Δ cells. For this, we used a direct-repeat (LY) and an inverted-repeat (SU) plasmid-based system [28]. As expected, rad7Δ cells show an increase in recombination upon UV-damage in both systems (13- and 35-fold, Figure S1). However, recombination frequencies did not increase upon UV irradiation in rna14-1 cells, suggesting that UV damage is efficiently repaired by NER. Notably, the rna14-1 rad7Δ double mutant shows UV-dependent increase in recombination frequency as compared to the rad7Δ mutants in the direct-repeat system -but not in the inverted-repeat system- suggesting that these cells suffer from increased genomic instability that is not linked to increased CPD repair deficiencies.
The cellular response to UV-induced damage involves, in addition to checkpoint activation, proteosomal degradation of RNAPII [3]. To check the functionality of the DDR in rna14-1 cells, we analysed the stability of Rpb1, the largest subunit of RNAPII, and activation of the Rad53 checkpoint protein upon UV irradiation by Western analysis (Figure 2, C and D). Interestingly, UV-induced Rpb1 degradation was less pronounced and severely delayed in rna14-1 cells as compared to the wild type. Activation of the DNA-damage checkpoint, monitored by the appearance of hyper-phosphorylated Rad53 upon UV irradiation was delayed in rna14-1 cells as compared to the wild type, in which Rad53 phosphorylation occurs immediately upon UV irradiation. In addition, the rna14-1 mutation did not increase the sensitivity to UV or 4-NQO of cells lacking either one of the DNA-damage checkpoint proteins Rad9 and Mec1 (Figure S2), suggesting that CFI might act within the canonical checkpoint pathways. To gain more insights into the function of CFI in the cellular response to UV-induced damage, Rpb1 stability and Rad53 phosphorylation were also analysed in cells bearing the rna15-1, hrp1-5 and pcf11-2 mutations (Figure S3). Both rna15-1 and pcf11-2 cells were partially impaired in UV-induced Rpb1 degradation while hrp1-5 cells behaved similarly to the wild type. However, Rad53 phosphorylation was delayed in the rna15-1 and hrp1-5 mutants but not in pcf11-2 cells. These interesting results suggest that UV-induced Rpb1 degradation might not depend on Rad53 activation. Previously, deletion of the DEF1 gene was shown to increase the sensitivity to UV in the absence of GG-NER without affecting DNA repair at the molecular level and to impair UV-dependent Rpb1 degradation [29]. Thus, we assayed viability and sensitivity of rna14-1 def1Δ, rna15-1 def1Δ, hrp1-5 def1Δ, and rat1-1 def1Δ double mutants to assess possible genetic interactions and observed strong synthetic sickness even in the absence of exogenous damage in all strains except hrp1-5 def1Δ (Figures 2E and S4). These interesting genetic interactions suggest that Def1 and CFI might have complementary functions for cell growth, which eventually rely on alternative ways to regulate RNAPII turnover. Although the penetrance of the different alleles differs depending on the analysed phenotype, our data indicate that CFI is required for the cellular response to UV-induced damage.
Sensitivity analysis of different termination mutants to distinct DNA damaging agents revealed that the rna14-1, rna15-1, and hrp1-5 mutants were sensitive to Phleomycin and to methyl methansulfonate (MMS) in contrast to the pcf11-2, rat1-1, and yhh1-3 cells, which were either slightly or not sensitive to those genotoxic agents (Figure 3A). Interestingly, the three alleles conferring significant sensitivity were those that increase the UV-sensitivity of rad7Δ mutants. To assess whether this phenotype was general rather than specific to GG-NER, we generated double mutants of rna14-1 with mutations in representative genes with known functions in the different DNA repair pathways, including homologous recombination (HR), non-homologous end joining (NHEJ), post-replicative repair (PRR), mismatch repair (MMR), base excision repair (BER) and NER (Figure 3B). Interestingly, the rna14-1 mutant showed synthetic growth defects even in the absence of exogenous damage with several repair mutants, including rad52Δ, ku70Δ, lig4Δ, and rad1Δ. These growth defects are further sustained by DNA content profiling FACS analysis (Figure S5). In addition, synthetic UV/4-NQO sensitivity was observed in all double mutants but rna14-1 ogg1Δ ntg1Δ ntg2Δ. Thus, our results indicate that Rna14 dysfunction makes cells unable to cope with high levels of DNA damage and rely on functional repair pathways even in the absence of exogenous damage.
To check whether these genetic interactions might arise from expression defects of DNA repair genes, mRNA expression was analysed by microarrays in rna14-1 and rna15-1 cells (Table S1). The results obtained with the two mutants were highly similar (Figure S6). Analysis of gene ontology terms of genes with higher (> 2-fold) and lower (< 2-fold) expression as compared to wild-type levels revealed that many genes involved in the DNA damage and/or stress response are induced in these mutants (Table S2), including genes such as OGG2, PRX1, DNL4, LIF1, RAD2 or MAG1. In addition, we found out that in rna14-1 or rna15-1 cells, the down-regulated genes were on the average longer than those of the entire genome, while the up-regulated genes were shorter (Figure S6), but DNA repair genes were not specifically down regulated. Thus the results rule out that the reduced capability of CFI mutants to withstand DNA damage is due to reduced transcription of repair protein encoding genes. On the contrary, the elevated expression of DNA damage and/or stress response transcripts suggests that CFI mutants may accumulate DNA damage or structures that impose a steric hindrance to DNA metabolic processes.
Transcription and replication need to occur in a coordinated manner in order to avoid conflicts that can result in genetic instability (reviewed in [30], [31]). To assess whether the CFI dysfunction affects replication, we first analysed sensitivity of several mutants to hydroxyurea (HU), a drug that slows replication down by reducing the pool of available deoxyribonucleotides (Figure 4A). Notably, the alleles that conferred sensitivity to HU were rna14-1, rna15-1, and hrp1-5, while the others did not at concentrations assayed. Since the expression of genes encoding ribonucleotide reductase components were not affected in rna14-1 and rna15-1 (Table S1), the observed HU sensitivity might reflect DNA replication impairment. Next we analysed plasmid loss in rna14-1 cells as a way to measure replication efficiency genetically (Figure 4B). Our results show that less than 5% rna14-1 cells maintained the pRS315 centromeric plasmid after about 10 divisions in non-selective medium as compared to the 50% value of wild-type cells. FACS analysis of cell cycle progression upon release from α-factor-mediated G1-arrest revealed that rna14-1 mutants remain trapped in G1 and suffer from a delay in S-phase entry as compared to the wild type (Figure 4C). For a specific analysis of initiation and progression of replication, we monitored BrdU incorporation upon release from α-factor-mediated G1-arrest at three different early origins (Figure 4D). DNA was immunoprecipitated with anti-BrdU antibody and BrdU enrichment at each locus was analysed by real-time qPCR with specific primers. Importantly, strong defects in replication were observed in rna14-1 mutants, as ARS activation was significantly reduced and occurred at later time points than in wild-type cells. Thus, cell-cycle progression is severely compromised in rna14-1 cells.
Because G1 to S-phase progression was markedly delayed in rna14-1 cells, we asked whether persistent G1/S checkpoint activation might be responsible for the observed cell-cycle delay. Deprivation of Sic1, a protein that is required for the G1/S checkpoint, suppressed the S-phase entry defects in the rna14-1 mutant upon release from α-factor-mediated G1-arrest as seen by FACS analysis (Figure S7). To evaluate the consequences of forcing S-phase entry in rna14-1 mutants by SIC1 deletion, we analysed phosphorylated H2A (H2A-P) levels by Western analysis (Figure 5A). Our results indicate that the rna14-1 sic1Δ mutant accumulates DNA damage, as seen by the large amount of H2A-P. We then analysed recombination and Rad52-foci accumulation to gain insight into the impact of G1/S-checkpoint bypass in rna14-1 cells. As rna14-1 sic1Δ shows severe growth defects at 30°C (Figure S8), recombination was scored at 26°C, a semi-permissive temperature for the rna14-1 mutant, in a direct-repeat (LYΔNS) as well as an inverted-repeat (TINV) plasmid-based system [28] (Figure 5B). A significant increase in recombination frequency was observed in the double rna14-1 sic1Δ mutants with respect to the frequencies of either single mutant in both systems. Rad52-foci accumulation was monitored in cells transformed with plasmid pWJ1344 expressing a Rad52-YFP fusion protein using fluorescence microscopy. As can be seen in Figure 5C, the percentage of S/G2 cells with Rad52-foci was significantly higher in the rna14-1 sic1Δ double mutant (≈35%) than in the single mutants (<20%). Altogether, these results indicate that a functional G1/S checkpoint is essential to avoid genomic instability and/or cell death in rna14-1 cells.
In this study, we asked whether transcription termination might contribute to DNA repair by TC-NER in S. cerevisiae. We found that the rna14-1, rna15-1, and hrp1-5 alleles of CFI confer increased UV and 4-NQO sensitivities in the absence of GG-NER, but surprisingly do not affect CPD repair in a transcribed gene. Importantly, we show that both checkpoint activation and RNAPII degradation are delayed in UV-irradiated CFI-deficient cells and that the rna14-1 mutation interacts genetically with mutations affecting several DNA repair pathway, including HR, NHEJ, MMR, PPR, and NER, in some cases even in the absence of exogenous DNA damage. Our data indicate that CFI participates in DDR in yeast and that this function is needed to cope with high amount of DNA damage. Additionally, we demonstrate that the rna14-1 mutation leads to severe cell cycle progression hindrance and that a functional G1/S checkpoint becomes essential in restraining genomic instability when CFI function is impaired.
Although the precise mechanisms underlying termination downstream of poly(A) sites and 3′-end processing of RNAPII-transcribed genes remains unresolved, it certainly requires cooperation among several factors, including CFI, CPF, Pap1, Rat1 and even the RNAPII holoenzyme (reviewed in [32], [33]). CFIA is progressively recruited to RNAPII during elongation and peaks at poly(A) sites [34], [35]. Its role in transcription termination and 3′-end processing is recapitulated by ongoing transcription past poly(A) sites and in vitro cleavage and polyadenylation defects in CFI mutants [36]–[38]. The CFIB factor Hrp1 binds throughout transcribed genes [39] and displays in vitro cleavage and polyadenylation defects when mutated [40], [41]. We found that CFIA rna14-1 and rna15-1 as well as the CFIB hrp1-5 alleles increased the UV and 4-NQO sensitivities of cells deficient in GG-NER and led to Phlemomycin and MMS sensitivities while the CFIA pcf11-2, CPF yhh1-3, and the rat1-1 alleles did not (see Figures 1 and 3A). On the other hand, UV-induced Rpb1 degradation is impaired in rna14-1, rna15-1 and pcf11-2 but not in hrp1-5 while Rad53-phosphorylation upon UV irradiation is delayed in rna14-1, rna15-1 and hrp1-5 but not in pcf11-2 cells (Figures 2C, 2D and S3). Thus it appears that the penetrance of each particular mutation depends on the assayed phenotype. Indeed, different pcf11 alleles differ in phenotype strength as seen by RNAPII chromatin immunoprecipitation (ChIP) on the ADH1 and PMA1 genes [42]. However, transcriptional read-through or 3′-end processing defects alone might not be sufficient to impair the DDR as ongoing transcription past poly(A) sites are also observed in yhh1-3 and rat1-1 mutants, and yhh1-3 is deficient in 3′-end cleavage and polyadenylation as well [22], [36], [43]. One possibility could be that the requirement of CFI function for the DDR could rely on intrinsic sensing activity or specific interaction with DDR factors, thus enabling CFI to coordinate transcription termination and DDR.
UV irradiation was shown to lead to 3′-end processing inhibition along with targeted RNAPII degradation in human cells, these responses seemingly being mediated by direct interaction between CstF, the functional homologue of yeast CFI, and BRCA1/BARD1 [44], [45]. The link between DDR and 3′-end processing is further supported by the observations that partial depletion of the CstF-50 subunit leads to increased UV sensitivity, reduced ability to ubiquitinate RNAPII in response to UV and defects in CPD repair in human cells [46]. Our results show a notable divergence with respect to the human system though, as no CPD repair defects were observed in yeast CFI mutants (see Figure 2A and 2B). Another difference between yeast and human is the observation that poly-adenylated mRNAs get stabilized upon UV irradiation in yeast [47], while transcript deadenylation takes place under damaging conditions in humans, mediated by DNA damage-dependent physical interaction between CstF and the PARN deadenylase [48]. In addition, it has recently been shown that targeted variation of poly(A) site usage occurs in response to 4-NQO treatment in yeast, possibly as a consequence of transient depletion of CPF subunits [49]. Altogether, these findings suggest that transcription termination factors participate in DDR, a multiple-sided system fundamental for cell survival under genotoxic stress conditions.
The cellular response to UV damage involves global down-regulation of transcriptional activity concomitantly with high expression of a subset of stress-induced genes and proteosomal-mediated degradation of RNAPII major subunit Rpb1. Notably, UV-induced Rpb1 degradation is delayed in CFI-deficient cells (see Figures 2C, 2D and S3), RNAPII turnover being thus impaired. Interestingly, transcription termination factors - including CFI - interact with the transcription initiation factor TFIIB and this interaction is required for the formation of gene loops both in yeast and humans [50]–[53]. Gene looping has been proposed to enable the efficient recycling of RNAPII and to contribute to transcription regulation by acting on promoter directionality and transcriptional memory (reviewed in [54], [55]). It is thus conceivable that gene looping may also function to control transcription and RNAPII turnover under DNA damaging conditions. This idea is supported by recent work showing that TFIIB may function as a general transcriptional switch in humans, as it is dephosphorylated during genotoxic stress thus losing its interaction with CstF, while direct interaction between CstF and the p53 tumor suppressor ensures the recruitment of termination factors to the promoter of stress-induced genes [56].
The persistence of stalled RNAPII on transcribed genes is known to impede the progression of the replication machinery and to be one of the causes underlying transcription-associated recombination (TAR) (reviewed in [30], [31]). Recently, inhibition of Rho-dependent transcription termination in bacteria has been shown to induce double-strand breaks depending on replication, suggesting that Rho might function in the release of obstructing RNAP during replication [57]. It is possible that CFI might act on paused RNAP, whether or not stalled at a DNA damage, contributing to its displacement and thus allowing progression of an oncoming replication fork. Over the last few years, growing evidence supports a role for co-transcriptionally formed RNA-DNA hybrids (R-loops) as a source of TAR (reviewed in [58]). Noteworthy, several transcription termination and 3′-end processing mutants have been shown to accumulate R-loops in yeast (including pcf11-2 and rna15-58) [59]. It is thus possible that stalled RNAPIIs accumulate at DNA damages or other structures such as R-loop in CFI mutants, leading to steric hindrances to the replication machinery that would account for the observed cell cycle progression defects (see Figure 4). The mechanisms by which stalled RNAPIIs or structures presenting steric hindrance to replication are sensed to activate the G1/S cell cycle checkpoint, which is required to restrain genetic instability in rna14-1 cells (see Figure 5), are currently unknown. Interestingly, the Sen1/SETX helicase - a component of the NRD transcription termination complex - prevents R-loop accumulation at transcription termination sites both in yeast and humans [60], [61]. In addition to its association with transcribed units, yeast Sen1 is also found at replication forks, contributing to prevent deleterious outcomes of the putative collisions between the transcription and replication machineries [62]. Noteworthy, Sen1 interacts physically with the NER repair protein Rad2 and the sen1-1 mutation increases the UV sensitivity of cells lacking RAD2 [63], suggesting further connections between transcription termination, replication, and DNA repair.
Altogether, our results support a model in which CFI dysfunction impairs DDR, probably leading to the accumulation of endogenous DNA lesions, and hinders DNA replication possibly due to the accumulation of RNAPs, whether or not stalled at DNA damages, thus rendering the G1/S checkpoint mandatory to prevent genomic instability (see Figure 6). Our findings emphasize the importance of coordinating transcription termination, DDR and replication in the maintenance of genomic stability and suggest that CFI plays a fundamental function in the coupling of these processes.
All strains used were isogenic to W303, and are listed in Table S3. Newly generated strains were obtained either by direct transformation or by genetic crosses. Plasmids used for recombination tests were pRS314-LYΔNS, pRS316-TINV, pRS314-LY and pRS314-SU [28].
For cell survival, yeast cells were grown in YEPD rich medium to an OD600 of 0.6. 10-fold serial dilutions were dropped on YEPD plates, irradiated with the indicated dose of UV-C light, and incubated in the dark at 30°C for 3 days. For the 4-NQO, Phleomycin, MMS, CPT and HU sensitivity assays, the serial dilutions were dropped on YEPD plates containing the indicated amounts of genotoxic agents and incubated in the dark at 30°C for 3 days. UV survival curves were performed as described [9]. UV-C irradiation was performed using a BS03 UV irradiation chamber and UV-Mat dosimeter (Dr. Gröbel UV-Elektronik GmbH).
CPD repair at the RPB2 gene was analysed as described [64]. Briefly, cells were grown at 30°C in YEPD rich medium, irradiated in SD medium w/o amino acids with 200 J/m2 UV-C light (BS03 UV irradiation chamber), the medium supplemented to YEPD rich and the cells incubated at 30°C in the dark for recovery. DNA from the different time-points was extracted, cut with NsiI and PvuII restriction enzymes (Roche) and aliquots were either treated with T4-endonuclease V (Epicentre) or left untreated. DNA was electrophoresed in 1.3% alkaline agarose gels, blotted to Nylon membranes and hybridized with radioactively labelled strand-specific DNA probes, which were obtained by primer extension. Sequences of the primers are listed in Table S4. Membranes were analysed and quantified with a PhosphorImager (Fujifilm FLA5100). The average of the initial damage generated was 0.025 CPD/kb. To allow direct comparison between different strains, repair curves were calculated as the fraction of CPDs removed versus time. The initial damage was set to 0% repair.
Cells were grown at 30°C in YEPD medium to an OD660 of 0.6. Total RNAs were purified (RNeasy Midi kit, Qiagen) and expression profiling performed using the Affymetrix platform (see Table S1). The relative RNA levels for all yeast genes were determined using an Affymetrix microarray scanner and processed using the robust multiarray average method. Statistical data analyses were performed using the limma package (affylmGUI interface) of the R Bioconductor project (http://www.bioconductor.org). For each strain, microarray analysis was conducted in triplicate. All values presented represent the average of these three determinations. Genes were considered significantly up- or down-regulated when their expression values were > or < 2-fold, respectively (parameters: false discovery rate-adjusted p-value<0.01, B-statistic value>2, and average log2intensity A>7). The expression data for each mutant has been deposited in NCBI's Gene Expression Omnibus (accession number GSE50947).
Plasmid loss was monitored as the percentage of cells that lost centromeric plasmid pRS315 upon growth in non-selective media. Individual transformants were inoculated in 5 ml YEPD and grown at 30°C to OD660 0.6. Cells were plated on YEPD or SC-leu to determine the percentage of plasmid loss. Six individual transformants were analysed for each genotype.
Recombination frequencies were determined as the average value of the median frequencies obtained from at least three independent fluctuation tests performed at 26°C each from six independent colonies according to standard procedures [28].
Isogenic wild-type and rna14-1 strains deleted for the BAR1 gene and carrying several copies of the Herpes simplex thymidine kinase (TK) under the control of the strong constitutive GPD promoter were obtained by genetic crosses with strain SY2201 (E. Schwob). Cells were grown in YEPD, incubated for 2.5 h with 0.125 µg/ml α-factor, washed twice with pre-warmed YEPD and released into S phase by addition of 1 µg/ml pronase. BrdU (200 µg/ml) was added to the cultures prior to release. Cell cycle progression was monitored by flow cytometry on a FACSCalibur (BD Bioscience) using CellQuest software. Chromatin immunoprecipitation was carried out as described [65] with minor modifications. Briefly, Sodium Azide (0.1%) was added to each sample and cells were broken in a multi-beads Shocker (MB400U, Yasui Kikai, Japan) at 4° in lysis buffer (50 mM HEPES-KOH pH 7.5, 140 mM NaCl, 1 mM EDTA, 1% triton X-100, 0.1% sodium deoxicholate) and sonicated. Immunoprecipitation was performed using anti-BrdU antibody (MBL) attached to magnetic beads coated with Protein A (Invitrogen). Input and precipitated DNA were analysed by RT qPCR (7500FAST Applied Biosystems). Relative BrdU incorporation at a given region was calculated relative to the signal at a late replicating region (Chr. V, position 242210–242280, [66]) in the same sample. Primer sequences are listed in Table S4.
Rad52-YFP foci from log-phase cells transformed with plasmid pWJ1344 were visualized with a DM600B microscope (Leica) as previously described [67] with minor modifications. Individual transformants were grown to early-log-phase at 26°C, incubated at 30°C for 4 hours, fixed for 10 minutes in 0.1 M KiPO4 pH 6.4 containing 2.5% formaldehyde, washed twice in 0.1 M KiPO4 pH 6.6, and resuspended in 0.1 M KiPO4 pH 7.4. A total of 617 wild type, 947 rna14-1, 733 sic1Δ, and 820 rna14-1 sic1Δ cells derived from at least three different transformants were analysed.
Detection of Rpb1, Rad53, H2A-P, and β-actin was accomplished by Western analysis of TCA-precipitated proteins separated in 4–20% Cristerion TGX gradient PAGE (Biorad). Antibodies 8WG16 (Rpb1, Covance), sc-20169 (Rad53, Santa Cruz Biotechnology), ab15083 (H2A-P, Abcam) and ab8224 (β-actin, Abcam) were used. For quantification, secondary antibodies conjugated to IRDye 680CW or 800CW (LI-COR) were used, the blot scanned in an Odyssey IR scanner and analysed with Image Studio 2.0 software (LI-COR). For Western analysis after UV irradiation, cells were grown in YEPD rich medium to mid-log-phase, resuspended in SD media lacking amino acids to an OD660 of 0.6 and irradiated with UV-C light in a BS03 UV irradiation chamber (Dr. Gröbel UV-Elektronik GmbH) at 100 J/m2. Medium was supplemented to YEPD rich and cells incubated in the dark at 30°C for recovery.
|
10.1371/journal.pcbi.1000594 | Challenges Predicting Ligand-Receptor Interactions of Promiscuous Proteins: The Nuclear Receptor PXR | Transcriptional regulation of some genes involved in xenobiotic detoxification and apoptosis is performed via the human pregnane X receptor (PXR) which in turn is activated by structurally diverse agonists including steroid hormones. Activation of PXR has the potential to initiate adverse effects, altering drug pharmacokinetics or perturbing physiological processes. Reliable computational prediction of PXR agonists would be valuable for pharmaceutical and toxicological research. There has been limited success with structure-based modeling approaches to predict human PXR activators. Slightly better success has been achieved with ligand-based modeling methods including quantitative structure-activity relationship (QSAR) analysis, pharmacophore modeling and machine learning. In this study, we present a comprehensive analysis focused on prediction of 115 steroids for ligand binding activity towards human PXR. Six crystal structures were used as templates for docking and ligand-based modeling approaches (two-, three-, four- and five-dimensional analyses). The best success at external prediction was achieved with 5D-QSAR. Bayesian models with FCFP_6 descriptors were validated after leaving a large percentage of the dataset out and using an external test set. Docking of ligands to the PXR structure co-crystallized with hyperforin had the best statistics for this method. Sulfated steroids (which are activators) were consistently predicted as non-activators while, poorly predicted steroids were docked in a reverse mode compared to 5α-androstan-3β-ol. Modeling of human PXR represents a complex challenge by virtue of the large, flexible ligand-binding cavity. This study emphasizes this aspect, illustrating modest success using the largest quantitative data set to date and multiple modeling approaches.
| Promiscuous proteins generally bind a large array of diverse ligand structures. This may be facilitated by a very large binding site, multiple binding sites, or a flexible binding site that can adjust to the size of the ligand. These aspects also increase the complexity of predicting whether a molecule will bind or not to such proteins which frequently function as exogenous compound sensors to respond to toxic stress. For example, transporters may prevent absorption of some molecules, and enzymes may convert them to more readily excretable compounds (or alternatively activate them prior to further clearance by other detoxification enzymes). Nuclear hormone receptors may respond to ligands and then affect downstream gene expression to upregulate both enzymes and transporters to increase the clearance for the same or different molecules. We have assessed the ability of many different ligand-based and structure-based computational approaches to model and predict the activation of human PXR by steroidal compounds. We find the most effective computational approach to identify potential steroidal PXR agonists which are clinically relevant due to their widespread use in clinical medicine and the presence of mimics in the environment.
| Promiscuous proteins generally bind a large array of diverse ligand structures. These proteins include enzymes like cytochrome P450s (e.g. CYP3A4, EC 14.13.97), transporters such as P-glycoprotein (ABCB1), the human ether-a-go-go related gene (hERG, Kv11.1) potassium channel and nuclear hormone receptors (NHRs) such as the pregnane X receptor (PXR; NR1I2; also known as SXR or PAR) [1]. This promiscuous binding may be facilitated by a very large binding site, multiple (overlapping) binding sites, or a flexible binding site that can adjust to the size of the ligand. Intrinsic disorder in the protein may also have a role [2],[3]. These proteins described above are also particularly important as xenobiotic sensors and represent key mechanisms to respond to toxic stress.
The human PXR [4]–[6] transcriptionally regulates genes involved in xenobiotic metabolism and excretion, as well as other cellular processes such as apoptosis [7]–[11]. Human PXR has a very broad specificity for ligands as exemplified by the structurally diverse array of activators including endogenous (bile acids, steroid hormones, fat-soluble vitamins) and exogenous (prescription and herbal drugs, and environmental chemicals) compounds. Activation of human PXR can cause drug-drug interactions [4],[5] or result in physiological effects ranging from ameliorating cholestatic injury to the liver, altering bone homeostasis, and causing cell proliferation [12]. As PXR represents a potential target for pharmacologic modulation in disease, it is therefore becoming even more important to develop methods that can identify whether a molecule is likely to be a PXR agonist [13]. Currently there are five high-resolution crystal structures of human PXR [14]–[18] available in the Protein Data Bank (PDB) (and another structure to be deposited [19]). The structures have provided atomic level details that have led to a greater understanding of the ligand binding domain (LBD) and the structural features involved in ligand-receptor interactions [9], [10], [12]–[15]. The co-crystallized ligands include the natural products hyperforin (active component of the herbal anti-depressant St. John's wort) and colupulone (from hops), the steroid 17β-estradiol, the synthetic compounds SR12813, T1317 and the antibiotic rifampicin. These ligands span a range of molecular sizes (M.Wt range 272.38 – 713.81Da, mean 487.58±147.25Da, ) and are predicted as generally hydrophobic (calculated ALogP [20] 3.54–10.11, mean 5.54±2.41). The cavernous ligand binding pocket (LBP) with a volume >1350 Å3 accepts molecules of these widely varying dimensions and chemical properties, and is likely capable of binding small molecules in multiple orientations [21]. This complicates overall prediction of whether a small molecule is likely to be classified as a PXR agonist using traditional structure-based virtual screening methods like docking [13],[22]. With regard to this, we have previously shown that the widely used structure-based docking methods FlexX and GOLD performed relatively poorly in predicting human PXR agonists [7],[16] and this is perhaps not surprising based on the observations described above.
An alternative method, which has been found to be valuable elsewhere in drug discovery, particularly when there may not be an available crystal structure of the target protein, uses a ligand-based approach. In this case a series of small molecule structures with PXR agonist activity data can be used to facilitate a structure activity relationship (SAR). When the biological activity data is continuous this will enable a quantitative structure activity relationship (QSAR) [23]–[25]. One widely used computational technology produces pharmacophores [20]–[23], which represent models that encode the key chemical features important for biological activity. Human PXR agonist pharmacophore models have been shown to possess hydrophobic, hydrogen bond acceptor and hydrogen bond donor features, consistent with the crystallographic structures of human PXR ligand-receptor complexes [26]–[29]. These pharmacophore models have predominantly used structurally diverse ligands in the training set and have the limitation in most cases of compiling data from multiple laboratories using different experimental protocols, ultimately forcing binary classifications of ligands for the training sets (i.e., activating versus non-activating). Most of the models so far use EC50 data, a measure of receptor transactivation. Although binding assays have been done with human PXR, they are problematic given the low affinity of most PXR activators. As a result, there is little radioligand binding data in the literature other than competition experiments with radiolabeled SR12813.
To date there have been few attempts to build ligand-based models around a large structurally narrow set of PXR activators. The absence of large sets of quantitative data for PXR agonists has restricted QSAR models to a relatively small universe of molecules compared to the known drugs, drug-like molecules, endobiotics and xenobiotics in general [30]. The PXR data limitation has resulted in the use of various machine learning methods (e.g support vector machine, recursive partitioning etc.) when the biological data is binary in nature (e.g. activating or binding versus non-activating / non-binding) [13],[22],[25],[30].
As part of an ongoing analysis of NHRs [25]–[28], we have generated a large cadre of experimental data for classes of steroidal compounds, namely androstanes, estratrienes, pregnanes and bile acids/salts [31]. The advantages of using steroidal compounds for QSAR are that they are amenable to common alignments based on the steroidal backbone. For, example steroids represented the first datasets used for comparative molecular fields analysis (CoMFA) [32] and have been widely used as a benchmark for other methods such as comparative molecular similarity analysis (CoMSIA)[33]. Pharmacophore methods, in contrast, generally do not require the rigid alignment methods and have found use with more diverse structure sets [28],[31]. Using this large quantitative data set of PXR activators, we applied various ligand-based computational methods including Bayesian modeling with 2D fingerprints. We also compared the results from QSAR approaches to molecular docking into the six available human PXR crystal structures.
Modeling of a broad specificity receptor such as PXR represents a challenge for in silico modeling and it is invaluable to know what approaches prove successful, if any. Ideally, these methods will also translate to modeling approaches for other broad specificity enzymes, transporters and ion channels [1], or other promiscuous proteins [34]. We are not aware of any similar studies using a comparative approach to predicting ligand-protein interactions for promiscuous proteins. This study also provides further insights into PXR-steroid interactions which have not been well studied [19] and is clinically relevant due to the widespread use of steroidal compounds and steroid mimics (e.g. oral contraceptives [35], for inflammation and as cancer treatments etc.) in clinical medicine [36], as well as the increasing problem of environmental contamination by endocrine disruptors [24].
All compounds shown in Table S2 were docked to the six human PXR crystal structures using GOLD which we have used previously for docking diverse compounds into the human PXR structure [22]. All six crystal structures superimposed with a backbone root mean squared deviation of 0.5 Å suggesting that they had very similar structures and their co-crystallized ligands bound to the same binding pocket (Figure S1). The docking scores for all the compounds (Table S2) were in the range of 36 to 77 for all the crystal structures and their corresponding Tanimoto similarity scores to 5α-androstan-3β-ol and the crystal ligand 17β-estradiol using MDL public keys were between 0.4 and 1.
To evaluate docking results, we compared docking scores for classifying compounds as activators or non-activators of PXR. Using an EC50 value of 10 µM as a cutoff the compounds listed in Table S2 were classified as activators (30 compounds) and non-activators (89 compounds). These results were compared to the classification obtained from the docking studies. The overall accuracy (Q values) were in the range of 35 to 55 % for models that used 5α-androstan-3β-ol based similarity scores as weights to the goldscore, while the Q values were in the range of 47 to 58% for models that were generated with goldscores weighted with 17β-estradiol based similarity scores (Table 1). The Matthews coefficient C showed a modest prediction rate with the best score for docking of compounds to PXR crystal structure 1M13. Further changing the cutoff values to either 100 µM or 40 µM did not improve the prediction rates. The Q value for a model computed by averaging all the models with 5α-androstan-3β-ol weighted goldscore was 46% and for the average model with 17β-estradiol the weighted goldscore was 51%.
Although the overall performance of docking produced rather modest results for classification the results for individual classes of compounds was better than average. In the best classification model (compounds docked to crystal structure 1M13 and weighted with 17β-estradiol based similarity scores), 20 out of 30 PXR activators and 49 out of 89 non-activators were predicted correctly. Among the androstanes, 6 out of 11 compounds were predicted correctly as activators and 9 out of 14 compounds were classified as non-activators (Table S2). Among the bile salts, all 4 activators and 22 out of 46 non-activators were predicted correctly. Among the estratrienes, 5 out of 7 activators were predicted correctly, while the 4 non-activators were predicted as activators (Table S2). The reason for this mis-classification was due to the high similarity scores of the estrogens with 17β-estradiol. In the pregnane class, 4 out of 7 activators and 16 out of 20 non-activators were correctly classified (Table S2). Some examples of molecules in their binding modes with PXR structure 1M13 are shown in Figure 1.
All 115 compounds shown in Table S2 were used to generate a Bayesian classification model [37], using a definition of active as a compound having an EC50 for PXR activation of less than 10 µM. Using molecular function class fingerprints of maximum diameter 6 (FCFP_6) and 8 interpretable descriptors (AlogP, molecular weight, rotatable bonds, number of rings, number of aromatic rings, hydrogen bond acceptor, hydrogen bond donor and polar surface area) a model was developed with a receiver operator curve (ROC) statistic for leave one out cross validation of 0.84. In addition to the leave one out cross validation, further validation methods were undertaken. After leaving 20% of the compounds out 100 times the ROC is 0.84±0.08, concordance 73.2 %±8.94, specificity 69.14%±12.12, and sensitivity 84.11%±18.04. The Bayesian method appears to have good model statistics for internal cross validation of steroids. These statistics suggest the model is stable and not over-trained as the ROC values are essentially identical to that obtained with leave one out cross validation.
We have additionally used this model to classify a previously used diverse molecule test set [13],[22]. After removing the steroids from the test set, the Bayesian PXR model was used to rank 123 molecules (65 activators and 58 non activators). Out of the top 30 molecules scored and ranked with this model 20 (75%) were classified as activators (EC50 <100 µM) (Table S3). Even though the cutoff for activity for the Bayesian model is more stringent it still appears to be able to predominantly pick out the key molecular features that contribute to activity in non-steroidal compounds.
The Bayesian model with FCFP_6 descriptors also enabled the visualization of substructure fingerprints (Figure 2) that either contributed positively or negatively to the activity classification. It appears that all positive contributing substructures are essentially hydrophobic, while negatively contributing features possess hydroxyl or other substitutions which are likely not optimally placed to facilitate interactions with hydrogen bonding features in PXR. Therefore possession of these hydrogen bond acceptor and donor features indicated in the steroidal substructures appears to be related to loss of PXR activation. The method does not readily identify where these groups should be added in contrast to methods like docking [13],[38].
A major challenge in CoMFA and CoMSIA modeling is alignment of molecules, which must be defined by the user. As described in Text S1, multiple alignment approaches were attempted. Despite the use of multiple alignments, the best CoMFA and CoMSIA models consistently showed a large difference between the correlation R2 and cross-validated (XV-R2), whether modeling the entire set of steroidal compounds or the various subsets (androstanes, bile salts, pregnanes). This suggests that the CoMFA and CoMSIA models do not generalize beyond the molecules in the training set, even for a subset of steroidal compounds (Text S1, Tables S4, S5, S6, S7 and Figures S2, S3, S4, S5, S6, S7).
Using the pharmacophore approach for the individual steroids, the training set r values were quite low but increased upon inclusion of excluded volumes with variable weight and tolerances (0.81–0.93) (Table S8). All PXR pharmacophores (Figure S8) had at least 2 hydrophobes and a hydrogen bond acceptor in common (Text S1). Using the pharmacophores derived from training sets based on subsets of steroidal compounds (e.g., androstanes only) to predict the other respective subsets did not result in reliable correlations (data not shown), suggesting that highly specific pharmacophores were generated or this may be due to the addition of the excluded volumes which limits the chemical space of molecules mapping to the features. These class-specific pharmacophores may therefore only be useful for making predictions of very closely related molecules and even crossing steroidal classes may be extrapolating too far beyond the training sets.
4D-QSAR performed somewhat better than CoMSIA and CoMFA in modeling the compounds in the training sets using three atom alignments (Table S9). One potential advantage of 4D-QSAR relative to standard 3D-QSAR methods is the ability to consider an ensemble of different ligand conformations, theoretically increasing the chances of defining the active conformation. The best 4D-QSAR models are found in Table S10 and Figure S9, and generally predict steric/non-polar interactions between ligand and receptor. Although the XV-R2 for the best 4D-QSAR models are better than for CoMFA and CoMSIA models of the same training sets, the 4D-QSAR were poorly predictive of the activity of compounds in the test set (Table S10).
4D- and 5D-QSAR have the advantage of being able to select the bioactive conformation from a pool of possible binding modes in parallel to the QSAR modeling stage. We have tested three different alignment protocols in conjunction with the 5D-QSAR technique Raptor.
In the alignment protocols (2) and (3) the protein crystal structure was used as a forbidden (excluded) region. A penalty was added to the similarity score for alignment solutions that overlapped with the protein, thus physically impossible solutions were removed from the alignment. As significant protein flexibility is observed on the side chain level, all crystal structures were aligned using PyMol [39]. Side chains that have different rotamer states for different co-crystallized ligands were removed from the forbidden region definition.
Our multidimensional QSAR study (software Raptor [40]) was based on the same set of 115 molecules as described in the CoMFA and CoMSIA studies. The dataset was split into 95 training set compounds, and 20 test set compounds identical to the separation used in the CoMFA and CoMSIA studies. For 33 compounds only an upper limit for their Ki values has been experimentally determined. These molecules defined the “threshold class” (26 training, 7 test). A threshold value of 100 µM was chosen considering that the lowest affinities were measured for this dataset at approximately this value. To allow for topological and physicochemical variation at the true biological receptor with different ligands bound, the Raptor results were averaged over 10 individual models defining a surrogate conformational family.
For alignment (1) we were not able to derive QSAR models with predictive models for leave-5-groups-out (r2CV-5) or cross-validation values (i.e.>0.3). This is not surprising, as the identification of bioactive binding modes using docking is difficult for this system (see docking results). If we use an alignment with only the top-1 or top-2 solutions, we most probably end up with an alignment containing incorrect binding modes. Using the top-10 or top-20 binding modes generates too large a variety of contacts between ligand and binding site model that the QSAR algorithm is not able to extract the critical interactions throughout the binding site modeling phase.
For alignment (2) a QSAR model with a r2CV-5 value of 0.55 could be generated, but with no observed correlation for the test set. For alignment (3) a QSAR model with an r2CV-5 of 0.56 was derived with a predictive r2 for the test set of 0.45. The superior model based on alignment (3) was due to the focused class-based alignment process (Figure 3). The maximum deviation of predicted from experimentally measured EC50 is 5.6 and 3.0 fold for training and test set, respectively. Significantly higher regression coefficients can hardly be expected for this dataset considering the fact that the threshold compounds have to be removed from the calculation of the regression coefficients yielding a rather small range in EC50 of 2.2 log units (Figure S10, Table S11). This is in contrast to the CoMFA and CoMSIA simulations where the threshold compounds have been assigned an EC50 value of 10,000 µM yielding a range of 4.1 log units. All except one of the 33 threshold compounds have been predicted with an EC50 value lower than the given threshold or maximally a factor of 6.6 fold higher. Only 5α-Androstane was predicted to have a 46 fold higher value than the threshold. Thus, the model was able to predict the affinity of compounds accurately and at the same time was able to classify weak- or non-binding molecules correctly.
It has been suggested that PXR forms a heterotetramer and exhibits a range of motions which are key for its functioning and preparing for coactivator binding at the Activator Function (AF-2) site [41]. The large and promiscuous ligand binding pocket of PXR accepts molecules of widely varying sizes (Table S1), and is likely capable of binding small molecules in multiple orientations. Furthermore, movement of regions of this pocket may be translated elsewhere in the protein to influence protein-protein interactions. Thus, the identification of the bioactive conformation of a ligand binding to PXR (and the effect it might have as an agonist, antagonist or allosteric antagonist [10]) and development of a ligand alignment based on these conformations represents a challenge for any computational technique. A realistic ligand alignment, however, is the basis for a reliable 3D-QSAR model. Computational methods including QSAR (3D, 4D and 5D), pharmacophores and machine learning classification models for PXR can assist in rapid prediction of whether a compound is likely to be an agonist (activator), however each method has its limitations and advantages (Table 2). For example a previous study used human PXR activation data for 30 steroidal compounds (including 9 bile acids) to create a pharmacophore with four hydrophobic features and one hydrogen bond acceptor [27]. This pharmacophore contained 5α-androstan-3β-ol (EC50 0.8 µM) which contains one hydrogen bond acceptor, indicating that in contrast to the crystal structure of 17β-estradiol (published EC50 20 µM) bound to human PXR with two hydrogen bonding interactions [19], hydrophobic interactions may therefore be more important for increased affinity [27]. This and other pharmacophores have been used to predict PXR interactions for antibiotics [35] which were verified in vitro, suggesting one use for computational approaches in combination with experimental methods.
To our knowledge there has been no comparative analysis of the steroidal classes with respect to their use as PXR agonists. The use of the Bayesian classification with 2D fingerprints represents a low computational cost approach [42] which has been used frequently with large molecule datasets [43]–[46]. Using 2D-molecular fingerprint descriptors identified regions in the training set molecules that were predominantly hydrophobic and that were important for PXR activation. Substructures with free hydroxyls as hydrogen bonding features were associated with compounds that were not activators. This is in general agreement with other studies which have used docking to try to help design out PXR activation [38]. This model was able to successfully rank a large test set (Table S3) of non-steroidal molecules, indicative that the molecular descriptors adequately captured the global properties of PXR agonists and suggests some utility.
The current study suggests that while it is generally possible to create 3D-QSAR (CoMFA, CoMSIA, Catalyst) and 4D-QSAR models that can be cross-validated, these models perform poorly when used to predict external molecules. Only the 5D-QSAR model generated displays some success in predicting external test set steroidal compounds. Three main differences between the 5D-QSAR and the 3D-QSAR studies that might contribute to the difference in performance are the less rigid alignment using Symposar [40], the possibility to present a ligand in more than one binding pose and the better treatment of weak or non-binding compounds.
Pharmacophore models for the 4 classes of steroidal compounds possessed some of the features in the published human PXR crystal structures, however the models contained two or three hydrophobic regions (rather than four as shown previously)[27],[28],[31] and one to two hydrogen bond acceptors or a hydrogen bond acceptor and hydrogen bond donor (compared to one hydrogen bond acceptor as shown previously). This might suggest that the steroids evaluated occupy just a part of the ligand binding pocket while larger molecules like rifampicin occupy most of the binding pocket and have subsequently many more interactions with the protein [17]. The addition of the excluded volumes to the pharmacophores was shown to improve the correlation for the training sets and likely acts in a similar manner to using the crystal structures in 5D-QSAR.
Consistent with the QSAR findings were those from docking studies that though modest in success overall, fare much better with individual classes of compounds. The classification was performed using two similarity weighted scoring schemes: one based on a highly potent compound 5α-androstan-3β-ol and the other based on a structurally relevant compound 17β-estradiol. The goal was to test the utility of biasing the scoring scheme with either a structurally relevant compound or a functionally significant compound.
However, in this case 17β-estradiol and 5α-androstan-3β-ol share nearly 75% structural similarity (using MDL Keys and Tanimoto similarity coefficient). The results from the classification studies showed that biasing the scoring scheme with a structurally relevant compound (17β-estradiol) produced classification rates with sensitivity and specificity values averaging at 52% and 50% respectively with slightly better prediction accuracy (Table 1). These results unfortunately cannot be compared with our recent docking study [47] as a different co-crystal ligand was used for the scoring scheme. Although the structure biased scoring scheme performed better among all the compounds, both the scoring schemes performed equally well when individual classes were considered. In the case of androstanes, 6 out of 11 compounds were correctly predicted as activators in docking studies. 5α-Androstan-3β-ol that had the lowest EC50 value (described earlier) was predicted to be an activator in all structures. 5α-Androstan-3β-ol binds with very high docking scores and has a hydrogen bond interaction with His407, a key interaction of PXR (Figure 1A). This interaction was consistent among all the androstane activators. However, epitestosterone sulfate has an EC50 of 3.39 µM and was misclassified in the combined model using predictions from all structures as a non-activator. Docking studies show that epitestosterone sulfate has a consistently reversed docking pose (when compared with 5α-Androstan-3β-ol) in all the models and the sulfate group is predicted to make a hydrogen bond interaction with His407, as opposed to the steroid ester in 1M13 structure (Figure 1B). A few other misclassified activators were docked in reversed poses and often had favorable hydrogen bonding partners such as sulfates that probably influence the binding mode of these steroids. This is a surprising and novel finding of this study and other researchers should be aware of this when docking similar compounds with this functional group.
Among the bile salts, all four activators were correctly predicted and the ligands bind in a conserved mode with the steroid esters participating in favorable interactions with the side chain of His407 and Arg410, and the steroid rings with hydrophobic groups such as Leu411, Leu239 and Phe281 (Figure 1C). The pregnanes had similar activation patterns as the bile salts and docking studies could predict 4 out of the 9 compounds correctly. Among the misclassified compounds, levonorgestrol was predicted to be an activator in three models, and a non-activator in three models and hence could not be classified with high confidence. Levonorgestrol has an EC50 of 4.30 µM and is predicted to have favorable interactions with hPXR as shown in Figure 1D. Despite this, the similarity weighted scoring functions generally performed well in classifying activators as described in the examples above and by the sensitivity values in Table 1. The paucity of available PXR binding data may limit some of the insights from docking experiments performed to date.
It is not surprising that CoMFA and CoMSIA do not perform well as they use rigid alignments of the molecules. This is potentially a seriously limitation given that the binding pocket of PXR may accommodate multiple orientations of the steroids (Figure 1A vs. Figure 1B). Theoretically, 4D- and 5-QSAR should perform better by considering an ensemble of ligand conformations and in fact 4D-QSAR does well within subsets (especially androstanes) but like all methods extrapolates poorly. 5D-QSAR appears to perform the best with the test set. Alignment independent methods like Catalyst which can deal with structurally diverse molecules can generate pharmacophores for the individual classes of compounds but their inter-class predictivity is limited. Another alignment independent method such as using 2D fingerprints and descriptors with the Bayesian classification approach may represent a fast approach to screen for potential PXR agonists, but like all methods their applicability domain [48],[49] is dependent on the training set. In this case the set of steroids would be expected to limit the utility of such models to a relatively narrow class of compounds, although it may be picking up key features in more diverse molecules (Table S3) suggesting overlap in the chemical space.
This study shows the inherent difficulty of producing predictive ligand or structure-based computational models for PXR. Some of the methods used are ligand alignment dependent while others are alignment independent, and each has limitations when used with flexible proteins. These computational models also confirm some of the molecular features (hydrophobicity and hydrogen bond acceptors) identified in previous models and structures, while using a large quantitative dataset to create new QSAR, classification and pharmacophore models to test docking and scoring. The study represents an initial step comparing multiple methods focused on steroidal compounds rather than a more diverse series of drug-like molecules. Using a more diverse series of molecules would have been expected to present even more difficulty for the alignment dependent methods such as CoMFA and CoMSIA. There are also many more commercial computational methods that could be evaluated and compared, although we have used several 3D, 4D, 5D-QSAR methods, machine learning with 2D descriptors, pharmacophore and GOLD docking and scoring methods in this study. The results from these methods could be used in combination as part of a consensus approach or Pareto optimization [50]. The provision of the 115 molecule human PXR dataset is potentially useful as a benchmark PXR set for testing further methods in future. For example, flexible docking methods [51] could be used as well as algorithms that could differentiate multiple binding mechanisms [52].
In conclusion, there are many promiscuous proteins [34] where the modeling of ligand-protein interactions is complicated by a large binding site, multiple binding pockets, protein flexibility or all of the preceding. We have applied several different computational approaches which could also be applied to other proteins like CYPs, transporters and ion channels. This work is therefore more broadly applicable in an attempt to predict whether molecules bind in such flexible proteins, and which methods perform the best. Depending on the desired use of such information, different modeling methods may be appropriate and required. While 2D methods do not encode 3D information like shape [53] they are fast and they can highlight important features likely interacting with the protein. 3D-5D methods provide more shape based information but they are fragile, with a narrow applicability domain and may not be able to differentiate close analogs. Docking is also limited unless key interactions with the protein are already known. Our results suggest that even in the presence of multiple crystal structures, the full range of protein motions may not be captured. As we have previously shown, when docking classification predictions are correct the binding conformation information alone may be instructive [13]. This current analysis indicates that using many different computational approaches (both alignment dependent and alignment independent) may be necessary and expectations should be scaled accordingly if some do not work with such promiscuous proteins. Even with their respective limitations, these methods have provided some useful information of general interest that could be applicable beyond PXR.
Human PXR activation was determined by a luciferase-based reporter assay as has been previously described [21],[33],[34]. The datasets modeled in this study were collected by a consistent protocol and have been previously published [31],[54]. Experimental data for four classes of steroidal compounds, namely androstanes, estratrienes, pregnanes and bile acids/ salts are shown in Table S2.
All molecules described in Table S2 were used for docking experiments. The molecules were docked into these six crystallized structures of human PXR (PDB IDs 1M13, 1NRL, 1SKX, 2O9I, 2QNV and one structure co-crystallized with 17β-estradiol that is not in the PDB identified here as EST). In all cases, the crystal structure ligand was removed, and hydrogen atoms were added to the amino acids. All amino acids within 6 Å of the co-crystallized ligand were identified as the binding site. The docking program GOLD (ver 4 [55]) was used for docking all compounds to the binding sites of each PXR crystal structure. GOLD uses genetic algorithm to explore the various conformations of ligands and flexible receptor side chains in the binding pocket. Further, 20 independent docking runs were performed for each ligand. The docked complexes were scored with goldscore [55] and then rescored using similarity weighted scoring scheme (SWscore). For each ligand, the best ranking conformation's goldscore denoted by Si was used to derive the SWscore shown in equation 1. The similarity scores Wi were computed based on 2D similarity encoded in MDL fingerprint keys calculated using Discovery Studio 2.1 (Accelrys, San Diego, CA, USA). The Tanimoto coefficient was used as the metric to compare the molecular fingerprints. The coefficients varied between 0 and 1, where 0 meant maximally dissimilar and 1 coded for maximally similar. The Tanimoto coefficient between fingerprints X and Y has been defined to be: [number of features in intersect (A, B)]/[number of features in union (A,B)], where A and B are two compounds.
So the SWscore is given by, SWscore = Wi*Si, where Wi was the similarity score of compound i against 5α-Androstan-3β-ol which had the best EC50 value of 0.8 µM for PXR or 17β-estradiol which had a steroid core that was present in most of the compounds. Further, the quality of the scoring function was assessed using standard statistical indicators namely sensitivity (SE), specificity (SP), overall prediction accuracy (Q) and Matthews correlation coefficient (C) (Table 1) and were derived as described previously [22].
Bayesian models were generated using Discovery Studio 2.1 (Accelrys, San Diego, CA) Laplacian-corrected Bayesian classifier [37],[42],[43],[45],[56]. FCFP_6 fingerprints, AlogP, molecular weight, number of rotatable bonds, number of rings, number of aromatic rings, number of hydrogen bond acceptors, number of hydrogen bond donors and molecular fractional polar surface area were calculated from the input sdf file using the “calculate molecular properties protocol”. The “create Bayesian model protocol” was used for model generation and a custom protocol for validation (leave out 20% 100 times) was used.
5D-QSAR studies were performed using Raptor [40]. Raptor includes the possibility of representing each ligand molecule as an ensemble of conformations, orientations, stereoisomers and protonation states (4D-QSAR), thereby reducing the bias in identifying the bioactive conformer. In addition, it explicitly allows for induced fit by a dual-shell representation of the three-dimensional binding-site model, onto which the physicochemical properties (hydrophobicity and hydrogen-bonding propensity) are mapped (5D-QSAR). The inner shell is tailored using the most potent ligand of the training set, the outer shell accommodates the topology of all molecules from the training set. The adaptation of both field and topology of the receptor surrogate to each ligand is achieved by combining a steric adjustment to the topology of every ligand and a term due to the attraction or repulsion between ligand and receptor model. The latter is obtained by correlating their physicochemical properties (hydrophobicity and hydrogen-bond propensity) in 3D space. Since the mapping of properties onto the shells is not unambiguously determinable, different models with similar predictive power can be identified. Raptor generates a family of receptor models. Such model families may be interpreted to represent the various configuration states of the true biological receptor. The obtained binding affinities are averaged over the individual models.
The underlying scoring function for evaluating ligand-protein interactions includes directional terms for hydrogen bonding (ΔGHbond), hydrophobicity (ΔGHphob) as well as terms for the cost of the topological adaptation (ΔGIF) and the changes in entropy (TΔS) upon ligand binding: ΔGbinding = ΔGconstant + ΔGHbond + ΔGHPhob − TΔS + ΔGIF .
Experimental determination of binding affinity for weak inhibitors is often prevented due to limited solubility or limited sensitivity. Thus, only an upper limit (‘threshold’) for Ki values is accessible. To prevent artificial assignment of affinities in a QSAR study including weak binders, the Raptor concept allows the use of a threshold option: the optimization algorithm forces the model to reproduce the binding affinity of the weak- and non-binding ligand molecules to be lower than the experimental limit. Obviously, compounds which are experimentally measured to bind weaker than a threshold Ki(t) and are correctly classified during the model optimization, no penalty is added to the lack-of-fit value, if, on the other hand, the binding affinity of the ligand is predicted higher than the threshold, the lack-of-fit function applies a penalty proportional to ΔGbinding(t) − ΔGbinding.
4D sets of alternative conformations for each ligand as input for Raptor were performed with Symposar [57]. In Symposar the ligand molecules are superimposed onto one or several template molecules, first, on the basis of fuzzy-like 2D substructure similarities and, subsequently, in 3D space with respect to their similarity of physicochemical fields. This two-step process combines the speed of a 2D similarity search with the accuracy and authenticity of protein-ligand interactions in 3D space. The molecules are thereby treated as flexible and are fully relaxed at the end of the alignment process.
|
10.1371/journal.pcbi.1002935 | Detecting DNA Modifications from SMRT Sequencing Data by Modeling Sequence Context Dependence of Polymerase Kinetic | DNA modifications such as methylation and DNA damage can play critical regulatory roles in biological systems. Single molecule, real time (SMRT) sequencing technology generates DNA sequences as well as DNA polymerase kinetic information that can be used for the direct detection of DNA modifications. We demonstrate that local sequence context has a strong impact on DNA polymerase kinetics in the neighborhood of the incorporation site during the DNA synthesis reaction, allowing for the possibility of estimating the expected kinetic rate of the enzyme at the incorporation site using kinetic rate information collected from existing SMRT sequencing data (historical data) covering the same local sequence contexts of interest. We develop an Empirical Bayesian hierarchical model for incorporating historical data. Our results show that the model could greatly increase DNA modification detection accuracy, and reduce requirement of control data coverage. For some DNA modifications that have a strong signal, a control sample is not even needed by using historical data as alternative to control. Thus, sequencing costs can be greatly reduced by using the model. We implemented the model in a R package named seqPatch, which is available at https://github.com/zhixingfeng/seqPatch.
| DNA modifications have been found in a wide range of living organisms, from bacteria to human. Many existing studies have shown that they play important roles in development, disease, bacteria virulence, etc. However, for many types of DNA modification, for example N6-methyladenine and 8-oxoG, there is not an efficient and accurate detection method. Single molecule real time (SMRT) sequencing not only generates DNA sequences, but also generates DNA polymerase kinetic information. The kinetic information is sensitive to DNA modifications in the sequenced DNA template, and therefore can be used for detecting a wide range of DNA modification types. The usual detection strategy is a case-control method, which compare kinetic information between native sample and a control sample whose modifications have been removed. However, generating a control sample doubles the cost. We proposed a hierarchical model, which can incorporate existing SMRT sequencing data to increase detection accuracy and reduce coverage requirement of control sample or even avoid the need of a control sample in some cases. We tested our method on SMRT sequencing data of plasmids with known modified sites and E. coli K-12 strain to demonstrate our method can greatly increase detection accuracy and reduce sequencing cost.
| Modifications to individual bases like 5-methylcytosine, 5-hydroxymethylcytosine, and N6-methyladenine in DNA sequences are an important epigenetic component to the regulation of living systems, from individual genes to cellular function. Single molecule, real time (SMRT) sequencing provides a high throughput platform for direct DNA modification detection without the need for special sample preparation procedures such as bisulphite treatment or restriction enzyme digestion [1]–[3]. In SMRT sequencing, each base identity is read when fluorescently labeled nucleotides are incorporated into a DNA sequence being synthesized by DNA polymerase [4]. In this case, because the incorporation events are being directly observed in real time, the duration between the pulses of light (referred to as inter-pulse duration or IPD) that indicate an incorporation event can be precisely measured. IPD measures are a direct reflection of the DNA polymerase kinetics. This kinetic parameter for the enzyme has been shown to be sensitive to a wide range of DNA modification events, including 5-methylcytosine, 5-hydroxymethylcytosine, and N6-methyladenocine [1]–[3], where variations in the kinetics are predictive of modification events.
For each position in the DNA sequence being synthesized, the IPD distribution is empirically determined as each read covering a given position yields an IPD value for that position, so that for each position there are a number of IPD observations. In these previous demonstrations [1], [2], kinetic variations were detected using a case-control method in which the IPDs at a given site in the native DNA from a sample of interest (case group) are compared to the IPDs in whole-genome amplified (WGA) DNA corresponding to the native DNA (control group). The WGA process erases all of the modifications by replacing any modified base with the corresponding standard base. The null IPD distribution can be determined from the IPDs in the control group and then the IPD distribution for the case group can be compared to this null distribution (Figure 1A). If the IPD values between cases and controls differ significantly, then a kinetic variation event is called. Because SMRT sequencing reads are strand specific with respect to the detection of these kinetic variation events, modifications can be inferred in a strand specific manner. This approach to detecting kinetic variation events works well when there is sufficiently high numbers of reads covering each position, but is much less reliable in low coverage cases due to the high variability of IPD measures (the IPDs are exponentially distributed). In addition, this case-control method requires sequencing a sample twice, so making these detections come at a significant cost.
In this paper, we examine the correlation between polymerase kinetics and sequence context to demonstrate that polymerase kinetics can be well predicted by local sequence context, suggesting that baseline kinetics can be established for any sequence context to use as a null distribution in testing for base modification events. We demonstrate that this correlation between local sequence context and enzyme kinetics is highly consistent across independent experiments carried out on DNA from different species. Given this, we hypothesized that IPDs from positions with the same sequence context, referred to as homologous positions, including those from historical control data, could be used jointly to better estimate the null IPD distribution. Towards that end, we develop a hierarchical model to combine IPDs across homologous positions to enhance the detection of kinetic variation events. The hierarchical model can work with or without control data. When control data are available, for a given position, the hierarchical model combines IPDs of control data and IPDs of homologous positions to estimate the null IPD distribution. We refer to this type of model as a hierarchical model with control data. When there is no control data available, for a given position, the hierarchical model estimates the null IPD distribution using only IPDs of homologous positions from historical data. We refer to this as a hierarchical model without control data. We test these two hierarchical models on two high coverage plasmid datasets and a medium coverage E. coli K-12 MG 1655 dataset: 1) plasmid DNA isolated from a strain of E. coli engineered to methylate the 4th carbon in cytosine residues, referred to as 4-mC, in the GATC context; 2) plasmid DNA isolated from a strain of E. coli engineered to methylate the A residue in the GATC context, referred to as 6-mA; and 3) DNA isolated from a wild type E. coli reference strain (K-12) (Table 1). We show that the hierarchical model with control data significantly increases the detection accuracy compared with the case-control design on all of the datasets. The hierarchical model without control data also achieves a good accuracy for N6-methyladenocine, which has a strong signal-to-noise ratio (i.e. impact on the enzyme kinetics), but does not work well for methylcytosine, whose signal-to-noise ratio is relatively weak. In the case of the E. coli K-12 dataset, we were able to detect roughly 80% of the 6-mA events in the GATC context at a 5% FDR (False Discovery Rate) using the hierarchical model with control data, a context known to be methylated in a vast majority of the occurrences of the GATC motif in this strain [5]. In addition to detecting these known methyladenine events in the GATC context, we demonstrate the detection of thousands of kinetic variation events that occur at positions not previously described as having known methylation motifs, suggesting more extensive patterns of modification than had been previously observed.
We note that while the IPDs are observed to be exponentially distributed, tests based on this assumption are more sensitive to extreme outliers. Thus, we adopt a Box-Cox transformation to make the IPDs follow an approximate normal distribution (Figure 2), making it more robust to outliers. Formally, we used the following transformation,Because the chemistry of SMRT sequencing is being constantly improved, there are two different types of chemistry, FCR and C2, represented in our datasets. The IPD characteristics of these different chemistries are quite different, so we used different and , which were estimated for each set. For data using the FCR chemistry, we used and . For data using the C2 chemistry, we used and . The and parameter values were chosen empirically such that the skewness distribution was approximately centered at 0.
A number of factors can influence enzyme kinetics in addition to sequence context (see below) and DNA modification, including reagent lot, temperature, SMRTcell lot and instrument operator. Just as we observe batch effects and other experimental noise factors with other technologies such as microarrays and RNA-seq that impact gene expression values, so these different effects can have strong effects on the IPD. Therefore, IPDs from different experiments are not necessarily directly comparable. For the current version of the Pacific Biosciences RS DNA sequencing instrument, DNA molecules are sequenced in zero mode waveguides (ZMWs) located on a SMRTcell [6], with pulses of light in different color channels corresponding to the bases being incorporated into the sequence being synthesized. These signals are detected and recorded by a CCD camera operating at 100 Hz, resulting in a movie containing up to 150,000 pulse streams corresponding to the different ZMWs on the SMRTcell. Overall, the IPD distribution can be significantly different between movies even for identical DNA samples (Figure 1B). Therefore, we applied a simple centering approach to normalize the IPD data before modification detection.where is any Box-Cox transformed IPD in a movie, and N is the number of alignable bases in that movie. In the rest of the paper, we refer to the normalized IPD as simply the IPD.
The kinetic rate of DNA polymerase is known to be sensitive to sequence context [7]. Given the ability of SMRT sequencing to observe many thousands of individual molecules of DNA polymerase as they carry out DNA synthesis, we examined the relationship between the kinetic rate (estimated from the IPDs) and sequence context. For each position, the position-specific kinetic rate is defined as the mean of its Box-Cox transformed IPDs (Methods). Sequence context is defined as the sequence flanking the incorporation site of interest, the boundaries of which are explored below. To avoid ambiguity caused by modification events, we explored enzyme kinetics using whole-genome amplified (WGA) E. coli K-12 data (E. coli WGA-FCR in Table 1), given the WGA process erases all chemical modifications. From the K-12 dataset, we extracted positions in which the single strand coverage was greater than 35 reads. We then applied MART [8], a non-linear tree based regression method, to estimate the relationships between polymerase kinetics and sequence context. Here, position-specific kinetic rate is the response variable and sequence context is the predictor variable. The proportion of the variation in the response variable that can be explained by the predictor variable (i.e., the value), was used as the measure of dependence of enzyme kinetics on sequence context. We explored these relationships over different sequence context lengths and found that grows as the number of bases upstream of the incorporation site increases, but becomes saturated at 7 bases upstream. The bases downstream from the incorporation site have much smaller impact on the enzyme kinetic rate, with positions more than 2 bases downstream from the incorporation site having no observable impact on the values (Figure 3A). Roughly 80% of the IPD variation can be explained by a 10 base pair sequence context (7 bases upstream and 2 bases downstream from the incorporation site).
We refer to the average Box-Cox transformed IPDs corresponding to positions with the same sequence context as the context effect. We examined the consistency of the context effect between two independent experiments: 1) WGA data from the E. coli K-12 strain, and 2) WGA data from M. pneumoniae (E. coli WGA-C and M. pneumoniae WGA-C2 in Table 1). While these experiments were performed completely independently, carried out by two different groups at two geographically separated sites, the context effects were strikingly similar (Figure 3B), with 80% of the IPD variation in one set explained by variation in the second set. Importantly, we compared context effects between two experiments using the same chemistry (FCR chemistry), as the consistency of context effects will not hold when comparing experiments using different chemistries. Thus, in all the experiments carried out herein, only datasets with the same chemistry are used together.
Given that a large percentage of variability of position specific enzyme kinetic rates can be explained by sequence context (Figure 3A), IPDs of homologous positions can be combined together to estimate the null IPD distribution. In addition, because of the high consistency in context effects across different experiments, IPDs of homologous positions in historical WGA datasets can also be incorporated to enhance the power to detect kinetic variation events (Figure 4). However, the IPD distributions of homologous positions will not be exactly the same, and so, false positive calls may be introduced if the null IPD distribution is estimated without considering the heterogeneity in the IPD distributions that can exist between homologous positions. To deal with this type of heterogeneity, we developed an hierarchical model to incorporate IPDs of homologous positions in a robust fashion. In the hierarchical model, Box-Cox transformed IPDs of homologous positions were assumed to follow normal distributions, with differences in mean and variance allowed between these distributions. The mean and variance parameters were treated as random variables and were assumed to follow the same prior distribution. For the hierarchical model with control data, the model was fitted by both IPDs from the control data and as well as IPDs from all homologous positions. For the hierarchical model without control data, the model was fitted using only homologous positions in historical data. Then, we adopt a likelihood ratio to evaluate how likely a position is modified (See Methods).
To assess the utility of the hierarchical model in detecting kinetic variation events, we compared the naive case-control design with the hierarchical model using data from two data sets in which the sites that were modified were known a priori. The first set was generated from plasmid DNA isolated from a strain of E. coli engineered to methylate the 4th carbon of cytosine residues in each GATC context (referred to as the 4-mC set), and the second was generated from plasmid DNA isolated from a strain of E. coli engineered to methylate adenine residues in each GATC context (referred to as the 6-mA set). We use E. coli WGA-FCR and M. pneumoniae WGA-FCR in Table 1 as historical data and only contexts that have more than 5 positions with larger than 10x coverage in the historical data are used. We explored both the [−7,+2] contexts (7 bp upstream, 2 bp downstream of the incorporation site) and [−6,+1] contexts, and found that their performances were similar (Figure 5). However, roughly one third of the positions in these datasets did not have the corresponding [−7,+2] context in the historical data. Therefore, to make the comparisons fair, we only considered positions that had a corresponding context in the historical data. To maximize the number of sequence contexts in one dataset that would be represented in another, we restricted the sequence context to 8 bases(the [−6,+1] context) for the remainder of our study. For the hierarchial model with control data (see Methods), when the sequence coverage of the control sample is relatively low (15x35x single strand coverage), the hierarchical model compared to the case-control method is seen to increase the sensitivity by 10%30% under the same FDR (Figure 5). Performance of the case-control method and hierarchical model become similar as sequencing coverage of the control sample increases. For the hierarchical model without control (See Methods), the accuracy is comparable to the case-control method for 6-mA, but does not perform as good for 4-mC.
We further tested our method on datasets with partial modifications in which only a fraction of the molecules with respect to a given position were modified. Given one of the limitations with the modeling approach presented herein is that it assumes in the native sample a given position is either fully modified or not, understanding the sensitivity of this assumption on the detection rates is important. For each of the 6-mA set and 4-mC datasets, we constructed an artificial native sample, where only a certain proportion of reads were sampled from the native sample (referred to as the modification proportion), while others were sampled from the control data. We tested the performance of our method on three different modification proportions: 50%, 70% and 90%. As expected, the detection accuracy decreases as the modification proportion decreases, however, in all cases our method was still able to make detections even when the fully modified assumption was clearly violated. We further note that it is possible to achieve accuracy that is comparable to the case of fully modified positions by increasing the sequence coverage of the native sample (Figure S1).
To further evaluate the performance of the hierarchical model, we explored data from the E. coli K-12 strain in which there are not only modifications that are known to occur in certain sequence motifs (e.g. GATC), but also potentially novel modification events that cannot be explained by known motifs. We applied both hierarchical modeling with and without control data to the SMRT sequencing data from native the E. coli K-12 MG 1655 strain (E. coli native in Table 1). For the hierarchical model with control data, we used data from a WGA E. coli K-12 MG 1655 sample as control (E. coli WGA-C in Table 1) and a WGA M. pneumoniae data (M. pneumoniae WGA-C2 in Table 1), which is generated in another unrelated experiment, as historical data. For each given position, the IPD distribution in the native sample was compared to the null IPD distribution, which was estimated by fitting a hierarchical model that combines the IPDs of the corresponding positions in the control sample and IPDs of all of the homologous positions. Homologous positions were identified in two different ways: 1) find homologous positions in the WGA M. pneumoniae data, and 2) find homologous positions in the control data. For the hierarchical model without control, we estimated the null distribution by fitting the hierarchical model by the homologous positions in the WGA M. pneumoniae data only. A position was called modified if the generated likelihood ratio exceeded a certain threshold (see Methods).
As most adenines in the GATC context are expected to be methylated in wild type E. coli K-12 MG 1655, we detected modifications in the regions within 20 bp around adenine positions in the GATC context to evaluate how well 6-mA could be detected. Here, the FDR is estimated as the ratio between the number of significant adenines detected that are not in the GATC context and the total number of significant adenines detected. We note that it is certainly possible that there are modified bases outside of the GATC context, so that treating only adenines detected in the GATC context as true positives and all other bases as true negatives, out estimated FDR can be considered as a conservative estimation, i.e. the actual FDR is lower than this. The receiver operating characteristic (ROC) curve (Figure 6A) shows that 95% of adenine of GATC can be detected under FDR of 5% by using the hierarchical model with control. The hierarchical model greatly increases the detection accuracy in this instance compared to the case-control method. We can also detect 6-mA without the control data, where the accuracy is lower than hierarchical model with control, but the results are comparable to the naive case-control method. If we apply this detection approach on genome-wide scale, we detect many putative modification events, with the ROC curve (Figure 6B) showing that at the 5% FDR we not only detect 80% of the adenine in the GATC context by using hierarchical model with control, but we also identify about 2000 other positions in other contexts that may reflect off target activity of the methyltransferase that makes the 6-mA modifications in the GATC context or perhaps reflects the activity of other enzymes capable of inducing base modifications. In the genome-wide study, to estimate FDR, we detected DNA modifications in another WGA sample (E. coli WGA-N in Table 1), where no modification should be found, and FDR is estimated by ratio between number of DNA modifications detected in the WGA sample(E. coli WGA-N in Table 1) and number of modifications detected in the native sample (E. coli native in Table 1).
We examined the correlation between DNA polymerase kinetics and sequence context quantitatively and found that roughly 80% of the variation in the enzyme kinetics as measured by IPD variance can be explained by sequence context. Our data support that the most informative regions of sequence context for the enzyme kinetics at a given incorporation site is the region 7 bp upstream and 2 bp downstream of the incorporation site. In addition, we found that this context dependence is extremely consistent between independent SMRT sequencing experiments carried out using the same chemistry. IPDs of homologous positions, including those from historical control data can therefore be incorporated to improve DNA modification detection accuracy. However, because heterogeneity of the IPD distribution within the same sequence context can cause false positive events, we adopted a hierarchical model that can adaptively incorporate information from homologous positions. The hierarchical model is flexible in that it can be used with or without control data. We demonstrated that the hierarchical model with control data can greatly increase accuracy compared to the naive case-control method. For the types of modifications that have a relatively weak signal-to-noise ratio, such as 4-mC, the hierarchical model without control does not perform as good as the case-control method. This may be expected given the sequence context in such instances does not appear to explain all of the kinetic variation, with other factors such as fragment length and experimental condition perhaps dominating the estimation of the null distribution from historical data. However, for modification types with a strong signal-to-noise ratio, noisy null IPD distributions have a relatively small impact on accuracy.
Our results suggest that the hierarchical model can reduce the requirement of control samples and thus provide a significant cost benefit. For detecting modifications with a strong signal-to-noise ratio, one can generate low coverage control data or even avoid the generation of the control data altogether. It may be possible in the future as more SMRT sequencing data obtains, given the dependence of local sequence context on enzyme kinetics, to build null models specific to each sequence context to leverage as a control in detecting base modification events. We anticipate as well that as more sequence data obtains across different species with larger genomes than the prokaryotic genomes represented in our study, that we will be able to re-evaluate whether a more expanded sequence context around the incorporation site better explains the DNA polymerase enzyme kinetics. It may be that with an expanded set that considers 9 bases upstream of the incorporation site and 3 bases downstream, for example, a better explanation of the enzyme kinetics obtains. Further, as the historical datasets get larger, we may also find that the historical data on its own achieves the same results as the combined historical and control data in all contexts and for all modification types.
For the th position in the genome, we assume that its Box-Cox transformed IPDs follow a normal distribution, which iswhere is the jth Box-Cox transformed IPD of the th position in the genome, and is the position specific polymerase kinetic rate, which is by its sequence context. We used ( is the single strand coverage of the th position), which are the estimated position specific polymerase kinetic rates, as the response, and the corresponding sequence context as the predictor to build a non-linear regression model using the MART method [8]. The dataset we employed for this characterization is whole genome amplified E. coli K-12 data after outlier removal and coverage filtering. Each data point in the dataset is a pair of estimated position specific polymerase kinetic rates and sequence contexts, which represent the upstream and downstream bases of the position of interest. Performance of the regression approach was evaluated using 5-fold cross validation in which 80% of the data points were randomly selected as the training set, MART was trained on this set, and then the predicted responses were carried out for the remaining 20% of the data set. The was the statistic used to measure the performance and was calculated as , where is the sample size, , and is the predicted .
The hierarchical model without control data is a special case of the hierarchical model with control data, i.e. or vector is empty.
We evaluated case-control method and hierarchical model on two different datasets: 3589 bases long plasmid with 19 known 4-methylcytosines, and 3591 bases long plasmid with 23 known N6-methyladenines (Plasmid m4C native/control and Plasmid m6A native/control in Table 1). Whole genome amplified E. coli and M. pneumoniae data were used as historical data (E. coli WGA-FCR and M. pneumoniae WGA-FCR in Table 1). A detection is called correct only if its distance to the nearest true modified position is less than or equal to 5 bp. Different thresholds of were set, and the corresponding false discovery rate and true positive rate were calculated. False discovery rate and true positive rate are defined asrespectively. To get the ROC under different coverage, we randomly sampled reads without replacement 100 times to get average FDR under different TPRs.
The raw sequence data listed in Table 1 are available at http://www.ncbi.nlm.nih.gov/sra, under accession number SRA062773 and SRA058893. (SRA058893 was published in [10]).
|
10.1371/journal.pgen.1002869 | Comparative Analysis of the Genomes of Two Field Isolates of the Rice Blast Fungus Magnaporthe oryzae | Rice blast caused by Magnaporthe oryzae is one of the most destructive diseases of rice worldwide. The fungal pathogen is notorious for its ability to overcome host resistance. To better understand its genetic variation in nature, we sequenced the genomes of two field isolates, Y34 and P131. In comparison with the previously sequenced laboratory strain 70-15, both field isolates had a similar genome size but slightly more genes. Sequences from the field isolates were used to improve genome assembly and gene prediction of 70-15. Although the overall genome structure is similar, a number of gene families that are likely involved in plant-fungal interactions are expanded in the field isolates. Genome-wide analysis on asynonymous to synonymous nucleotide substitution rates revealed that many infection-related genes underwent diversifying selection. The field isolates also have hundreds of isolate-specific genes and a number of isolate-specific gene duplication events. Functional characterization of randomly selected isolate-specific genes revealed that they play diverse roles, some of which affect virulence. Furthermore, each genome contains thousands of loci of transposon-like elements, but less than 30% of them are conserved among different isolates, suggesting active transposition events in M. oryzae. A total of approximately 200 genes were disrupted in these three strains by transposable elements. Interestingly, transposon-like elements tend to be associated with isolate-specific or duplicated sequences. Overall, our results indicate that gain or loss of unique genes, DNA duplication, gene family expansion, and frequent translocation of transposon-like elements are important factors in genome variation of the rice blast fungus.
| Magnaporthe oryzae is the causal agent of rice blast that is mainly controlled with resistance cultivars. However, genetic variations in the pathogen often lead to overcoming R gene-mediated resistance in rice cultivars. In this study we sequenced two field isolates from China and Japan. In comparison with the laboratory strain that was previously sequenced, the field isolates have a similar genome size and overall genome structure. However, they have slightly more genes and contain a number of expanded gene families that are likely involved in plant-fungal interactions. Each of the isolates has specific genes, some of which affect virulence and some others are important for asexual development. The three strains differ noticeably in the distribution of transposon-like elements. Many of the transposable elements tend to be associated with isolate-specific or duplicated sequences. This study revealed genetic factors involved in genome variation of the rice blast fungus.
| Rice blast caused by the heterothallic ascomycete Magnaporthe oryzae (also known as Pyricularia oryzae) is one of the most destructive diseases of rice, which is a staple for over half of the world's population. This pathogen also infects wheat and other small grains, and poses major threats to global food security [1], [2]. In the past two decades, rice blast has been developed as a model system to study fungal-plant interactions. M. oryzae was the first plant pathogenic fungus to have its genome sequenced and made available to the public [3].
In most parts of the world, rice blast is controlled mainly with resistant cultivars. However, M. oryzae is notorious for its ability to overcome resistance based on race-specific R genes [4]–[6]. New cultivars often lose their resistance within a few years of introduction. Genetic variations in populations of the pathogen have been well-documented in many parts of the world [7], [8]. M. oryzae isolates are also known to lose virulence and female fertility during laboratory manipulations [1] and large chunks of genomic DNA can be lost spontaneously during cultivation on artificial media, such as the deletion of over a 40 kb region containing the BUF1 locus [9]. The laboratory strain 70-15 of M. oryzae was generated by backcrossing a progeny from a cross between a rice isolate and a weeping love grass (Eragrostis curvula) isolate with the rice isolate Guy11 from French Guyana [10], [11]. It has been used in many laboratories and was selected for genome sequencing [3]. Although most of the 70-15 genome should be from the rice pathogen after backcrossing with Guy11 several times, some weeping love grass pathogen sequences are likely retained. In comparison with Guy11, 70-15 is reduced in female fertility, conidiation, and virulence [12].
To determine the extent of genetic variation among isolates of M. oryzae, we sequenced two field isolates Y34 and P131. Y34 was isolated from Japonica rice in 1982 in Yunnan province, China, where both Indica and Japonica rice cultivars are cultivated [13], [14]. Due to rich genetic diversity in rice cultivars and centuries of rice cultivation, highly diverse rice blast pathogen populations exist in Yunnan [15], and hence Y34 was chosen as a representative from this region for sequencing. The other field isolate, P131, originated from Japan where Japonica rice cultivars are dominant [16], [17]. The isolates P131, Y34, and 70-15 differ in some cultural characteristics (Figure S1). These three isolates also carry different avirulence genes and vary in aggressiveness toward different rice cultivars (Table S1). In comparison with 70-15, both Y34 and P131 have slightly larger genomes. The two Asian field isolates share a higher degree of similarity and contain over 200 genes that are absent in 70-15. Many pathogenesis-related genes showed evidence of exposure to diversifying selection when comparing either field isolate (P131 or Y34) to the laboratory strain (70-15). Functional characterization of randomly selected genes specific to the field isolates revealed that they play diverse roles, some of which affect virulence and others important for conidiation and vegetative growth. Furthermore, thousands of loci with transposon-like elements were identified in each genome. Many of them tend to be associated with the distribution of unique sequences and translocation of duplicated genes.
The genomes of P131 and Y34 were sequenced with the Sanger (2-fold) and 454 sequencing technologies (18-fold). The combined sequence reads for P131 and Y34 were 793.94 Mb and 843.92 Mb, representing about 20- and 21-fold genome sequence coverage, respectively (Table 1). The 454 sequence reads were assembled into contigs and placed into scaffolds by the Newbler assembler with paired-end information from the Sanger reads. The assembled P131 genome consisted of 1,823 scaffolds with a combined length of 37.95 Mb. The N50 and maximum lengths of P131 scaffolds were 65 kb and 459 kb, respectively (Table 1). The Y34 genome was assembled into 1,198 scaffolds with a combined length of 38.87 Mb. The N50 and maximum length of Y34 scaffolds were 106 kb and 708 kb, respectively (Table 1). Over 95% of the sequence reads were assembled into scaffolds >5 kb in both isolates. Approximately 33% and 51% of P131 and Y34 sequences, respectively, were assembled into scaffolds longer than 100 kb. In addition, the mitochondrial genomes of P131 and Y34 were also assembled (Table 1). While P131 has an almost identical mitochondrial genome with 70-15, Y34 lacks two short fragments with a combined length shorter than 350 bp (Figure S2).
Because repetitive sequences comprise approximately 10% of the genome of the laboratory strain 70-15 (version 6), repetitive sequences in the new assemblies were masked out with the RepeatMasker program for comparative analyses. The resulting ATCG bases after masking were 37.6 Mb, 38.2 Mb, and 37.5 Mb, respectively, for P131, Y34, and 70-15 (Table 1), indicating that the core genomes of these three isolates were not significantly different in size. However, because repetitive sequences and singletons smaller than 2 kb were not included in this analysis, it remains possible that the complete genomes of these three isolates vary in abundance of repetitive sequences and actually have greater size differences.
Scaffolds of P131 and Y34 were aligned with the assembled genome of 70-15 (Figure 1). Overall, most of the 70-15 genome (96%) is also conserved in two field isolates. Only 0.45 Mb of sequence in 70-15 are absent from the two field isolates. In contrast, P131 and Y34 have 1.69 Mb and 2.56 Mb isolate-specific sequences, respectively. In general, isolate-specific sequences were dispersed throughout the genomes. For individual chromosomes, there are regions enriched for isolate-specific sequences (Figure 1). Blocks of such sequences can be found at both ends of chromosome IV and at single ends of chromosomes I, II, III, V, and VI. In M. oryzae, genetic variation and avirulence genes are known to be enriched near the telomeres [18], [19]. Comparative analysis of the genomes of these three M. oryzae isolates revealed that genes responsible for variations in virulence and adaptation to the environment may be concentrated at the chromosomal ends.
To locate and verify isolate-specific sequences in the field isolates, we used clamped homogenous electric fields (CHEF) gel electrophoresis to separate the chromosomes. Chromosome size polymorphisms were observed among these three isolates (Figure 2A). Whereas chromosome VII (the smallest chromosome) in 70-15 was estimated to be 4.3 Mb, the smallest chromosomes in Y34 and P131 were approximately 1.8 Mb and 2.5 Mb, respectively. When one P131-specific sequence, P131_scaffold00006_11, which was not mapped on the chromosome alignment was used as the probe, an aggregate band of chromosomes larger than 6.0 Mb was detected in P131 but not in Y34 nor in 70-15 (Figure 2B). When a similar blot was probed with an Y34-specific sequence, Y34_scaffold00824_1665, only the smallest chromosome of Y34 was hybridized (Figure 2B). These findings confirm that the field isolates contain isolate-specific DNA.
Because the assembly of P131 or Y34 relied on the alignment with the 70-15 genome, it was not possible to accurately map P131 and Y34 sequences that were absent from the 70-15 genome assembly. However, the P131 and Y34 sequences could be used to fill the sequence gaps (≥50 bp) in the 70-15 assembly. We identified the end sequences of the contigs or scaffolds flanking these gaps. After filtering out simple repeats, these sequences were used to search against the assembled P131 and Y34 sequences. If both upstream and downstream flanking sequences of one gap were mapped on the same contig in either P131 or Y34, the in-between sequences were used to fill the gaps of 70-15. A total of 55 gaps were filled with sequences from P131 or Y34 (Table 2). Among them, 35 gaps had the sequences present in both P131 and Y34 (Table 2). The total gap sequence filled in the 70-15 genome was 25.3 kb. We randomly selected 18 of these filled gaps of the 70-15 genome for verification. All of them were confirmed in 70-15 by PCR (Figure S3).
The number of predicted genes in the masked genomes of P131, Y34, and 70-15 was 12,714, 12,862, and 12,440 (Table 1), respectively. The average length of predicted proteins was over 400 amino acids. Y34 apparently has the largest genome size and gene content, which may contribute to its adaptation to the environment or to rice cultivars grown in Yunnan province, China. To identify the gene pool of these three strains, the predicted amino acid sequences of the total gene set from each isolate were used to search against the nucleotide sequences of other two isolates by TBLASTN. The large majority of M. oryzae genes (12,375 from P131, 12,431 from Y34, and 12,214 genes from 70-15) share sequence homology in pair-wise comparisons. Among these genes constituting the ‘core’ gene set of the M. oryzae genome (Figure 3A), 11.3% had no orthologous sequences in other organisms. Moreover, approximately 10.1% of these M. oryzae-specific genes were predicted to encode secreted proteins.
To improve gene annotation in 70-15, we identified the genes that were common to all three isolates and had similar sizes (difference less than 1%) between Y34 and P131 but were 50 amino acids or 20% longer or shorter in 70-15. A total of 340 genes meeting these criteria were then manually annotated. Among them, 135 genes in 70-15 had incorrect intron annotations. The number of genes with inaccurate start or stop codon predictions was 259 or 15, respectively (Table S2).
The number of genes shared only by the two field isolates (198 from P131 and 220 from Y34) was approximately twice that of those shared by either P131 or Y34 with 70-15 (Figure 3A), implying that the two Asian field isolates share a higher degree of similarity and with about 200 genes that are absent in 70-15. For isolate-specific genes, we found that 51, 136, and 71 genes were unique to P131, Y34, and 70-15, respectively (Figure 3A). All the genes randomly selected for verification were confirmed by PCR to be either shared by two isolates or unique to one specific isolate (Figure S4). As found in 70-15, isolates P131 and Y34 also had various copies of DNA helicase Q genes and LTR elements towards the chromosomal ends [3].
For the genes common in Y34 and P131 but absent in the automated annotation of 70-15, we used their amino acid sequences to search the 70-15 scaffolds. The resulting homologous sequences of 70-15 were then used to search against M. oryzae ESTs deposited in GenBank. A total of 81 candidate genes were identified in the 70-15 genome and ESTs (Table S3). Seventy-six of them encoded hypothetical proteins with no known homologs in GenBank. Some of these M. oryzae specific genes may be important for the virulence or fitness of the pathogen because all three isolates have these genes. The other five genes had orthologous sequences of unknown functions in Sordariomycetes but were absent in lower fungi, such as Zygomycetes and Saccharomycetales.
To further analyze genetic relatedness of these three isolates, the 10,074 clusters containing one protein from one isolate were selected and the resulting individual protein sequences from each isolate were combined for distance analysis with PHYLIP. As shown in Figure 3B, the two field isolates have a closer relationship to each other than with the laboratory strain 70-15.
Based on analyses of gene content, 51, 136, and 71 genes, respectively, were unique to P131, Y34, and 70-15. Overall, 13% of these isolate-specific genes encoded secreted proteins and 46% of them had no significant homolog in GenBank (Table S4). RT-PCR analyses were performed with 10 and 14 randomly selected P131- and Y34-specific genes, respectively. All the selected genes were confirmed to be expressed in mycelia (Figure S5). While most of the isolate-specific genes were dispersed through the genome, some were located within clusters (Figure 1; Table S4). For example, scaffolds 00875 and 01112 of Y34 contained five and eight of the Y34-specific genes, respectively. In P131, there were three isolate-specific genes each on scaffolds P131_scaffold01777 and P131_scaffold01784. Moreover, many of the isolate-specific genes with known chromosomal positions in P131 and Y34 were located near the chromosomal ends (within 500 kb), which is consistent with the distribution tendency of isolate-specific sequences (Figure 1).
To determine the biological function of these isolate-specific genes, nine Y34-specific genes and three P131-specific genes were selected for functional characterization. For majority of them, the resulting gene deletion mutants had no obvious changes in colony growth, conidiation, or virulence (Figure S6). Their functions in plant infection may be redundant or too minor to be detected under laboratory conditions. However, deletion of one P131 unique gene, P131_scaffold00208-2, resulted in a reduction in virulence in infection assays with seedlings of a susceptible rice cultivar (Figure 4A). Deletion of another P131 unique gene, P131_scaffold01777-7, resulted in approximately 10% growth reduction on oatmeal tomato agar plates (Figure 4B). Proteins encoded by these two P131-unique genes were predicted to be localized in the nucleus. Homologous sequences of these two genes were not found in other sequenced fungal species. Moreover, deletion of one Y34 unique gene, Y34_scaffold00875-3, resulted in approximately 36% reduction in conidiation (Figure 4C). Interestingly, deletion of one Y34-unique gene encoding a putative G protein-coupled receptor (GPCR)-like integral membrane protein with six transmembrane domains resulted in changes in pathogenicity on a rice cultivar carrying the Pi-7 R gene, suggesting that this Y34-unique gene might be the potential AVR Pi-7 gene (data not shown).
Among the genes shared by both field isolates P131 and Y34 but absent in 70-15, 19% had signal peptides for secretion and 12% had transmembrane domains (Figure 3; Table S5). About 70% of these genes had no functional annotation. Strain 70-15 may have lost these genes during the initial genetic cross or after generations of cultivation in the laboratory. For example, a gene encoding a CFEM-containing GPCR-like protein [20] and the avirulence gene AVR Pi-a [21] were present in the field isolates P131 and Y34 but not found in 70-15.
Duplication is one of the major mechanisms for evolutionary innovation. The total duplicated genomic DNA fragments (longer than 500 bp and greater than 90% identity) were 289 kb, 385 kb, and 825 kb in P131, Y34, and 70-15, respectively. A total of 16, 20, and 155 predicted genes in P131, Y34, and 70-15, respectively, were located in these duplicated sequences (Table S6).
Although duplicated DNA sequences were detected genome-wide in all three isolates, in general chromosomes II, IV, V, and VII had more duplicated DNA sequences than other chromosomes (Figure 5A). For individual chromosomes, the end regions tend to contain more duplicated DNA sequences than the central region. Comparative analysis indicated that P131, Y34, and 70-15 all contained isolate-specific duplicated regions (Figure 5A). However, the laboratory strain 70-15 had significantly more duplicated genes, including the AVR gene PWL2 [22] (Table S6). Other duplicated genes with known functions include LPS glycosyltransferases, MFS transporters, sugar transporters, and carboxypeptidases. Both intra- and inter-chromosomal duplications were observed, but more inter-chromosomal duplications were apparent, and only a small portion of duplication events were conserved in all three isolates (Figure 5A).
To identify gene families, the entire set of the predicted proteins from all three isolates were clustered with the OrthoMCL program. A total of 38,016 proteins were grouped into 14,189 clusters with each cluster representing a group of putative orthologs. Among these clusters, 195 gene families were identified with more than one member in at least one isolate (Figure 5B), suggesting that 1.37% of the M. oryzae genes may have been evolved by gene family expansion. Among 45 clustered loci duplicated equally in each isolate, 38, 6, and 1 gene loci were duplicated between two, three or four times, respectively, per isolate (Table S7). These gene families might have existed before the divergence of the three isolates. The majority of these gene families were predicted to be involved in synthesis and transport of nutrition and secondary metabolites, suggesting that they may be related to plant infection (Table S7). There were 87 clustered loci duplicated at different frequencies in three isolates (Table S8). Most of these gene families (61 out of 87) contained duplicated genes in only one isolate, and 17 gene families contained gene loci duplicated at least three times in one or more isolates (Table S8), suggesting that they have been expanded or contracted in different strains, possibly during environmental adaptations. For example, one putative calcium P-type ATPase gene was duplicated three times in P131 and Y34, and twice in 70-15. Members of this gene family have been demonstrated to be required for disease development and induction of host resistance [23], [24].
For loci duplicated in two isolates but absent in the third one, there were eight in P131 and Y34, five in P131 and 70-15, and twelve in Y34 and 70-15 (Figure 5B; Table S9). Most of these expanded gene families had unknown functions. To confirm the duplication events that were unique to the two field isolates, three genes were selected by Southern blot analysis. All of them were confirmed to be specifically duplicated in P131 and Y34 but not in 70-15 (Figure S7). There were seven, thirteen, and eighteen gene families specifically expanded in P131, Y34, and 70-15, respectively (Figure 5B; Table S10). Most of these isolate-specific gene families contained two or three duplicated members that had unknown functions or no known homologs in GenBank.
To analyze asynonymous and synonymous nucleotide substitutions, we first identified and removed orthologous genes with large deletions or insertions in any of the isolates from the list of common genes. In total, 9,184 highly conserved orthologs were used to identify nucleotide substitution events. Among them, 7,569 genes had neither synonymous nor asynonymous nucleotide substitution in pair-wise comparisons, indicating that most of the genes were well-conserved among different isolates. Only 428 genes had nucleotide substitutions between P131 and Y34, and 1,651 genes had nucleotide substitutions between 70-15 and P131 or Y34, further indicating that the field isolates had closer relationship with each other than with the laboratory strain. Genes with substitutions in the 70-15 versus P131/Y34 comparison could be categorized into four groups: 414 genes only with synonymous nucleotide substitutions, 697 genes only with asynonymous nucleotide substitutions, 124 genes with Ka/Ks<1, and 6 genes with Ka/Ks>1.
Overall, similar numbers of genes identical between Y34 and P131 but with nucleotide variations in 70-15 were thought to have undergone diversifying versus purifying selections. However, several functional categories of genes, such as those involved in cellular responses to stimuli and organophosphate metabolisms, had more members exhibiting diversifying selection in the two field isolates (Table S11). Several of the genes underwent diversifying selection in the 70-15 versus P131/Y34 comparison (Table S12), including ATG4, HEX1, MCK1, MoSNF1, PTH2, and RGS1, which are known virulence factors in M. oryzae [25]–[30]. Three of them encode putative CFEM-domain receptors that may be involved in recognizing different environmental and plant signals (Table S12).
Repetitive sequences were masked by Newbler for assembling 454 sequence data of P131 and Y34. To compare repetitive sequences of these two isolates, we assembled the Sanger reads of P131 and Y34 (approximately 2-fold genome coverage) and found that 10.8%, 10.3%, and 10.6% of the 70-15, P131, and Y34 genomes, respectively, were repetitive sequences, indicating that the abundance of repetitive sequences is similar among these three isolates. Transposable elements (TE) and their insertion sites (flanking sequences) were identified by RepeatMasker. Although the exact copy numbers vary, both field isolates contained all classes of transposable elements identified in 70-15 (Table 3). In general, 70-15 has more members of the LINE, Maggy, and RETRO5 LTR retrotransposons. The Pot2/Pot4 DNA transposons and the Pyret and Grasshopper LTR retrotransposons were more abundant in P131 and Y34. In addition, nine new clusters of repetitive sequences were identified by analysis with RepeatScout (Table 3). However, none of them was unique to the field isolates. While clusters 1, 4, 5, 6, and 7 were much more abundant in the field isolates, 70-15 had more copies of the cluster 2 repetitive elements (Table 3).
In comparison with 70-15, the two field isolates were more similar in the distribution pattern of repetitive sequences (Figure 1 and 6A). While chromosomal ends tend to have more repetitive sequences, all three isolates had much reduced numbers of TEs in the gene-rich regions of chromosomes III, V, and VI (Figure 6A). For the TEs that could be assembled into the genome sequences, approximately 27% of them had the same locations in all three isolates by comparison of their flanking sequences (Figure 6B). Y34 had more TEs with unique chromosomal positions (1,061) than P131 (830) or 70-15 (976). In addition to the 603 locations of TEs conserved among the three strains, Y34 and P131 also shared 281 TEs with the same chromosomal locations, which was fewer than the 377 between 70-15 and Y34 or the 341 between 70-15 and P131 (Figure 6B). While over two-thirds of the members of some TEs, including Occan, had conserved genomic locations, TEs such as Retro5 and Maggy differed significantly in their chromosomal positions between Y34 and 70-15. Similar results were obtained with the P131 and 70-15 comparison (Table 3). A total of 41.1% and 46.0% of TEs in 70-15 and P131, respectively, had conserved genomic locations. The Pot3, Maggy, Retro5, and Retro7 elements had the highest variation in chromosomal positions between 70-15 and P131.
We also analyzed the impact of TEs on the genome evolution by comparing two-fold coverage Sanger data of P131 and Y34 with the 70-15 assembly. A total of 35, 38, and 116 genes were disrupted by the insertion of TEs in P131, Y34, and 70-15, respectively (Table S13, S14, S15). Over 50% of the gene disruption events were caused by TEs belonging to MGL, Mg-SINE, or Pot2/Pot4. Strain 70-15 had a number of genes disrupted by cluster 7, cluster 9, Occan, and RETRO5 elements, which were not observed in P131 or Y34 (Figure 6C). Some of these genes may have been disrupted by transposition events occurring during generations of cultivation under laboratory conditions, and these genes may play roles in plant infection or survival in the field isolates but were not required for the laboratory isolate. In comparison with 70-15, the field isolates P131 and Y34 had more genes disrupted by SINE (Figure 6C), which may indicate that these SINE elements were more active in these two field isolates.
Among all the genes disrupted by TEs in three isolates, only approximately one third of them have known functions based on their orthologs in GenBank, and most of them are involved in protein metabolism, transportation, transcription, or lipid metabolism. The majority of the TE-disrupted genes encode hypothetical proteins with unknown functions. Interestingly, 23.8% of them contained putative signal peptide sequences, which is significantly higher than the average percentage of predicted extracellular proteins in the genomes of these three strains (Table S13, S14, S15). Some of them may function as effectors involved in fungal-plant interactions, such as AVR Pi-ta1 in 70-15 (Table S13). In addition, 14.7%, 14.2% and 15.8% of the TE-disrupted genes in 70-15, P131, and Y34, respectively, encoded proteins with putative nuclear localization sequences.
Intriguingly, the regions containing isolate-specific sequences or duplicated genes families were often near areas with high frequency of TEs (Figure S8). In 70-15, several TEs were found within 1.0 kb from 23 duplicated genes families, including the avirulence gene PWL2 (Table S16) although many of these duplicated sequences were not closely linked or located on different chromosomes. Taken together, it is likely that the transposition events of TEs might be related to translocation of duplicated DNA fragments and presence of isolate-unique sequences in these three strains.
In a number of eukaryotic organisms, comparative analysis of multiple genomes of the same species has been used to improve assembly and annotation and to identify genome variations [31]–[34]. The rice blast fungus is well-known for its natural genetic variation [1], [2]. In this study, we sequenced two field isolates of M. oryzae from Asia. Genome analysis indicated that these two field isolates are more closely related to each other than to 70-15, which is a laboratory strain derived from three backcrosses of rice pathogen Guy11 with a progeny of a cross involving a weeping love grass pathogen, and maintained for many years under laboratory conditions. The overall genome content and composition are similar among these three isolates, but the genomes of P131 and Y34 with only A/C/T/G and no N's were slightly larger than that of 70-15.
Although the 70-15 genome has been updated several times, it still has many gaps (www.broadinstitute.org/annotation/genome/magnaporthe_grisea). In this study, a total of 55 gaps of the 70-15 genome (version 6) were filled in with sequences from P131 and Y34, and the results were validated by PCR analyses of 70-15. This number of putative filled gaps with sequences from two isolates may seem low, but because of the short read length, the threshold set may have been too stringent. For 35 gaps, they were filled with consensus sequences found in both field isolates. For the gaps with sequences only available in either Y34 or P131, the filling sequence for 70-15 was less certain, but of high probability because the overall nucleotide sequence identity between 70-15 with P131 or Y34 was over 98%. Besides improving the genome assembly, the sequences of P131 and Y34 were used to improve the annotation of 70-15. We identified 81 genes that were not predicted in the automated annotation of the 70-15 genome sequence, and none of them were related to the sequence gaps. In addition, we identified potential annotation errors in 340 predicted genes of 70-15. Most of them were related to the problems with the prediction of the boundaries of introns and start or stop codons.
Our study revealed that each M. oryzae isolate had some unique genomic DNA sequences. Because genome sequences of P131 and Y34 were aligned with that of 70-15, it was impossible to locate most of the sequences unique to Y34 and P131 onto specific chromosomes or chromosomal regions. However, sequences unique to 70-15 were distributed over all seven chromosomes. Because 70-15 was derived from three backcrosses of rice pathogen Guy11 with a progeny of a cross involving a weeping love grass pathogen, we expected that a small portion of its genome was from the weeping love grass pathogen. The isolates Y34, P131, and 70-15 had 136, 51, and 71 unique genes, respectively. Therefore, less than 1% of the predicted genes were unique to each isolate and these genes play diverse roles, some of which might possibly contribute to the specificity of individual isolates. Some of the isolate-specific genes were clustered, suggesting that isolate-specific DNA fragments might be gained or lost during evolution. The P131-specific gene P131_scaffold00208-2 encoded a hypothetical protein without known homologs in other fungi. Deletion of this gene resulted in reduced virulence toward rice plants. Because it might be involved in plant infection, P131_scaffold00208-2 may play an isolate-specific role in suppressing or overcoming plant defense responses. These results suggest that some of the field isolate-specific genes may play important roles in plant infection.
In all three M. oryzae isolates, most of the duplicated genes are functionally unknown. Duplicated sequences are distributed all over seven chromosomes and appear to be enriched in the telomeric regions. For the duplicated genes with known functions, many of them are predicted to be involved in primary and secondary metabolism and interactions with the host (such as cutinases and Avr proteins), which is consistent with earlier observations with 70-15 [3]. Interestingly, several gene families involved in synthesis and transport of nutrients and secondary metabolites were expanded with different frequencies in these three isolates. Some of these duplicated genes may contribute to the adaption of M. oryzae to different environmental conditions.
Among the genes that had undergone diversifying selection in Y34 and P131 in comparison with 70-15, a number of them are known to be important for virulence, suggesting that such genes may have been under strong selection pressure in their natural field environments. There were six genes under positive selection in the two field isolates compared to 70-15. Two of them encoded two hypothetical proteins, a serine/threonine protein kinase, an acyltransferase, a putative catalytic domain of diacylglycerol kinase, and an aspartic-type endopeptidase. Three of them are located on chromosome I. In contrast, there were no genes showing positive selection in the comparison between the field isolates.
Because sexual reproduction has not been observed in the field, it is possible that translocations of the repetitive sequences may be one of the major sources for genome variation and rapid adaption to different host and environmental conditions. Consistent with this hypothesis, over 10% of the genome sequences were found to be repetitive sequences. In addition to TEs that have been identified in previous studies [3], nine new clusters of repetitive sequences were identified in all three M. oryzae strains in this study. Most of these TEs have different copy numbers in different isolates. Strikingly, among thousands of TE loci, less than 30% of them were conserved among these isolates, suggesting active transposition of these TEs in M. oryzae. Moreover, approximately 200 genes were totally disrupted by TEs in these three strains, and approximately 40% of them encoded extracellular or nuclear proteins, suggesting that transpositions of TEs may contribute to variations in host-microbe interactions and transcriptional regulation. Interestingly, TEs tended to be found near isolate-specific sequences and duplicated DNA fragments. It is possible that translocation of TEs is important for gain or loss of isolate-specific sequences and gene duplication events.
Overall, our results indicate that gain or loss of unique genes, duplications, gene family expansions, and translocations of TEs can be important factors for genome variation in the rice blast fungus. Among these factors, translocation of TEs may be the most important one because of its association with gene duplication and isolate-specific sequences. There are reports on comparative genomic analyses of plant pathogenic oomycetes and fungi, such as Phytophthora and Fusarium species [35], [36]. However, to our knowledge, this study is the first on comparative analysis of the field and laboratory strains of a plant pathogenic fungus, and this can give insights into the genome variations of the fungus under different environments.
For Sanger sequencing, genomic libraries with insertion size of 1.5 kb to 3.5 kb were constructed and sequenced at the Beijing Genomic Institute (BGI, Beijing, China). These two isolates were also sequenced with the GS-FLX and GS-FLX Titanium 454 platforms [37] at BGI that generated reads with an average length of 240- and 380-bp, respectively. Reads from Sanger and 454 sequencing were placed into scaffolds using the Newbler assembler (version 1.1.02.15, Roche).
The M. oryzae 70-15 genome sequence version 6 was downloaded from the Broad Institute (www.broad.mit.edu/annotation/genome/magnaporthe_grisea). The repetitive sequences in the assembled genomes of laboratory strain 70-15 and the field isolates P131 and Y34 were masked with RepeatMasker (Smit, AFA, Hubley, R & Green, P. RepeatMasker Open-3.0 at http://repeatmasker.org). Masked genome sequences of the three M. oryzae isolates were compared with the MUMMER package [38] to construct chromosome sequences for P131 and Y34 based on 70-15 data. Genomic sequences with nucleotide identity over 92% were considered to be conserved among different isolates.
De novo gene prediction of the P131 and Y34 genome sequences was performed with FGENESH [39], which was trained with 79 gene models of M. oryzae (kindly provided by Prof. Zhen Su at China Agricultural University). The tRNA genes were identified by tRNAscan [40]. Gene functions were predicted by comparison with the NCBI NR protein database (http://www.ncbi.nlm.nih.gov/) and the Pfam database [41]. InterPro [42] was used for gene ontology annotations. Membrane and sub-cellular localization were predicted by TMHMM 2.0 [43], SignalP3.0 [44], and WoLF PSORT [45].
Nucleotide sequences of the predicted genes of P131, Y34, or 70-15 were compared separately with genomic sequences of the other two isolates with TBLASTN [46]. Homologous genes with sequence identities of 100%, 80–100%, and 50–80% were defined as identical, similar, and divergent, respectively, while those below 50% were considered non-homologous. Sequences of genes unique to the field isolates were also queried against the unassembled reads of 70-15. Orthologous proteins were clustered with OrthoMCL [47]. Only the clusters containing one protein from each isolate were selected for distance analysis. Individual protein sequences from three isolates were concatenated and aligned with T-Coffee [48], and a distance matrix was calculated with PROTDIST from the PHYLIP package [49]. Finally, a neighbor-joining tree was constructed with NEIGHBOR from the PHYLIP package.
The coding sequences of orthologous genes conserved in all three isolates were aligned with ClustalW [50] to detect large deletions (>12-bp), frame shifts, and null mutations. Orthologous genes without large deletions, frame shifts, or null mutations in the open read frame were analyzed for Ks and Ka with the YN00 program in the PAML package [51].
The Sanger reads of P131 and Y34 were assembled with RePS [52] and analyzed for transposable elements with RepeatMasker. New repetitive elements were identified by RepeatScout [53]. For each transposable element (TE) identified in P131 or Y34, its flanking sequences of 30 to 100 bp were extracted and used to search against the 70-15 genome with Standalone BLASTN (e-value<10−5). Each TE and its corresponding region in 70-15 genome were aligned with BLAST2seq to assess whether it was conserved. To search for genes disrupted by TEs, unique flanking sequences of TEs in P131 or Y34 were used to search against 70-15 genes (e-value<10−20). The search results were removed if more than one hit was found. Similar analyses were performed with P131 and Y34.
The wild-type and mutant strains of 70-15, P131, and Y34 were cultured at 25°C on oatmeal tomato agar (OTA) plates and conidiation assessed [17]. Mycelia collected from two-day-old cultures in complete media (CM) shaken at 150 rpm were used for extraction of fungal DNA and protoplasts. Media were supplemented with 250 µg/ml hygromycin B (Roche, USA) or 400 µg/ml neomycin (Amresco, USA) to select hygromycin-resistant or neomycin-resistant transformants. Four-week-old seedlings of monogenic rice cultivars (Table S1) and eight-day-old seedlings of barley cultivar ‘E8’ were inoculated as previously described [17], [54]. Lesion development was examined 5–7 days after inoculation.
Chromosome-size DNA were prepared with protoplasts isolated from vegetative hyphae as previously described [55], [56], and separated on 0.65% Megabase agarose (Bio-Rad, USA) gels with a Bio-Rad DR III system with switching intervals of 60 min for 48 h, 55 min for 72 h, 45 min for 72 h, and 35 min for 72 h at 1.5 V/cm. Chromosomal DNA of Schizosaccharomyces pombe and Hansenula Wingei (Bio-Rad, USA) were used as the molecular weight markers.
To generate the P131_scaffold00208-2 gene replacement vector pKOPS208-2, its 0.97 kb upstream and 0.82 kb downstream fragments were amplified with primer pairs P131_scaffold00208-2KO_LBf plus P131_scaffold00208-2KO_LBr, and P131_scaffold00208-2KO_RBf plus P131_scaffold00208-2KO_RBr, respectively. The resulting PCR products were cloned into the KpnI-HindIII and EcoRI-SpeI sites of pKOV21 [56], [57]. After linearization with NotI, pKOPS208-2 was introduced into protoplasts of P131. Hygromycin resistant transformants were isolated and assayed for neomycin-resistance. The resulting transformants were screened by primer pairs P1/P11 and P2/P12. The putative deletion mutants were identified and confirmed by Southern blot analysis. The same approach was used to generate gene replacement constructs and mutants for isolate-specific genes: P131_scaffold00297-2, P131_scaffold00493-1, Y34_scaffold00875-1, Y34_scaffold00875-3, Y34_scaffold00857-6, Y34_scaffold01193-2, Y34_scaffold00005-1, Y34_scaffold01048-2, Y34_scaffold00105-1, Y34_scaffold00105-2, and Y34_scaffold00855-11. The primer pairs used for generating the gene replacement constructs and for mutant screening are listed in Table S17.
The genome sequence data of Y34 and P131 were deposited in the NCBI Genome Database (www.ncbi.nlm.nih.gov/genome) under the accession numbers AHZS00000000 and AHZT00000000, respectively. The nucleotide sequence data of repetitive sequences and transposable elements are available in the NCBI GenBank database under the following accession numbers: M77661 for Grasshopper, AB024423 for Maggy, AF018033 for MGL, AJ851229 for Mg-MINE, AF314096 for MGRL3, MGU35313 for Mg-SINE, AB074754 for Occan, AF314096 for Pot2, AF333034 for Pot3, AB062507 for Pyret, NC_009594 for Pot4, RETRO5, RETRO6, RETRO7, and Ch-SINE, JQ929664 for Cluster 1, JQ929665 for Cluster 2, JQ929666 for Cluster 3, JQ929667 for Cluster 4, JQ929668 for Cluster 5, JQ929669 for Cluster 6, JQ929670 for Cluster 7, JQ929671 for Cluster 8, and JQ929672 for Cluster 9.
|
10.1371/journal.ppat.1001160 | Large-Scale Field Application of RNAi Technology Reducing Israeli Acute Paralysis Virus Disease in Honey Bees (Apis mellifera, Hymenoptera: Apidae) | The importance of honey bees to the world economy far surpasses their contribution in terms of honey production; they are responsible for up to 30% of the world's food production through pollination of crops. Since fall 2006, honey bees in the U.S. have faced a serious population decline, due in part to a phenomenon called Colony Collapse Disorder (CCD), which is a disease syndrome that is likely caused by several factors. Data from an initial study in which investigators compared pathogens in honey bees affected by CCD suggested a putative role for Israeli Acute Paralysis Virus, IAPV. This is a single stranded RNA virus with no DNA stage placed taxonomically within the family Dicistroviridae. Although subsequent studies have failed to find IAPV in all CCD diagnosed colonies, IAPV has been shown to cause honey bee mortality. RNA interference technology (RNAi) has been used successfully to silence endogenous insect (including honey bee) genes both by injection and feeding. Moreover, RNAi was shown to prevent bees from succumbing to infection from IAPV under laboratory conditions. In the current study IAPV specific homologous dsRNA was used in the field, under natural beekeeping conditions in order to prevent mortality and improve the overall health of bees infected with IAPV. This controlled study included a total of 160 honey bee hives in two discrete climates, seasons and geographical locations (Florida and Pennsylvania). To our knowledge, this is the first successful large-scale real world use of RNAi for disease control.
| High rates of honey bee mortality continue to threaten food security and apicultural industries worldwide. At least some of these losses are likely the result of viral infections. Application of RNAi technologies in the treatment and management of disease promises new solutions to disease problems through the naturally occurring biological processes of living organisms. This study applied a novel dsRNA product developed specifically with the aim of improving honey bee health. The results demonstrate the successful application of RNAi strategies to improve disease tolerance. Honey bees were fed a dsRNA product, Remebee-I, in the presence of the Israeli Acute Paralysis Virus. Treatment resulted in larger colony populations and thus increased honey production. We show that IAPV specific homologous dsRNA successfully curbed the negative effects of IAPV infection in 160 honey bee hives in two discrete climates, seasons and geographical locations (Florida and Pennsylvania). We provide the first successful demonstration of the use of RNAi as a preventative treatment for an insect disease on such a large scale.
| The importance of honey bees as pollinators of crops to the global economy far surpasses their contributions in terms of honey production [1]. In all, 52 of the world's 115 leading agricultural crops rely on honey bee pollination to some extent. These crops represent approximately 35% of the human diet [2]. Insect pollination, which is provided predominately by honey bees, is estimated to have a value of US$ 212 billion [3]. Honey bee populations have been decreasing globally in recent years [4]. Since fall 2006, honey bees overwintering in the U.S.A. have faced unusually high rates of mortality, in part because of a phenomenon now known as Colony Collapse Disorder (CCD) [5]. Several hypotheses have been offered to explain CCD and existing and emerging pathogens have been implicated either directly or indirectly [6]. Colonies affected by CCD are infected with larger numbers of pathogenic organisms than control colonies, yet no single pathogen was found associated with all affected colonies [7]. In another effort, researchers did find that single-stranded RNA viruses, specifically picorna-like viruses, occurred at elevated levels in CCD colonies. These elevated levels of viruses may interfere with gene transcription, thus reducing immune response competence and pesticide detoxification capabilities, subsequently leading to premature death of infected bees [8].
Honey bees are susceptible to a host of picorna-like viruses, including the closely related Acute Bee Paralysis Virus (ABPV), Kashmir Bee Virus (KBV), and Israeli Acute Paralysis Virus (IAPV) [9], [10]. The latter of these three viruses was identified as a good marker for CCD in initial studies, especially when found in association with the microsporidia Nosema sp. [6]. While IAPV is probably not the sole cause of CCD [7], its ability to cause increased mortality in honey bees has been established [11].
The process of post-transcriptional gene silencing is thought to be an evolutionarily-conserved cellular defense mechanism used to prevent the expression of foreign genes and is commonly shared by diverse flora and phyla [12]. The presence of long double-stranded RNAs in cells stimulates the activity of a ribonuclease III, Dicer, which is involved in the processing of the double stranded RNA (dsRNA) into short interfering RNAs (siRNAs). The RNAi response also features an endonuclease complex, commonly referred to as an RNA-induced silencing complex (RISC), which mediates cleavage of target ssRNA having sequence complementary to the antisense strand of the siRNA duplex.[13], [14], [15], [16].
In a variety of organisms, exogenously applied dsRNA or their siRNA derivatives, can be used to arrest, retard or even prevent a variety of pathogens. In some of these organisms, such as plants and the nematode C. elegans, an amplification stage follows the initiation stage of gene silencing, involving an RNA dependent RNA Polymerase (RdRp), which may lead subsequently to degradation of RNAs outside the initial dsRNA region of homology [17]. RNAi can spread from the initial site of dsRNA delivery, producing interference phenotypes throughout the treated animal. To serve as a preventive or curative strategy, amplification and systemic spread of the silencing signal are both paramount. In some invertebrates, including honey bees, a systemic interference defective (SID) gene encodes a transmembrane protein that is an important participator in the systemic RNAi pathway. Apparently, these SID1-like proteins channel dsRNAs between cells, enabling systemic spread of the silencing signal [18], [19]. Although a canonical invertebrate RNA dependent RNA Polymerase (RdRP) homologue has not yet been described, there is evidence that such RdRp activity may occur via other enzymes, leading to amplification of the silencing signal in insects [20].
IAPV specific dsRNA (Remebee-IAPV or herein Remebee-I) was used successfully to prevent bees from succumbing to infection from IAPV in small scale lab experiments whereas bees fed Green Fluorescent Protein (GFP) dsRNA and virus died in a manner similar to the IAPV fed control bees [12]. Although these results were exciting per-se, transferring RNAi from a well characterized and efficient tool in the lab and making it successful in preventing the adverse effects of virus infection in the field, remains notoriously difficult.
We present the first large-scale real world successful use of RNAi for disease control. We attempted to determine if IAPV specific homologous dsRNA can be used to reduce impacts from IAPV infection in 160 honey bee hives in two discrete climates, seasons and geographical locations (Florida and Pennsylvania). To our knowledge, this is the first successful demonstration of the use of RNAi as a preventative treatment for an insect disease on such a large scale.
The field demonstration in FL was designed in a manner that permitted us to follow IAPV-infested bee colonies (some given Remebee-I and others not) for six weeks. One hundred standard colonies of honey bees were split into 5 groups with 20 colonies per group. Four groups were located within 100 m of one another (non-isolated) while a 5th group was isolated from the remaining four by at least 3.2 km to measure any environmental effects due to location.
Treatment allocations (20 colonies per treatment) were as follows:
Treatment 1 – no treatment – non isolated
Treatment 2 – Remebee-I only – non isolated
Treatment 3 – Remebee-I+IAPV – non isolated
Treatment 4 – IAPV only (fed in sugar water solution) – non isolated
Treatment 5 – no treatment– isolated
Honey (or net weight gain) is the ultimate proxy to the total active population of the hive. The non treated control produced the most honey in PA, but not in FL. In FL, colonies treated with Remebee-I+IAPV produced significantly more honey than colonies receiving IAPV alone (Figure 2, N = 40, p<0.03). In PA, the difference between the weight at the start and the end of the experiment (4 months) shows that the non infected controls gained the most weight (mean gain = 23.5kg), whereas Remebee-I+IAPV had gained slightly less (mean = 21kg). Both made significantly greater weight gains compared with the group receiving IAPV alone (mean = 16.3kg) (Figure 3 F = 2.7; df = 4.92; P = 0.034).
Subsequent trials done under a similar protocol were repeated in the winter of 2009–10 in FL and in California (CA). Samples of bees were collected just before IAPV challenge and 2-weeks post treatment. Northern analysis was done with IAPV specific sequence probes corresponding with the Remebee-I sequence. The results of these are presented in detail in Supporting Information S2. High levels of discrete Dicer Remebee metabolites are evident in Remebee-I treated hives prior to IAPV challenge up to four weeks after a Remebee-I application. Non- Remebee-I treated bees are mostly negative, but a low signal was detected in some colonies. Subsequent to IAPV challenge, levels of siRNAs and IAPV metabolites are highly elevated in both Remebee-I and in non-Remebee-I treated hives, showing that production of dsRNA is a natural defense mechanism in bees against IAPV infection.
Varroa levels are unaffected by treatment. The change in varroa prevalence on adult bees did not differ significantly between treatment groups in either the FL (F4,95 = 2.39; P = 0.056) or PA (F2,57 = 1.03; P = 0.3642) trials.
Nosema levels: While the change in nosema levels in the FL trial was not significantly different between treatment groups (F4,95 = 0.47; P = 0.7586), a highly significant difference in nosema spore levels was detected at the end of the PA trial (F2,57 = 8.62; P = 0.0005) (Figure 4). Indeed, within the duration of the experiment, nosema spores levels increased in the IAPV treated group, yet went down in both Remebee-I+ IAPV and uninfected control colonies.
A brief summary of findings is presented in Table 1.
The negative effects of IAPV on honey bee health and colony vigor is evidenced by lower honey weight gains (Figures 2 and 3). Environmental factors and terrain also influence the availability of forage and thus can seriously reduce or increase the effects of virus infection on a hive by altering foraging patterns. Large scale examination of the RNAi treatment under such varied conditions at two separate and diverse locations (i.e. FL and PA) was challenging. However, one would expect that 7–8 weeks after a young queen begins to oviposition prolifically (see Supporting Information S1 for brief overview), all hives would be overflowing with bees. This was evidently not the case in this situation, and the IAPV only group reversed in total population and bee/brood ratio. Thus, we observed relative de-population of the hive with probable greater loss of foragers in IAPV only infected hives. We conducted the trials described herein in spring and summer, whereas CCD is a mostly winter phenomenon [1], [7]. This may be the reason that we did not see many hives devastated by CCD. However, subsequent trials performed in the 2009–10 winter in Florida and California resulted in 40% and 60% collapse of hives, respectively (Hunter Wayne. and Oliver Randy., personnel communication). Beyond cold weather providing additional stress on the bees, some of the difference may be attributed to the viruses' suppressors of gene silencing. In plants, these viral suppressors of gene silencing have more optimal enzymatic kinetic coefficients under cold temperatures in relation to the silencing enzymes, often leading to more acute virulence [21]. This could help explain the initiation and overall devastation of hives in the U.S. following winter cold snaps (Dennis VanEngelsdorp unpublished observations).
Our hypothesis that Remebee-I would protect bees from IAPV infection was supported by multiple observations: First, in FL, the Remebee-I+ IAPV treated hives were the only colonies with significantly increasing numbers of bees during the study. In PA bees increased within all treatments with no significant differences. Some of the difference between the two trials could account for this difference in observations. In PA the virus was introduced twice into the colonies within a three day period (instead of once), and the total amount of virus introduced was thus much higher than the FL trial. The PA trial starter hives were weaker in strength, and had more collapses of hives across all treatments prior to infection than those starting the FL trial. Second, although bee to brood ratios started lower in Remebee-I treated hives it became stronger and remained so until the end of the trial. The change in the ratio is attributed to changes in the adult bee counts, since the capped brood coverage was the same between treated and non-treated bees, thus these hives also contained more adult foragers, which resulted in significantly more honey production over the IAPV only treatment (Figure 2,3).
Furthermore, in subsequent trials, molecular evidence now proves that Remebee-I is active in the Remebee-I + IAPV treated groups, as determined by the presence of siRNA (see Supporting Information S2). The strong presence of siRNAs probably restricts the severity of the disease in the bees leading to a longer life-span and subsequently to an overall greater number of bees, with more foragers and consequently a greater yield of honey. It is interesting to note the natural occurrence of these siRNAs in bees receiving IAPV challenge. Presence of these siRNA in non- Remebee-I treated hives prior to infection may be a result of natural virus infection prior to the challenge, or by transcription of integrated viral sequences in the bee genome [22].
IAPV specific dsRNA (Remebee-I) was used successfully to prevent bees from succumbing to infection from IAPV. The results further demonstrate the possibility to produce targeted treatments for bee pathogenic diseases. These field results demonstrate the successful application of dsRNA as a viable treatment to solve a real world problem, which may further lead to concerted efforts to utilize this ubiquitous natural mechanism, RNAi, for the benefit of the bees, beekeepers, and hopefully to other applications in agriculture and veterinary health.
To determine if IAPV can be silenced using RNAi technology, we had to (1) purify IAPV from honey bees, (2) infect honey bee colonies with IAPV and/or Remebee-I and (3) determine IAPV presence in experimental colonies.
Essentially as described in [11].
Approximately 40 adult forager honey bees were collected from 10 colonies in a Florida bee yard (apiary) where CCD had been reported. Each bee was processed individually and tested using rtPCR for the presence of Israeli Acute Paralysis Virus (IAPV), genome – NC_009025; Acute Bee Paralysis Virus (ABPV) genome – NC_002548; Kashmir Bee Virus (KBV) genome – NC_004807; Black Queen Cell Virus (BQCV), genome-NC_003784; Deformed Wing Virus (DWV) genome NC_004830. All bees had more than one virus detected so inoculum was prepared from bees which tested positive only for IAPV+KBV by homogenizing the bees with glass beads in small amounts of 10 mM buffer phosphate, pH 7.2 containing 0.02% DETCA (Sigma-Aldrich Cat #22,868-0). Inoculum was prepared by passing the virus solution through a syringe filter, 0.45 µm, to remove bacteria, after which ∼10 µl were administered by microinjection along the lateral side of the abdomen of ∼700 pupae using a Hamilton syringe with a 30Gx½ gauge sterile needle. Inoculated pupae were kept in petri dishes covered with slightly damp filter paper and maintained at 21–23°C for three days to permit virus replication.
On the third day, batches of about 50 pupae were homogenized with glass beads. Small amounts of 10 mM buffer phosphate (pH 7.2 contained 0.02% DETCA, Sigma-Aldrich Cat #22,868-0) were added to the homogenates. The homogenates were collected in a beaker volume adjusted to ∼350 ml with buffer (see above) and mixed. Each sample was split into two 250 ml centrifuge tubes and centrifuged at 300×g (∼1,400 rpm) on a GSA rotor for 20 min. The supernatant (S1) was collected and kept at 4°C for 3 d. The pellet (P1) was recovered and saved at 4°C. Since some precipitation was noticed after 3 d, the supernatant was centrifuged again as before for 10 min to remove debris. Supernatant (S1) next was transferred to 12 ultracentrifuge tubes (about 26 ml/tube) (Beckman Cat #355618) and centrifuged for 4 h, 4°C, at 37,000 rpm (∼124,500×g) (Beckman Type 50.2 Ti rotor, Beckman Optima L-70K Ultracentrifuge). After 4 h the supernatant (S2) was removed and saved. The pellet (P2) was resuspended in 10 mM phosphate buffer containing 4% Brij 58 (Aldrich Cat #388831) and 0.4% Sodium deoxycholate (Sigma-Aldrich D6750): about 1 ml of buffer was used per tube. It was necessary to insert a spatula to help pellet into solution and this was followed by vortexing the suspension. The content from each tube was transferred to clean 50 ml centrifuge tubes. The process was repeated twice but only buffer phosphate was added the final time. Because the final solution was very thick, buffer was added to increase the final volume to ∼30 ml and this was mixed by inversion. The tube was centrifuged for 15 min at ∼10°C, 800×g (Beckman Coulter Allegra 25R), to remove debris. The pellet (P3) was saved at 4°C. (the pellet saved as a backup). The supernatant (S2) was transferred into two clean 50 ml tubes and 13.2 g CsCl (Amresco Cat #0415) were added to each tube. To ensure the right CsCl concentration, 13.2 g CsCl were added to ∼10 g sample; however the final volume was adjusted to 24 ml with buffer and gently mixed (up/down). The second tube was set by adding CsCl to the remaining sample. This final preparation was transferred to two ultracentrifuge tubes ∼25 ml (Beckman Cat #355618) and centrifuged at 37,000 rpm (∼124,500×g), 18°C for 24 h. After 24 h centrifugation, the tubes were removed carefully from the rotor and the whitish virus band collected by insertion of a needle attached to a syringe. Two more fractions were recovered for analyses: (1) the “liquid” part left after removing the virus band and (2) the “pellet” (P4) attached to the bottom of tube: Each fraction was transferred to dialysis tubes (Thomas Scientific Cat #3787-F42) and dialyzed overnight against nanopure filtered water followed by 3–4 additional changes in water the following day. After dialysis, content from the tubes was collected in 15 ml clean tubes and the volumes were measured. A subsample of 20 µl from each fraction was tested for virus presence.
Adult bees were transferred to 1.5 ml centrifuge tubes. Tri Reagent (Sigma Cat #T9424), was added and individual bees were homogenized in 0.5 ml Tri reagent using disposable pestles and glass beads. Homogenates were frozen at −20°C if needed. Samples then were centrifuged 10 min at 12,000×g, at 4°C. The clear supernatant was transferred to a new tube and left at least 5 min at room temperature (RT). Next, 0.2 ml chloroform was added and samples were shaken vigorously. This was followed by a10–15 min incubation at RT. Tubes were centrifuged 15 min at 12,000×g at 4°C. The colorless upper aqueous phase was transferred to a new tube and 0.5 ml isopropanol was added. After mixing, samples were allowed to stand for 10 min then spun 10 min at 12,000×g at 4°C. The supernatant was removed and the pellet containing the RNA was washed with 1 ml 75% ethanol. After 5 min centrifugation at 7,500×g at 4°C, the RNA was allowed to dry (5–10 min) and reconstitute in ∼30 µl Nuclease free water (Qiagen). RNA concentrations were measured in a Nanodrop, ND-1000 Spectrophotometer. Samples were diluted in Nuclease free water.
The field demonstration in FL was designed in a manner that permitted us to follow IAPV-infested bee colonies (some given Remebee-I and others not) for six weeks. One hundred standard colonies of honey bees were split into 5 groups with 20 colonies per group. Four groups were located within 100 m of one another (non-isolated) while a 5th group was isolated from the remaining four by at least 3.2 km to measure any environmental effects due to location.
Treatment allocations (20 colonies per treatment) were as follows:
Treatment 1 – no treatment – non isolated
Treatment 2 – Remebee-I only – non isolated
Treatment 3 – Remebee-I+IAPV – non isolated
Treatment 4 – IAPV only (fed in sugar water solution) – non isolated
Treatment 5 – no treatment– isolated
Colonies were equalized according to standard protocols prior to the beginning of the study (frames of bees/brood moved between colonies until populations leveled) and were managed optimally for honey production. Data collected at the beginning, middle, and end of the study included: frames of adult bees, cm2 brood, the presence of other bee maladies (nosema, varroa, and tracheal mites), bee activity, honey production and IAPV presence/absence and titer. The study lasted 6 weeks from the date of colony inoculation with IAPV and was replicated in PA with the following modifications: Only Treatment groups 1, 3 and 4 were established and the trial lasted 12 weeks after inoculation to enable the bees to take advantage of a honeyflow (Tables 2 and 3). In PA the virus was introduced twice into the colonies within a three day period (instead of once as in FL), and the total amount of virus introduced was thus much higher than the FL trial. We calculated all test/sampling dates below from the date of the last treatment with Remebee-I. Controls 1, 2, and 5, accounted for ‘within treatments’, a Remebee-I alone treatment to evaluate any potential detrimental effects to bees, and a distant control to measure environmental effects in the absence of IAPV.
Colonies were equalized at the beginning of the studies, starting and ending colony strength parameters were compared using ANOVA recognizing treatment as the main effect (PROC GLM). Honey gains in treated colonies in FL and PA were compared identically. However, to compare colony size measures in the PA trial and for levels of nosema and varroa mites, a Before-After Control-Impact (BACI) design [33], [34] was used. A BACI design us a way of comparing data that are measured before treatment with data obtained after treatment. In general, it can be described as a repeated measures analysis of variance (ANOVA) which is performed using colonies as replicates and the covariance structure that best suits the data (PROC MIXED, SAS Institute). Each variable is measured at the start of the experiment to show existing conditions before treatment and then after a treatment. The analysis then looks at whether the change in variable measures was different between treatment groups. A repeated measures analysis of variance [34] was performed using colonies as replicates and an unstructured covariance structure was performed using SAS statistical software (PROC MIXED) [35].
Israeli Acute Paralysis Virus (IAPV) genome – NC_009025 (RefSec); Acute Bee Paralysis Virus (ABPV) genome – NC_002548 (RefSec); Kashmir Bee Virus (KBV) genome – NC_004807 (RefSec); Black Queen Cell Virus (BQCV), genome-NC_003784 (RefSec); Deformed Wing Virus (DWV) genome NC_004830 (RefSec); RNA dependent RNA Polymerase protein (C. elegans)– NP_492131 (RefSec); SID-1 protein (A. mellifera) – XP_395167 (RefSec); GFP nucleotide sequence – U87625 (GenBank).
|
10.1371/journal.pntd.0003320 | Antibody-Dependent Enhancement Infection Facilitates Dengue Virus-Regulated Signaling of IL-10 Production in Monocytes | Interleukin (IL)-10 levels are increased in dengue virus (DENV)-infected patients with severe disorders. A hypothetical intrinsic pathway has been proposed for the IL-10 response during antibody-dependent enhancement (ADE) of DENV infection; however, the mechanisms of IL-10 regulation remain unclear.
We found that DENV infection and/or attachment was sufficient to induce increased expression of IL-10 and its downstream regulator suppressor of cytokine signaling 3 in human monocytic THP-1 cells and human peripheral blood monocytes. IL-10 production was controlled by activation of cyclic adenosine monophosphate response element-binding (CREB), primarily through protein kinase A (PKA)- and phosphoinositide 3-kinase (PI3K)/PKB-regulated pathways, with PKA activation acting upstream of PI3K/PKB. DENV infection also caused glycogen synthase kinase (GSK)-3β inactivation in a PKA/PI3K/PKB-regulated manner, and inhibition of GSK-3β significantly increased DENV-induced IL-10 production following CREB activation. Pharmacological inhibition of spleen tyrosine kinase (Syk) activity significantly decreased DENV-induced IL-10 production, whereas silencing Syk-associated C-type lectin domain family 5 member A caused a partial inhibition. ADE of DENV infection greatly increased IL-10 expression by enhancing Syk-regulated PI3K/PKB/GSK-3β/CREB signaling. We also found that viral load, but not serotype, affected the IL-10 response. Finally, modulation of IL-10 expression could affect DENV replication.
These results demonstrate that, in monocytes, IL-10 production is regulated by ADE through both an extrinsic and an intrinsic pathway, all involving a Syk-regulated PI3K/PKB/GSK-3β/CREB pathway, and both of which impact viral replication.
| IL-10 has multiple cellular functions, including anti-inflammatory and immunomodulatory effects. Clinical studies have demonstrated that the serum levels of IL-10 are significantly increased in DENV-infected patients with severe disorders. However, the molecular mechanism underlying DENV-induced IL-10 production is still unresolved. In this study, we demonstrate a molecular mechanism for DENV-induced IL-10 production, which may be exacerbated by ADE through Fcγ receptor-mediated extrinsic and intrinsic pathways, leading to IL-10/SOCS3-mediated advantages for viral replication. With or without Fcγ receptor- or CLEC5A-mediated DENV infection, a common Syk/PKA-regulated PI3K/PKB activation results in a decrease in GSK-3β activity followed by an increase in CREB-mediated IL-10 expression not only in THP-1 monocytic cells but also in human monocytes. Taken together, we demonstrate a potential regulation and a pathological role for ADE-induced IL-10 overproduction during DENV replication. Therefore, inhibiting immunosuppression by targeting the IL-10 pathways identified in this study may help to prevent the progression of severe dengue diseases.
| Four serotypes of dengue virus (DENV) – a mosquito-borne human pathogen belonging to the family Flaviviridae and the genus Flavivirus – infect an estimated 50 million people annually and cause a spectrum of illnesses, ranging from mild dengue fever (DF) to the more severe dengue hemorrhagic fever (DHF) and dengue shock syndrome (DSS) [1]. However, it is unclear which antiviral strategies are most appropriate for treating DENV progression, as many aspects of DENV pathogenesis remain controversial, including viral load, virulence, cytotoxicity, the nature of the immune response, autoimmunity [2], [3], and the potential effects of common diseases such as allergies, diabetes, and hypertension [4], [5]. There are no licensed antiviral drugs for DENV treatment. Administration of chloroquine (a 9-aminoquinoline) exerts direct antiviral effects by inhibiting the pH-dependent steps of flavivirus replication, although this drug is failed to inhibit the duration of viremia and antigenemia in DENV patients [6]. Balapiravir (4'-azidocytidine) is developed for the treatment of chronic hepatitis C Virus infection by a nucleoside analogue of RNA-dependent RNA polymerase; however, this drug does not alter the kinetics of viremia and NS1 antigenemia in DENV patients [7]. During the early acute phase of DENV infection, oral prednisolone is not related to prolongation of viremia or other pathogenic effects [8]. A recent trial showing that the α-glucosidase inhibitor celgosivir (6-O butanoyl prodrug of castanospermine) has antiviral activity by modulating the host's unfolded protein response, but it does not significantly reduce viral load or fever burden in DENV patients [9]. The development of a DENV vaccine would represent a powerful new tool for preventing DENV infection. Although a safe vaccine is not yet available, a number of candidate vaccines and strategies for strengthening vaccine efficiency are under active investigation [1], [10], [11].
DENV is an enveloped, single-stranded RNA virus that contains several types of structural proteins, including envelope protein (E), precursor membrane protein, and capsid protein, as well as several types of nonstructural (NS) proteins, including NS1, NS2A, NS2B, NS3, NS4A, NS4B, and NS5 [12]. All of the DENV proteins function in the viral biology and pathogenesis. The DENV E protein is the viral receptor for cell binding and fusion [13]. The cellular targets of DENV include monocytes/macrophages, dendritic cells, B cells, T cells, basophil/mast cells, endothelial cells, epithelial cells, and hepatocytes [14]. DENV infects and/or interacts with cells through a variety of cell-surface molecules, including heparan sulfate [15], integrins [16], dendritic cell-specific intracellular adhesion molecule 3 grabbing non-integrin (DC-SIGN) [17], C-type lectin domain family 5 member A (CLEC5A) [18], and heat shock proteins [19]. An alternative route for DENV infection is receptor-mediated endocytosis, following viral-cell receptor interaction [14]. The generation of antibodies (Abs) against the DENV E protein is fundamental for the host defense; however, such immune responses may increase the risk of developing DHF/DSS upon re-infection, primarily due to the canonical effects of antibody-dependent enhancement (ADE) – a phenomenon in which non-neutralizing anti-E Abs cross-react with the heterogeneous serotypes of DENV and facilitate their binding with Fcγ receptor-bearing cells to cause severe infection [20]. Combined with DENV-induced autoimmunity [2], [3], these effects could represent the primary challenges to DENV vaccine development.
DENV pathogenesis can be affected by many viral factors, including viral load, serotype, and virulence [21], [22]. However, the immunopathogenesis of DENV infection is caused by host-specific immune responses, including immune cell activation (CD4 positive T cells), cytokines (interleukin (IL)-1β, IL-2, IL-6, IL-10, IL-13, IL-18, macrophage migration inhibitory factor, tumor growth factor-β, tumor necrosis factor-α, and interferon (IFN)), chemokines (IL-8, monocyte chemoattractant protein-1, and regulated and normal T cell expressed and secreted), complement activation (C5a and C5b-9), inflammatory mediators (high mobility group box 1), and autoimmunity (auto-Abs against platelets, endothelial cells, and coagulants), all of which have been identified as hallmarks of DHF/DSS [1]–[3], [13], [20], [23], [24]. IL-10 has been proposed to play a role in DENV pathogenesis due to its immunosuppressive functions during IFN resistance and persistent viral infection. Epidemiological studies have demonstrated that IL-10 serum levels are higher in DHF/DSS patients than in DF patients with acute illness [25]–[31]. However, following ADE of DENV infection in monocytes [20], [32]–[34], it has been speculated that the Fcγ receptor might facilitate the IL-10 response to induce the expression of the suppressor of cytokine signaling (SOCS) 3, a downstream effector of IL-10 that mediates immunosuppression [35]. The existence of an intrinsic model for ADE [20], [33], differing from the canonical extrinsic ADE pathway, enhances DENV infection by facilitating IL-10/SOCS3-mediated benefits to escape from antiviral IFN responses, such as type I IFNs production [34] and T cell activation [36]. In this study, we investigate the molecular regulation of IL-10 production in monocytes infected with DENV, both directly and via ADE.
The reagents and antibodies used were polybrene, PKA inhibitor H-89, PI3K inhibitor 2-(4-Morpholinyl)-8-phenyl-4H-1-benzopyran-4-one hydrochloride (LY294002), and PKC inhibitor bisindolylmaleimide (Bis) (Calbiochem, San Diego, CA); 4,6-diamidino-2-phenylindole (DAPI), GSK-3 inhibitor BIO, heparin lyase III, chondroitin ABC lyase, O-linked glycosylation inhibitor benzyl-α-GalNAc, N-linked glycosylation inhibitor tunicamycin, Syk inhibitor BAY61-3606, dimethyl sulfoxide (DMSO), and mouse mAb specific for β-actin (Sigma-Aldrich, St. Louis, MO); recombinant human IL-10 (PeproTech, Rocky Hill, NJ); Abs against DENV NS4B and E (GeneTex, San Antonio, TX); Abs against SOCS3, phospho-CREB at Ser133, CREB, phospho-PKB at Ser473, PKB, phospho-GSK-3β at Ser9, GSK-3β, β-catenin, and Mcl-1 (Cell Signaling Technology, Beverly, MA); Abs against TIM1, Axl, and IL-10 (R&D Systems, Minneapolis, MN); Abs against isotype control IgG (Millipore, Billerica, MA); donkey anti-goat IgG conjugated with horseradish peroxidase (HRP) (Santa Cruz Biotechnology, Santa Cruz, CA) and goat anti-rabbit IgG conjugated with HRP (Chemicon International, Temecula, CA); rabbit anti-mouse IgG conjugated with HRP (Abcam, Cambridge, MA); and Alexa Fluor 488- and Alexa Fluor 594-conjugated goat anti-mouse and goat anti-rabbit (Invitrogen, Carlsbad, CA). All drug treatments were assessed for cytotoxic effects using cytotoxicity assays prior to experiments. Non-cytotoxic dosages were used in this study.
Human monocytic THP-1 cells were routinely grown on plastic in RPMI Medium 1640 (RPMI; Invitrogen Life Technologies, Rockville, MD), with L-glutamine and supplemented with 10% heat-inactivated fetal bovine serum (FBS; Invitrogen Life Technologies), 50 units of penicillin, and 50 µg/ml of streptomycin. Baby hamster kidney (BHK) cells and C6/36 cells were cultured in Dulbecco's modified Eagles medium (DMEM; Invitrogen Life Technologies) containing FBS. Monocyte-enriched peripheral blood mononuclear cells (PBMC) were isolated from healthy volunteers by density-gradient centrifugation using Ficoll-paque Plus (GE Healthcare, Piscataway, NJ), washed twice with red blood cell lysis buffer (eBioscience, San Diego, CA), resuspended in RPMI 1640 medium supplemented with 10% heat-inactivated fetal bovine serum, and maintained at 37°C in an atmosphere containing 5% CO2 while allowing adherence on uncoated polystyrene flasks during 90 min for monocyte enrichment. Non-adherent cells were gently removed by washing, after which the adherent cells were collected to perform the DENV infection experiment [37]. The protocols and procedures were approved by the institutional review board of the National Cheng Kung University Hospital (No. A-ER-102-123). Four serotypes of DENV (DENV1 8700828, DENV2 PL046 and 454009A, DENV3 8700829A, and DENV4 59201818) were maintained in C6/36 cells. Monolayers of C6/36 cells were incubated with DENV at a MOI of 0.01 and incubated at 28°C in 5% CO2 for 5 days. The virus supernatant was further filtered with 0.22 µm filter, and then stored at −80°C until use. Virus titer was determined by plaque assay, using the BHK cell line.
Cells were resuspended at a concentration of 5×105 cells/ml in appropriate medium with DENV (MOI = 1) and incubated for 90 min at 37°C. Then, the cells were washed once with RPMI medium, resuspended at a concentration of 5×105 cells/ml, and incubated at 37°C with 5% CO2. Monoclonal anti-E 50-2 IgG1 Ab, which can recognize viral E protein and shows both neutralization and enhancement activity, was used to induce ADE of the DENV infection, as described previously [38]. To prepare the UV-inactivated virus, DENV was exposed to a 15 W UV lamp at a distance of 10 cm for 1.5 h. The viral supernatants were checked using plaque assays.
BHK-21 cells were plated into 12-well plates (2×105 cells/well) and cultured in DMEM under CO2-enriched conditions. After adsorption with a serially diluted virus solution for 1 h, the solution was replaced with fresh DMEM containing 2% FBS and 0.5% methyl cellulose (Sigma-Aldrich). Five days post-infection, the medium was removed, and the cells were fixed and stained with crystal violet solution containing of 1% crystal violet, 0.64% NaCl, and 2% formalin.
Both procedures are described elsewhere [39]. In brief, total cell lysates were extracted and proteins were separated using SDS-polyacrylamide gel electrophoresis and then transferred to a polyvinylidene difluoride membrane (Millipore). After blocking, blots were developed with the indicated Abs and developed using an ECL Western blot detection kit (Pierce Chemical, Rockford, IL), according to the manufacturer's instructions. Following densitometer-based quantification and analysis using ImageJ software (http://rsbweb.nih.gov/ij/), the relative densities of the identified proteins were calculated.
After treatment, we used a commercial ELISA kit (88-7106-77, eBioscience) to detect the concentration of human IL-10 in cell-conditioned culture medium, according to the manufacturer's instructions. For the luciferase reporter assay, the cells were transiently co-transfected using a GeneJammer reagent (Stratagene, La Jolla, CA), with an IL-10 promoter-driven luciferase reporter (0.2 µg), kindly provided by Dr. Yu-Ming Wang, Institute of Bioinformatics and Biosignal Transduction, National Cheng Kung University, and 0.01 µg of Renilla luciferase-expressing plasmid (pRL-TK; Promega, Madison, WI). Twenty-four hours after the transfection, the cells were infected with DENV for 24 h, lysed, and then harvested for luciferase and Renilla measurement, using a luciferase assay system (Dual-Glo; Promega). For each lysate, the firefly luciferase activity was normalized to the Renilla luciferase activity to assess transfection efficiencies.
After DENV infection, we used a commercial ProFluor® PKA Assay kit (V1240, Promega) and a PIP3 Mass ELISA kit (K-2500s, Echelon Biosciences, Salt Lake City, UT) to detect the activity of PKA and PI3K in THP-1 cells, according to the manufacturers' instructions.
To detect expression of CREB and DENV2 E protein, cells were fixed with 4% paraformaldehyde, permeabilized with 0.5% Triton X-100, and washed twice with ice-cold PBS. Cells were stained with anti-CREB and DENV E Abs, and then with Alexa 488-conjugated goat anti-mouse IgG and Alexa 594-conjugated goat anti-rabbit IgG. DAPI (5 µg/ml) was used for nuclear staining. Cells were visualized under a fluorescent microscope (BX51; Olympus, Tokyo, Japan) or a laser-scanning confocal microscope (SPII; Leica Mikrosysteme Vertrieb, Bensheim, Germany). The three-dimensional images reconstructed from a series of confocal images, along with the z-axis of the cells and the analysis of z-stacks, were reconstructed using the Leica Confocal Software. For flow cytometric analysis, cells were fixed and stained with anti-TIM1, Axl, and NS4B Abs as described elsewhere [40], and then incubated with a mixture of Alexa Fluor 488-conjugated secondary Ab. Cells were analyzed using flow cytometry (FACSCalibur; BD Biosciences, San Jose, CA) with excitation set at 488 nm; emission was detected with the FL-1 channel (515–545 nm). Samples were analyzed using CellQuest Pro 4.0.2 software (BD Biosciences). Small cell debris was excluded by gating on a forward scatter plot.
Protein was downregulated using lentiviral expression of short hairpin RNA (shRNA) targeting IL-10 (TRCN0000058462 containing the following respective shRNA target sequences: 5'-GCCTACATGACAATGAAGATA-3'), GSK-3β (TRCN0000010551 containing the following respective shRNA target sequences: 5'- CACTGGTCACGTTTGGAAAGA-3'), and a negative control construct (luciferase shRNA, shLuc). The shRNA clones were obtained from the National RNAi Core Facility, Institute of Molecular Biology/Genomic Research Center, Academia Sinica, Taipei, Taiwan. Lentiviruses were prepared and cells were infected according to previously described protocols [39]. In brief, THP-1 cells were transduced by lentivirus, with an appropriate multiplicity of infection, in complete growth medium supplemented with polybrene (Sigma-Aldrich). After transduction for 24 h and puromycin (Calbiochem) selection for 6 days, protein expression was monitored using Western blot analysis. CREB and CLEC5A expression was silenced using commercialized siRNA (clone #1, CREB1-HSS102262 containing the following respective siRNA target sequences: 5′-UUACGGUGGGAGCAGAUGAUGUUGC-3′ and 5′-GCAACAUCAUCUGCUCCCACCGUAA-3′; clone #2, CREB1-HSS102264 containing the following respective siRNA target sequences: 5′-UUGCUGGGCACUAAGAUCUGCUGUC-3′ and 5′-GACAGCAGAUCUUAGUGCCCAGCAA-3′ for CREB silencing and CLEC5A-HSS119041 containing the following respective siRNA target sequences: 5′-AAUAAGCCCAGAGAUGAUCAUGUGC -3′ and 5′-GCACAUGAUCAUCUCUGGGCUUAUU-3′ for CLEC5A silencing) (Invitrogen). Transfection was performed by electroporation using a pipette-type microporator (Microporator system; Digital Bio Technology, Suwon, Korea). After transfection, THP-1 cells were incubated for 18 h in RPMI medium at 37°C before infection. A nonspecific scrambled siRNA kit (StealthTM RNAi Negative Control Duplexes, 12935-100; Invitrogen) was the negative control.
Data obtained from three independent experiments are presented as the mean ± standard deviation (SD). Statistical analysis of data analyses were performed using Prism version 5 (GraphPad Software, San Diego, CA). Two sets of the data were analyzed by an unpaired Student's t test. Three or more sets of data were analyzed by one-way ANOVA with Tukey's multiple-comparison posttest. Statistical significance was set at P<0.05.
DENV has a variety of cellular targets, the most common being mononuclear phagocytes [14], [41]. Furthermore, IL-10 production is upregulated in monocytes following ADE of DENV infection [20], [32]–[34]. Human monocytic THP-1 cells were infected with DENV serotype 2 PL046, as demonstrated by plaque assays (Figure 1A, upper panel), and the quantitative data of Western blotting revealed the time-kinetic expression of viral NS4B protein (Figure 1A, middle and lower panels), which was first detectable 12 h post-infection and increased significantly (P<0.001) by 48 h post-infection. ELISA showed that either DENV infection alone or treatment of ultraviolet-inactivated DENV was sufficient to significantly increase IL-10 production in THP-1 cells (P<0.001) (Figure 1B), as well as the significant (P<0.001) expression of its downstream target SOCS3 as determined by Western blotting (Figure 1C). Without DENV infection, supernatants of C6/36 cells did not cause IL-10 production (Figure S1 in Text S1). In human peripheral blood monocytes, DENV infection also significantly (P<0.001) caused viral replication (Figure 1D, left panel) and IL-10 production (Figure 1D), right panel. To further demonstrate the essential role of IL-10 in SOCS3 expression, a lentiviral-based short hairpin RNA (shRNA) approach was used. In IL-10 silenced cells, the DENV-induced IL-10 expression was significantly abolished (P<0.001) (Figure 1E, left panel), accompanied by a decrease in DENV-induced SOCS3 expression (Figure 1E, right panel). These results demonstrate that DENV can induce an IL-10 response in monocytes through an infectious process and/or attachment.
Although we determined that DENV infection and/or attachment induces IL-10 production, the molecular mechanisms underlying the IL-10 expression were unclear. IL-10 production can be regulated by a variety of transcription factors [42]. We therefore investigated the activity of one of these in DENV-infected THP-1 cells: CREB, a transcriptional factor that activates IL-10 expression [43]–[45]. The quantitative data of Western blotting demonstrated that DENV infection significantly (P<0.01) activated CREB by inducing Ser133 phosphorylation in a time-dependent manner (Figure 2A). Immunostaining for expression of the DENV E and CREB proteins revealed that CREB translocated to the nucleus following DENV infection or stimulation by ultraviolet-inactivated DENV (Figure 2B). Plasmids expressing small interfering RNAs (siRNAs) specific to CREB (siCREB) were used to silence CREB expression in THP-1 cells, and this knockdown of CREB significantly (P<0.001) decreased DENV-induced IL-10 production (Figure 2C, upper panel), accompanied by SOCS3 down-regulation (Figure 2C, lower panel). CREB phosphorylation is mediated by PKA, PI3K/PKB, and PKC [42]. Treatment with both the PKA inhibitor H-89, which selectively inhibits only PKA, and the PI3K inhibitor LY294002 significantly (P<0.001) decreased DENV-induced IL-10 production (Figure 2D, upper panel) and efficiently reduced CREB phosphorylation (Figure 2D, lower panel); treatment with the broadly acting PKC inhibitor bisindolylmaleimide-1 (Bis) had no effect. Furthermore, inhibiting PKC by using myristoylated PKC inhibitor also did not cause a decrease in the DENV-induced IL-10 (Figure S2 in Text S1). Next, we investigated the potential regulation of PI3K/PKB by PKA, which has been demonstrated previously [46]. Using activity assays, we determined the time courses of PKA (Figure 2E) and PI3K (Figure 2F) activation, revealing an early activation of PKA by 1 h post-infection. Western blotting confirmed that DENV infection induces PKB phosphorylation at Ser473 (Figure 2G). Notably, pharmacologically inhibiting PKA and PI3K, but not PKC, differentially inhibited DENV-induced PKB phosphorylation (Figure 2H), suggesting that PKA, at least in part, acts upstream of PI3K/PKB. These results demonstrate that DENV infection activates PKA, PI3K/PKB, and CREB in a sequential manner, leading to IL-10 production in monocytes.
In addition to PKA and PI3K/PKB, CREB is also regulated by GSK-3β, which decreases the stability of CREB by phosphorylating CREB at Ser129 [47]. Furthermore, GSK-3β function – which is controlled by the PKA [48] and PI3K/PKB signaling pathways [49] – is important for CREB activity [50] and IL-10 production [51], [52]. Therefore, we hypothesized that GSK-3β may be inactivated during DENV-induced IL-10 production. Indeed, the quantitative data of Western blotting showed that GSK-3β was significantly (P<0.01) inhibited by phosphorylation at Ser9 [53] in DENV-infected THP-1 cells (Figure 3A, upper panel); this GSK-3β inhibition was also accompanied by an accumulation of β-catenin and Mcl-1 protein, substrates that are negatively regulated by GSK-3β [54], [55] (Figure 3A, lower panel). Furthermore, treating cells with the GSK-3 inhibitor BIO significantly (P<0.001) enhanced DENV-induced IL-10 production (Figure 3B) and effectively increased CREB phosphorylation at Ser133 (Figure 3C). shRNA-based GSK-3β silencing (Figure 3D) caused an increase in CREB phosphorylation at Ser133, and pharmacological inhibition of PKA and PI3K resulted in CREB dephosphorylation (Figure 3E). These findings indicate an important but not strictly necessary role for GSK-3β in regulating CREB activity during DENV infection. In addition, the DENV-induced inactivation of GSK-3β was found to be regulated by both PKA and PI3K/PKB (Figure 3F). Taken together, these data indicate that following DENV infection, both PKA and PI3K/PKB inhibit GSK-3β activity to coordinately facilitate CREB-mediated IL-10 production in monocytes.
Next, we investigated the role of host receptors in regulating DENV-induced IL-10 production. Heat-inactivated DENV failed to induce IL-10 production in THP-1 cells (Figure S3 in Text S1), whereas ultraviolet-inactivated DENV induced IL-10 normally, indicating the essential role of structural proteins for this aspect of infection. Various host receptors expressed on cell surfaces have been reported to bind the DENV E protein [14]. We used heparan lyase III to cleave the extracellular heparan sulfate, and benzyl-α-GalNAc and tunicamycin to block O- and N-linked glycosylation, respectively. Although glycosylated heparan sulfate was previously reported to function as a DENV receptor [15], neither heparan sulfate cleavage nor inhibition of glycosylation resulted in inhibition of DENV-induced IL-10 production (Figure 4A). DC-SIGN and β3-integrin are host cell receptors for DENV infection [16], [17]; however, there was no expression of these proteins on the surface of THP-1 cells (Figure S4 in Text S1). Additionally, the results of competitive assays that utilized neutralizing Abs to block these proteins confirmed the independent roles of these receptors for DENV-induced IL-10 production in THP-1 cells (Figure S5 in Text S1). The cell-surface phosphatidylserine receptors TIM-1 and Axl, which were originally identified as surface receptors for the recognition of apoptotic cells, were recently identified as potential DENV entry receptors [56]. Immunostaining, followed by flow cytometric analysis, revealed low levels of expression of TIM-1 and Axl in THP-1 cells (Figure 4B); however, inhibition of TIM-1 and Axl by neutralizing Abs did not reduce DENV entry (Figure 4C, upper panel) or DENV-induced IL-10 production (Figure 4C, lower panel). CLEC5A, a member of the C-type lectin superfamily, plays a crucial role in the DENV infection-associated cytokine response [18]. We showed that CLEC5A expression was higher in DENV-susceptible THP-1 cells than in other monocytic HL-60 and U937 cells (Figure S6 in Text S1). Furthermore, A positive relationship between CLEC5A, infectious ability, and IL-10 production was also shown. Knockdown of CLEC5A expression in THP-1 cells (Figure 4D) partly, but significantly (P<0.001), decreased IL-10 production following DENV infection or stimulation by UV inactivated-DENV (Figure 4E). Next, we investigated the potential regulation of CLEC5A-regulated signaling during IL-10 production. Activation of the tyrosine kinase Syk positively regulates DNX activating protein 12, the downstream adaptor protein of CLEC5A [57]; pharmacological inhibition of Syk with the selective inhibitor BAY-61-3606 significantly (P<0.001) reduced DENV-induced IL-10 production (Figure 4F). These results demonstrate a novel role for Syk signaling in DENV-induced IL-10 production in monocytes, most likely in a CLEC5A-regulated manner.
Experiments have demonstrated that ADE not only facilitates DENV entry through the Fcγ receptor but also increases DENV-induced IL-10 expression [20], [32]–[34]. To investigate the regulation of IL-10 production during ADE of DENV infection, anti-E (clone 50–2) monoclonal (m) Abs were used to induce ADE, as described in a previous study [38]. A plaque assay confirmed the enhanced infection of DENV by ADE (P<0.05; Figure 5A, upper panel), and immunostaining-based flow cytometric analysis (P<0.001; Figure 5A, middle panel) and Western blotting (Figure 5A, lower panel) confirmed the increased levels of viral NS4B expression in THP-1 cells following ADE of DENV infection. As compared with DENV infection alone, the effects of ADE were further evidenced by increased IL-10 production (P<0.01; Figure 5B, upper panel) and transcriptional activation of IL-10 (P<0.01; Figure 5B, lower panel). Treating anti-E (clone 50–2) alone also did not cause IL-10 production. Using activity assays, we confirmed that ADE significantly increased DENV-activated PKA (P<0.01; Figure 5C, upper panel) and PI3K activity (P<0.001; Figure 5C, lower panel). Furthermore, Western blotting demonstrated increased phosphorylation of PKB at Ser473 and GSK-3β at Ser9 in THP-1 cells under ADE of DENV infection (Figure 5D). Consistent with DENV infection alone, silencing of CREB expression (Figure 5E, left panel) significantly (P<0.01) reduced IL-10 production following ADE (Figure 5E, right panel). Notably, Syk is also required for Fcγ receptor signaling and may be involved in IL-10 regulation [20]. Pharmacological inhibition of Syk, PI3K, and PKA significantly (P<0.001) reduced IL-10 production following ADE of DENV infection (Figure 5F). Furthermore, the Syk inhibitor BAY-61-360 also decreased ADE-induced phosphorylation of PKB at Ser473, GSK-3β at Ser9, and CREB at Ser133 (Figure 5G). These results demonstrate that Syk regulates the PI3K/PKB/GSK-3β/CREB pathway during ADE-induced IL-10 production in monocytes.
No significant differences in the ability of the four DENV serotypes to induce IL-10 production following DENV infection in THP-1 cells were observed between the four serotypes (Figure 6A). Although intrinsic ADE has been hypothesized to facilitate IL-10 production, most likely through intracellular signaling of the Fcγ receptor II [20], [32]–[34], the virus-cell interaction is also enhanced extrinsically in the canonical ADE pathway by increasing the infection rate in Fcγ receptor-bearing cells. Next, the effects of different DENV viral loads on IL-10 induction and signal regulation were investigated. Notably, a high viral load of DENV infection alone, as demonstrated by plaque assays (P<0.001; Figure 6B), significantly (P<0.001) induced IL-10 production (Figure 6C) in a multiplicity of infection (MOI)-dependent manner, although phosphorylation of PKB at Ser473 and GSK-3β at Ser9 was not increased (Figure 6D). Notably, the results of IL-10 production in THP-1 cells infected with DENV alone under a high MOI condition or with ADE infection of DENV under a lower MOI were similar. These results indicate that viral load may affect DENV-induced IL-10 production in monocytes, independent of serotype.
IL-10 serum levels are higher in patients with DHF/DSS [25]–[31], and furthermore, ADE-enhanced IL-10/SOCS3 expression may interfere with the antiviral response to IFN [20], . Based on our earlier findings, we next investigated the importance of IL-10 signaling for DENV replication. With or without ADE, replication of DENV in THP-1 cells, as determined by plaque assays (P<0.001; Figure 7A, upper panel) and Western blotting of NS4B expression (Figure 7A, lower panel), was inhibited by the presence of neutralizing IL-10 Abs. With or without ADE, DENV replication, as determined by plaque assays and Western blotting of NS4B expression, was completely blocked by genetically silencing CREB (P<0.001; Figure 7B) and pharmacologically inhibiting Syk, PI3K, and PKA (P<0.01; Figure 7C). Consistent with these findings, inhibition of Syk, PI3K, and PKA also reduced DENV replication, as determined by plaque assays and Western blotting of NS4B expression, in human peripheral blood monocytes (P<0.001; Figure 7D). These inhibitors did not cause cell cytotoxicity (Figure S7 in Text S1). These results show that IL-10 facilitates DENV replication and that altering IL-10 regulation can affect viral replication in monocytes.
Both the physiological and the pathogenic roles of IL-10 are immunosupressive. IL-10 not only suppresses inflammation during immune resolution but also affects pathogen clearance and helps alleviate immunopathology [58]. In particular, during microbial infection, IL-10 plays an essential role in relieving IFN-γ- and TNF-α-mediated immunopathology [59], [60] as well as IFNs-mediated antiviral responses. In addition to the canonical extrinsic ADE pathway, which facilitates DENV virus-cell interactions, an intrinsic ADE pathway exists that may play a role during persistent viral infections by inducing IL-10-mediated immune suppression, particularly on IFNs responses [20], [32]–[34]. ADE of DENV infection requires the Fcγ receptor [34], but the molecular regulation of Fcγ receptor signaling for IL-10 expression remains unclear; therefore, we chose to further investigate this pathway in monocytes. As summarized in Figure 8, combined with the previous studies which ADE facilitates IL-10 production [20], [32]–[34], we speculated that ADE of DENV infection and/or attachment induces not only the Fcγ receptor/Syk-facilitated (i.e., intrinsic) pathway but also the Fcγ receptor/CLEC5A partly/Syk-mediated (i.e., extrinsic) pathway, which both lead to PKA/PI3K/PKB activation, followed by CREB-mediated IL-10 production. In this study, we did not exclude the involvement of CLEC5A/Syk signaling for the intrinsic pathway of IL-10 production under ADE. In addition, we demonstrated that PKB phosphorylates GSK-3β and decreases its activity, enhancing CREB stability and inducing IL-10 production, consistent with previous studies [43]–[45], [51], [52]. However, the regulation of PKA by DENV receptors and effectors requires further investigation. Following IL-10 induction, SOCS3 is upregulated and may facilitate attenuation of T cell responses [36] and induction of IFN resistance to enhance viral replication [20], [32]–[34]. By interfering nitric oxide generation, which is regulated by STAT1/IRF1 signaling and confers anti-DENV activity, IL-10 may cause immunosuppression through SOCS3 expression [32], [34]. Based on our findings, IL-10 appears to play a permissive role with respect to DENV pathogenesis, and regulating IL-10 production may therefore provide cellular protection against DENV infection, even under ADE conditions. However, this hypothesis needs to be approved in vivo by using an appropriate animal model.
Following DENV infection, host cells secrete a variety of immune mediators to mediate anti-viral responses and inflammatory activation. Severe cases of DHF/DSS caused by DENV infection are clearly the result of interactions between viral and host responses. Interest in IL-10 is increasing for a variety of reasons. First, IL-10 may be a useful prognostic tool because IL-10 serum levels are positively correlated with dengue disease severity, particularly in DHF/DSS patients [25]–[31], and because IL-10 displays immunosuppressive properties during DENV infection [20], [32]–[34]. The role of the expression and activation of CD25-positive regulatory T cells, as well as relevant genetic polymorphisms, in IL-10 overproduction were previously evaluated [61], [62]; however, these studies did not provide a strong link between IL-10 and DHF/DSS progression, particularly in the context of ADE. Therefore, several possibilities remain: ADE causes IL-10 overexpression, IL-10 production facilitates ADE, or both processes are synergistically coupled. Furthermore, it is also possible that aberrant production of IL-10 could be the result of intrinsic regulation by ADE of DENV infection [32], [34]. Based on this hypothesis, the IL-10 response elicited by ADE may increase resistance to antiviral IFN-mediated immune surveillance. This study therefore favorably explores the pathogenic regulation of IL-10 signaling in ADE infection of DENV. However, the expression and the role of IL-10 can be pathologically as well as physiologically. Although clinical studies have shown that the increased levels of IL-10 can be detected and show impact on dengue disease statement. It should not be exclude that the increased levels of IL-10 may be the result of a regulatory effect more than pathogenic during later phase of the disease progression.
To escape immune surveillance, the DENV can induce several immunosuppressive mediators through unknown mechanisms, including oncostatin M (OSM; an activator of SOCS3 expression), protein inhibitor of activated STAT1 (PIAS1; a negative regulator of STAT1), SOCS1, and SOCS3 [34]. SOCS3 plays a key role downstream of IL-10 signaling in modulating the immune response [63]. With respect to IFN resistance, SOCS3 binds to IFGR1 to inhibit STAT1 interactions, and numerous human viruses, such as hepatitis C virus (HCV), human immunodeficiency virus, and Epstein-Barr virus, can facilitate SOCS3 expression [64]–[67]. Increased serum IL-10 levels as well, as increased SOCS3 expression in PBMCs, have been demonstrated previously in severe DHF/DSS patients [32], [34]. Furthermore, DENV-induced IL-10 may interfere the production of IFN-γ in T cell activation [36]. These findings confirm that IL-10 is important for DENV infection and replication. Although IL-10 may act upstream of SOCS3 in particular, a variety of transcription factors are involved in IL-10 production, including CREB, NF-κB, and MAF [68]. Regarding the early detection of IL-10 at 24 h post-infection, the factors which show crosstalk with CREB are speculated to be important for regulating IL-10 expression. In this study, we observed that DENV infection significantly activates CREB and that CREB silencing reduces DENV replication. Importantly, CREB is a transcription factor involved in the regulation of glucose homeostasis, growth factor-dependent cell survival, and immune modulation [42]. Furthermore, a pathogenic role for CREB has been demonstrated during HBV pathogenesis; the HBV X protein activates viral gene transcription by interacting with CREB, leading to activation of the HBV enhancer I [69]. CREB can also induce expression of early growth response-1, which facilitates herpes simplex virus-1 replication [70]–[72]. In addition to its role in IL-10/SOCS3-mediated immunosuppression, we hypothesize that CREB plays a role in transcriptional regulation during DENV infection and replication, which requires further investigation.
We demonstrated that DENV infection induces PKA/PI3K/PKB-mediated CREB phosphorylation, followed by IL-10 production. However, PKC was not required for these effects, even though PKC has been speculated to act downstream of Fcγ-receptor signaling during ADE of DENV infection [20], [33]. TLR3 recognizes the dsRNA form of the replicated DENV virus [73]. Following stimulation with UV-inactivated DENV, THP-1 cells still produced IL-10, suggesting the existence of IL-10 production mechanisms that are independent of viral replication. As heat-inactivated DENV did not induce IL-10 production (Figure S3 in Text S1), this suggests that the viral E protein is essential for IL-10 signaling during DENV infection. A previous study reported that the DENV E protein activates PKA through an unknown mechanism [74]. PKA is a cAMP-dependent kinase, and the activation of guanine nucleotide binding protein-coupled receptors increases intracellular cAMP levels to facilitate PKA-mediated CREB activation [75]. Therefore, other receptors and effectors of DENV-activated cAMP/PKA activation should be investigated in future studies. Signaling by integrins, which are potential receptors for DENV binding and entry, typically induces PKA-mediated CREB activation following an increase in intracellular cAMP levels [76]. However, we demonstrated that blocking integrins did not inhibit DENV-induced IL-10 production, and THP-1 cells did not express β1 and β3 integrins. We also eliminated DC-SIGN and heparan sulfate as mediators of IL-10 induction. A study of influenza A virus-infected cells demonstrated that cyclooxygenase (COX)-2-mediated prostaglandin E2 (PGE2) expression facilitates cAMP/PKA/CREB activation [77], and indeed, PGE2 receptors can rapidly increase cAMP levels [78]. Notably, DENV infection also induces COX-2 expression, followed by PGE2 production [79]. We hypothesize that DENV-induced soluble effectors may contribute to the activation of cAMP/PKA/CREB/IL-10 signaling.
In addition to ADE of DENV infection, both the Fcγ receptor and CLEC5A trigger Syk activation [57], [80], [81]. Syk is required for PI3K/PKB activation [82], [83], and PKA may also regulate PI3K/PKB [46], but the potential regulation of PKA by Syk remains unclear. However, the intracellular domain of the Fcγ receptor is required for ADE-facilitated DENV infection, indicating that intracellular signaling is initiated by the Fcγ-receptor complex [84]. Our findings speculate that both the extrinsic (Fcγ receptor/CLEC5A partly) and intrinsic (Fcγ receptor) ADE pathways can trigger Syk activation, followed by activation of the PI3K/PKB signaling axis. However, the presented data from this study has not excluded the possibility that Syk activation is also mediated by CLEC5A even under the intrinsic ADE infection. The contribution of the Fcγ receptor and viral receptors is not clear and needs further investigation. While this finding shows an important role of Syk for DENV infection, an opposing role of Syk is proposed by its role in facilitating type I IFN response and the expression of type I IFN-stimulated genes [85]. For ADE infection, a possible regulation of FcR-activated Syk may be negatively regulated by coligation with leukocyte immunoglobulin-like receptor B1 to inhibit type I IFN response [86]. A different regulation of Syk signaling by DENV and FcR is speculated for viral infection. Furthermore, this does not excluded the possibility of synergistic activation of IL-10 regulation by these two pathways during ADE of DENV infection and/or by other receptors, which may link to Syk activation.
ADE of DENV infection may induce Fcγ receptor/Syk- and/or CLEC5A partly/Syk-mediated PI3K/PKB activation. Activated PKB phosphorylates and activates CREB at Ser133, but it also phosphorylates and inactivates GSK-3β, which stabilizes CREB (GSK-3β phosphorylates CREB at Ser129, causing CREB downregulation) [47]. Consistent with previous studies [48]–[52], we observed that DENV-induced synergistic activation of PKA and PKB inhibited GSK-3β, leading to CREB-mediated IL-10 production. In monocytes/macrophages, PKC signaling is required for FcγR-mediated endocytosis [87]. Notably, PKC-induced GSK-3β inactivation facilitates the expression of IL-10 following LPS stimulation [88]. These results are inconsistent with the regulation of GSK-3β by DENV infection identified in this study, where PKC is not required for DENV-induced GSK-3β inactivation, as well as IL-10 production. GSK-3β modulation is known to be important for cell growth, differentiation, apoptosis, and inflammatory activation [89]. Inactivation of GSK-3β increases not only the stability of CREB for IL-10 production but also enhances the activity of β-catenin and Mcl-1 for cell growth and survival [54], [55]. Hepatitis C virus NS5A protein induces β-catenin accumulation by inactivating GSK-3β [90]. During Helicobacter pylori infection, PKB-mediated GSK-3β inactivation plays an essential role in Wnt signaling activation and cell proliferation [91]. Following DENV-induced GSK-3β inactivation, the physiological and pathological roles of accumulated β-catenin and Mcl-1 may contribute to the pathogenesis of DENV infection and replication.
In conclusion, an excessive or poorly timed IL-10 production may allow the pathogen to escape immune surveillance during DENV pathogenesis. This study demonstrates a molecular basis for IL-10 induction during DENV infection, as well as during ADE of DENV infection in human monocytes. For strengthening the significance of our findings, patients' serum and/or humanized antibodies are suggested to be examined for verifying the consistent pathway caused by Fc receptors' affinity. Current study showed that DENV infects macrophages and causes mild IL-10 production [92]. However, DENV infection does not induce significant IL-10 release from immature myeloid dendritic cells [93]. Not only in monocytes, it is crucial to check the identified pathways in other IL-10-producing cells and/or the natural targeting cells in vivo. With respect to a possible pathogenic role for aberrant IL-10 production in DHF/DSS patients, targeting the Syk/PKA/PI3K/PKB/GSK-3β/CREB signaling axis may represent a viable therapeutic strategy for combating the progression of severe dengue diseases.
|
10.1371/journal.pgen.1004913 | Genetic Mapping of MAPK-Mediated Complex Traits Across S. cerevisiae | Signaling pathways enable cells to sense and respond to their environment. Many cellular signaling strategies are conserved from fungi to humans, yet their activity and phenotypic consequences can vary extensively among individuals within a species. A systematic assessment of the impact of naturally occurring genetic variation on signaling pathways remains to be conducted. In S. cerevisiae, both response and resistance to stressors that activate signaling pathways differ between diverse isolates. Here, we present a quantitative trait locus (QTL) mapping approach that enables us to identify genetic variants underlying such phenotypic differences across the genetic and phenotypic diversity of S. cerevisiae. Using a Round-robin cross between twelve diverse strains, we identified QTL that influence phenotypes critically dependent on MAPK signaling cascades. Genetic variants under these QTL fall within MAPK signaling networks themselves as well as other interconnected signaling pathways. Finally, we demonstrate how the mapping results from multiple strain background can be leveraged to narrow the search space of causal genetic variants.
| Wild yeast strains differ in phenotypes that are controlled by highly conserved signaling pathways. Yet it remains unknown how naturally occurring genetic variants influence signaling pathways in yeast. We have developed an approach to facilitate the mapping of genetic variants that underlie these phenotypic differences in a set of wild strain. Our mapping approach requires minimal strain engineering and enables the rapid isolation of mapping populations from any strain background. In particular, we have mapped genetic variants in twelve highly diverse yeast strains. Further, we demonstrate how the mapping results from these twelve strains can be used jointly to narrow the number of genetic variants identified to a set of putative causal variants. We identify genetic variants in genes with various roles in cell signaling. Our results illustrate the interplay of different signaling pathways and which signaling genes are more likely to contain variants of large phenotypic impact.
| Cellular survival is dependent on the ability to sense and respond to changing environmental conditions. Mitogen-activated protein kinase (MAPK) signaling cascades are ubiquitous in eukaryotic organisms and enable them to react to extracellular stimuli [1]. MAPK cascades are composed of three sequentially acting kinases [2], which upon sensing an extracellular stimulus trigger a cellular response by activating transcription factors and other regulatory proteins [2], [3]. The yeast Saccharomyces cerevisiae has long been used as a eukaryotic model system to elucidate the principles of these signaling pathways, as many of the core signaling proteins are conserved from yeast to human [3].
In yeast, MAPK pathways facilitate response to environmental cues such as the presence of mating pheromones and stresses such as cell wall damage and high osmolarity [4]. Adaptation to high osmolarity is conducted by the HOG (high osmolarity glycerol) MAPK pathway. The human homolog of the MAPK Hog1, p38α, not only mediates the response to hyperosmolarity as well [4], but also plays key roles in inflammation and cancer [5], [6]. Insults to the cell wall are sensed by the cell-wall integrity (CWI) pathway, which is anchored by the MAPK Slt2, homolog of the mammalian MAPK7 [7]. While these pathways have been scrutinized in great detail, it remains largely unknown how they are affected by genetic variation among different yeast isolates. Response and resistance to MAPK-activating stress conditions are highly variable among different yeast isolates [8], [9], [10]. While genes of the core MAPK cascades are highly conserved across species, upstream regulatory components, such as stress sensors, and downstream targets, such as transcription factors, exhibit high levels of divergence [11], [12]. Yet, it is unknown how genetic differences in such elements of the MAPK pathways contribute to the phenotypic differences between isolates of the same species. We sought to examine how sequence variation among yeast isolates results in quantitative differences in MAPK-dependent phenotypes.
Yeast is an ideal model system for the study of quantitative traits. Quantitative trait locus (QTL) mapping studies in S. cerevisiae have characterized the genetic architecture of complex traits, including global gene expression and resistance to small molecules [13], [14], [15], [16], as well as shed light on the sources of “missing heritability” [17]. We have previously developed a bulk-segregant-analysis approach (BSA) [18], [19], named X-QTL, which utilizes extremely large mapping populations for increased mapping power [20]. Such mapping populations are subjected to selection to enrich for alleles that contribute to a trait of interest. We have recently used this method to determine the genetic underpinnings of protein abundance variation [21]. The majority of mapping studies to date have interrogated a pair of strains at a time, with only some recent studies expanding to crosses among four parent strains [22], [23]. Isolation of individual recombinant haploids, as used in linkage mapping studies, is very labor intensive, and current X-QTL protocols require extensive strain engineering. Consequently, mapping studies have queried a fraction of the S. cerevisiae species-wide genetic and phenotypic variation and have left much of the genetic architecture of quantitative traits across the species unexplored.
Here, we introduce a novel X-QTL approach and use it to map variants that alter MAPK-dependent traits in a panel of diverse yeast isolates. Our method takes advantage of plasmid-borne fluorescent markers to leverage the precision and high throughput of fluorescence-activated cell sorting (FACS) to rapidly generate large mapping populations of recombinant haploids. We charted the genetic architectures of salt (sodium chloride) and caffeine tolerance in a set of strains that capture much of the species-wide genetic diversity. Salt tolerance is mediated by the HOG MAPK pathway, while exposure of yeast to caffeine results in the activation of the CWI MAPK pathway [7]. In addition to these two pathways, we leverage the ability of our new mapping approach to isolate mapping populations of either mating type to examine the effect of the MAPK mating pathway on growth. We have identified QTL that contain genes with various roles in MAPK signaling and other signaling pathways. Our results illustrate how genetic variation within S. cerevisiae shapes MAPK-dependent traits. Finally, we illustrate how a round-robin cross design can be leveraged to identify causal genetic variants.
Our original X-QTL approach relied on the expression of the auxotrophic marker HIS3 under the control of the MATa-specific STE2 promoter for the isolation of MATa recombinant haploids [24]. However, most isolates of S. cerevisiae are generally prototrophs, lacking the auxotrophic mutations required for the original X-QTL approach. In addition, auxotrophic mutations can have large phenotypic effects [14]. For instance, the HIS3 deletion required under the original X-QTL approach results in sensitivity to various salts, which is not desirable when investigating natural variation in salt resistance [25]. Furthermore, the use of the STE2 promoter restricted X-QTL mapping populations to segregants of the MATa mating type. Here, we designed a novel approach that relies on mating-type specific expression of plasmid-borne fluorescent reporters, enabling the isolation of haploids of either mating type without the need to integrate the reporters into the genome (Fig. 1A). We placed the fluorescent markers mCitrine and mCherry under the control of the STE2 promoter and its MATα-specific equivalent, the STE3 promoter, respectively [24]. We combined these two fluorescent markers in a single construct and created a set of plasmids with drug resistances that permit facile introduction into any strain background (see Methods for details).
As expected, the presence of this construct resulted in green fluorescence in MATa strains and red fluorescence in MATα strains (Fig. 1B & C). Diploids carrying the marker construct were non-fluorescent, as neither promoter is active in diploid cells [24], [26]. After sporulation and germination, haploid progeny express the fluorescent marker corresponding to their mating type. We tested this aspect of the reporter construct by introducing it into the diploid hybrid between the well-characterized BY and RM strains and generating a pool of BY/RM recombinant haploids (see Methods for details). Flow cytometry revealed that this pool encompassed three cell populations: non-fluorescent cells and, in roughly equal proportions, green and red cells (Fig. 1D). We used fluorescence-activated cell sorting to isolate populations of green and red cells and then sequenced them in bulk to determine the BY and RM allele frequencies in each population. The BY parent strain contributes the MATa allele of the MAT locus, located on chromosome 3, while the MATα allele originates from the RM parent. As expected, the MATa allele was highly enriched in the green fluorescent recombinant haploids, while the MATα allele was enriched in the red population (Fig. 1E). This result demonstrates that the fluorescent markers can be used to isolate populations of MATa, as well as MATα, haploids.
In addition to the expected large allele frequency differences at the MAT locus, we identified several allele frequency skews that were highly concordant between the two populations (Fig. 1E). These skews correspond to previously identified growth QTL [17], [20]. For example, the skew on chromosome 12 is located at the gene HAP1; the BY allele of this gene results in partial loss of function and poorer growth compared to the RM allele, leading to the observed reduction in frequency of the BY allele. The agreement between the allele frequency distributions in the two mating type pools indicates that the isolated cell populations can be used to map QTL with excellent reproducibility in both mating types.
We set out to use our approach to dissect phenotypes mediated by different MAPK pathways. We used two stress conditions to query two different MAPK-signaling networks: high salt concentrations and caffeine. High salt conditions lead to activation of the HOG signaling cascade, while caffeine triggers the cell-wall integrity pathway.
We determined sodium chloride and caffeine tolerances for a collection of 65 S. cerevisiae isolates [27] and selected twelve diverse strains that spanned a range of resistances for our mapping studies (Fig. 2A, S1 Table). These twelve strains vary at approximately 80% of the common SNPs (allele frequency ≥0.05) segregating within the larger strain collection. We designed a round-robin crossing scheme in which each of the twelve strains was crossed to two other strains, for a total of twelve crosses (Fig. 2B). While the round-robin crossing scheme is the most efficient way to intercross a set of strains while maintaining equal contributions of each strain [28], it has only been used to map traits in a small number of organisms [29], [30]. Most mapping studies are based on crosses between resistant and sensitive strains, which implicitly assumes that resistance alleles are shared between resistant strains. By contrast, we randomly assigned the order of strains within the round-robin design without regard to strain phenotypes. As such, the crossing scheme included crosses between parent strains with different, as well as with similar, resistance phenotypes.
To assess the genetic complexity of salt and caffeine resistance, we isolated up to twenty individual segregants from each cross. We measured growth of these segregants and the parent strains under high salt and caffeine stress conditions (S1 Fig.). Three of the 24 phenotype distributions of the cross progeny were approximately bimodal, suggesting one large-effect determinant of resistance in these crosses. Other crosses exhibited phenotype distributions consistent with directional genetics and transgressive segregation [31]. In two crosses, the mean progeny phenotype was significantly lower than the midparent phenotype (S2 Table), indicative of interactions between the underlying causal loci [31], [32]. These findings illustrate the genetic complexity of the resistance phenotypes among wild isolates.
Next, we established a pipeline to map QTL between any pair of strains. Sequences of parent strains were used to generate lists of SNPs that segregated within a cross and could serve as genetic markers for QTL mapping. Mapping populations of both mating types were isolated and grown under permissive as well as selective conditions (YPD, YPD containing 0.5 M sodium chloride, 1 M sodium chloride, 15 mM caffeine or 20 mM caffeine). The resulting populations were bulk-sequenced to assess allele frequencies at SNP sites segregating in a particular cross. We used the MULTIPOOL software to calculate LOD scores testing whether allele frequencies at a given position in the genome in each selected population differed from those in the corresponding control population (S2 Fig.) [21], [33].
We used replicate BY/RM mapping experiments to determine the robustness and reproducibility of our mapping approach (see Materials & Methods for details). Allele frequency distributions of replicate experiments were extremely similar: comparisons between replicates did not exceed a LOD of 2.62 (S3A–S3B Fig.). At a LOD score threshold of five, QTL detection was highly reproducible, with 90.1% of QTL reproducibly detected in replicate experiments (S3C–S3D Fig., S3 Table). Reproducibility was also high between selections based on mapping populations of opposite mating types (88.3%).
We applied our mapping strategy to identify QTL that influence MAPK-dependent resistance traits in the twelve Round-robin crosses. We used the MATa and MATα selections for each cross as replicate mapping experiments (see S4A–S4B Fig. for combined LOD plots for each cross). We used a stringent criterion that a QTL must exceed a LOD threshold of five in both selections, and identified 155 QTL across the twelve crosses and the different conditions. Of QTL identified in only one of the two matched experiments, the majority had LOD scores close to the LOD threshold (median LOD 6.7 for unreplicated QTL vs. 12.3 for replicated QTL, Fig. 4A). Indeed, for 93.5% of unreplicated QTL the direction of the underlying allele frequency skews was the same in the matched MATa and MATα experiments (permutation based p<0.001). This suggests that many of these QTL stem from small effect variants that narrowly escaped detection in one of the two experiments.
Next, we asked how QTL are distributed and shared among the Round-robin crosses (Fig. 3 A & B, S4 Table). The salt QTL fell into 37 unique genomic regions, while the caffeine QTL comprised 23. There were 13 QTL regions that recurred between the conditions suggesting they maybe due to pleiotropic variants that affect both traits. We found an average of 6.2 salt resistance QTL (range 1 to 13) and 3.7 caffeine resistance QTL (range 0 to 7) per cross. Interestingly, the number of QTL detected per cross was independent of the genetic distance of the parent strains involved (divergence & salt QTL: Pearson correlation 0.154, divergence & caffeine QTL: 0.087). However, when the QTL of the two conditions were considered together per cross, this resulted in a weak correlation with the extent of parental divergence (0.257). We tested if the number of QTL per cross was a function of the phenotypic difference of the parent strains, since a previous study found that a higher number of QTL could be identified between yeast strains of similar phenotypes [34]. We found that this was the case for the salt resistance QTL (phenotypic difference & salt QTL: Pearson correlation −0.517), but not for caffeine resistance QTL (phenotypic difference & caffeine QTL: 0.395). While increasing the phenotypic similarity of yeast parent strains through the experimental fixation of large effect QTL can permit the identification of additional small effect QTL [35], [36], this trend is condition-specific for crosses between different wild isolates.
Because each genetic variant is introduced into the Round-robin scheme at least twice, each QTL is expected to be identified at least twice assuming 100% power and no genetic interactions. Contrary to this expectation, previous studies involving sets of interconnected crosses have found that the majority of QTL are ‘context-dependent’, that is their identification was dependent on a specific cross and not just on the parent strains involved [23], [34], [37]. Indeed, 33 of the 60 QTL regions we identified across the two conditions were found in only one cross (Fig. 4B). The overall lower LOD scores of these ‘context dependent’ QTL as a class partially explained their lack of detection in additional crosses (Fig. 4C; Pearson correlation 0.417, p = 2.6×10−6). In addition, interactions between the QTL and the genetic background likely contribute to context dependency [37].
We asked if the QTL mapping had resulted in any functional enrichment in the underlying genes. While genes within caffeine resistance QTL were only marginally enriched for the GO-term ‘Response to Stimulus’ (p = 0.077), genes found under salt resistance QTL were significantly enriched for ‘Cellular Response to Stimulus’ (p = 0.006) and ‘Kinase Activity’ (p = 0.002) (S5 Table). This suggests that our mapping experiments point towards genetic variants that impact cellular signaling pathways.
Several of the identified QTL tower among the genetic architectures of the two resistance traits (Fig. 3A & B, S4 Table). Tolerance of high salt concentrations is dominated by variation at the ENA locus on chromosome 4, which we detected in eleven out of twelve crosses. The ENA genes encode sodium efflux pumps that, in the presence of high salt, are transcriptionally activated by the HOG pathway [38]. European S. cerevisiae isolates carry an ENA variant that resulted from an introgression from the yeast species S. paradoxus and exists in varying copy numbers, referred to as ENA1 through ENA5. In contrast, non-European strains carry a single copy of the ancestral ENA6 gene [39]. Both the introgressed variant and its copy number have been shown to be associated with high salt tolerance [10], [40]. We determined the copy numbers of the two ENA variants within our parent strains and found that strains with high salt resistance indeed carried multiple copies of the introgressed ENA gene (S6 Table). Interestingly, the identified ENA QTL reflected differences beyond the two divergent ENA variants and copy number. We identified ENA QTL in crosses between parents with equal copy numbers of the introgressed variant (Cross 9), as well as between non-European isolates carrying alleles of the ENA6 gene that differed from each other by three missense mutations (Cross 12). Our findings corroborate the importance of the ENA locus in shaping the genetic architecture of salt resistance [10], [40] and further illustrate the multi-allelic nature of this locus.
The hallmarks of S. cerevisiae caffeine resistance architecture are the QTL containing TOR1 and TOR2 on chromosomes 10 and 11, respectively (Fig. 3B). In contrast to other eukaryotes, yeast carries two copies of TOR. As in higher organisms, TORC complexes respond to nutrient signals to regulate a plethora of cellular processes through two functionally distinct complexes. In yeast, Tor1p is specific to TORC1 complexes, while Tor2p is able to participate in both TORC1 and TORC2 complexes [41]. Caffeine inhibits Tor1p function [42], leading to the activation of the CWI pathway [43]. We identified a TOR1 QTL in eight of the twelve crosses, illustrating the large impact of TOR1 on the genetic architecture of caffeine resistance. TOR2 QTL were found in five crosses, and their identification likely speaks to Tor2p's ability to functionally compensate for Tor1p in TORC1 complexes. However, the QTL we detected as encompassing TOR genes were not mutually exclusive. In the four cases where we found both TOR1 and TOR2 QTL within the same cross, resistance alleles originated from the same or different parent strains an equal number of times. Furthermore, TOR negatively regulates transcription factors that are activated by the HOG pathway in response to stress (reviewed in [11]). Indeed, we uncovered salt resistance QTL containing TOR2 in three crosses (S4 Table).
In addition to these QTL shared among many of the Round-robin crosses, several other QTL had large effects in a smaller number of crosses and contain genes with well-established roles in cellular signaling. A caffeine resistance QTL shared between crosses 2 and 3 encompasses MSS4, which encodes the essential Phosphatidylinositol-4-phosphate 5-kinase required for activation of the CWI MAPK pathway [44]. Two additional large effect QTL highlight the interconnectivity of the cellular signaling network. These QTL on chromosomes 10 and 15 were pleiotropic and strongly influenced growth in the presence of both high salt and caffeine. The QTL on chromosome 10 contains CYR1, the adenylate cyclase essential for cAMP production and as such cAMP-PKA signaling [45]. Similar to TOR, the cAMP-PKA pathway responds to nutrient signals to modulate a myriad of cellular processes, including the negative regulation of stress response transcription factors (reviewed in [11]). The QTL on chromosome 15, on the other hand, includes a gene that represents a functional counterpart to this aspect of cAMP-PKA signaling: WHI2 encodes a phosphatase required for activation of the general stress response [46].
While trait variation among S. cerevisiae isolates is great in comparison to related fungal species [10], the evidence for adaptive alleles is scant [9], [47]. Rather than being strongly shaped by their source environment, yeast phenotypes have been proposed to be a result of population history [10]. The ENA locus illustrates this pattern. The introgression of the S. paradoxus ENA variant is widespread in European isolates regardless of their source environment. Our parent strains included European clinical and vineyard isolates, yet these ecological niches had no bearing on the ENA QTL we identified in crosses involving these strains. We wondered if alleles, denoted by the strain they derived from, tracked with source environment for any of the other QTL. We examined QTL detected in non-overlapping crosses (S7 Table), which excludes QTL likely due to single instances of loss- or gain-of-function mutations, and found two QTL where the grouping of alleles was unlikely to be a result of shared population history. Caffeine resistance alleles that gave rise to TOR2 QTL belonged to wine and clinical strains from a diverse range of geographic sources, while the corresponding sensitive alleles stemmed from natural isolates, such as oak strains. A salt resistance QTL on chromosome 15, containing YGK3, which encodes a stress response controlling kinase, exhibited a similar pattern: a European and an American clinical strain contributed resistance alleles, while diverse natural isolates supplied sensitivity alleles. Further studies are needed to clarify whether these two cases are coincidental or whether they represent cases of environmental adaptation.
Yeast mating, akin to the stress responses examined above, is regulated by a MAPK signaling pathway. In addition to providing us with independent selection experiments, our ability to isolate mapping populations of either mating type allowed us to search for variants that influence growth through a genetic interaction with the mating type locus. Fitness differences between the two mating types have been reported [48], [49], and we have shown that the MAT locus influences the differential expression of numerous genes [13]. In order to detect MAT-dependent growth QTL, we directly compared the allele frequencies resulting from corresponding MATa and MATα selection experiments.
For the round-robin crosses, we first compared the permissive YPD selections for each mating type. Among the twelve crosses, we identified six significant QTL (S8 Table). One of these QTL appeared in two non-neighboring crosses and encompassed the mating pathway regulator GPA1 located on chromosome 8. Polymorphisms within Gpa1, which negatively regulates the mating pathway, can impair its function and hence result in costly activation of the pathway in the absence of mating pheromone [50]. Importantly, the GPA1 allele selected for in one of our QTL experiments (Fig. 5) contains a known fitness-associated variant [50], [51]. We previously uncovered a genetic interaction between the MAT locus and GPA1 that affects the expression of other mating pathway genes [13]. While this interaction did not result in a growth QTL in the BY/RM cross, our identification of MAT-dependent GPA1 QTL suggests that this interaction can have an effect on growth in other genetic backgrounds. In addition to the GPA1 QTL, we identified a QTL containing SST2, which encodes the GTPase-activating protein for Gpa1 [52].
Next, we asked if mating type had any influence on the allele frequency distributions arising under the two MAPK-inducing stress conditions. We found 19 instances of MAT-dependent stress resistance QTL (S8 Table). While it is difficult to determine whether these QTL represent an instance of MAT-dependence or failure to replicate a weak QTL, 14 of these QTL were identified in independent experiments using different dosages of the same stress condition or in both stress conditions. Among the latter group, we found that the GPA1 and SST2 mating type dependent QTL persisted in the presence of either stress condition. The mating and HOG MAPK pathways share the MAPKKK Ste11, and several studies have investigated how signaling specificity is achieved despite this shared signaling component [53], [54]. In particular, faithful signaling along either pathway could be maintained by either mutual inhibition [53] or pathway insulation [54]. Our results suggest that the baseline fitness cost of the mating pathway [50] remains unaltered in the face of stressors that activate other MAPK cascades, which is consistent with the insulation of the mating MAPK from other MAPK signaling cascades [54].
The Round-robin design enables a strategy to identify candidate causative variants that is not available using pair-wise crosses. Specifically, the segregation of individual variants in relation to the presence or absence of QTL in multi-parental crosses can be leveraged to reduce the search space of candidate causative variants [55], [56].
Many sequence variants occur in several of the Round-robin crosses and, assuming additive effects independent of genetic background, should have similar consequences in these different crosses. We generated de novo assemblies of the parent strain genomes [57] and cataloged the non-synonymous coding variants in each QTL. We determined whether a given variant segregated within each cross by examining if the respective parent strains carried different alleles of the variant. For each QTL, we then assessed to what extent the segregation of a particular variant in each cross was associated with the detection of the QTL in the crosses (S4 Table). We used the maximum LOD scores within a particular QTL interval for each cross as a quantitative measure of QTL detection. This association analysis resulted in a 17-fold reduction in the number of variants to be considered as potentially causative for each QTL. We confirmed causality of individual sequence variants in two QTL to illustrate how this approach can aid in the identification of causal variants.
We first focused on the QTL on chromosome 15 that contained the stress response regulator WHI2 [46] and strongly influenced the resistance to both salt and caffeine in crosses 7 and 8 (Fig. 6A). Both crosses shared CLIB219 as a parent strain, and in both sets of experiments, the CLIB219 allele was strongly selected against, indicating that it leads to stress sensitivity. As this QTL did not appear in this form in other crosses besides the two involving CLIB219, the causative variant likely was private to CLIB219. Four variants out of 64 variants in the region were private to CLIB219 (Fig. 6B). One of the variants was a frame-shift mutation in WHI2. Allele replacements of the CLIB219 allele of WHI2 resulted in reduced growth in the presence of high salt and caffeine and thus confirmed that it was indeed the causal variant (Fig. 6C).
Next, we turned to a QTL with a substantially more complex allelic pattern. The TOR1-containing caffeine resistance QTL on chromosome 8 was identified in eight crosses. Specific alleles were selected for or against depending on the particular cross, which was a clear indication that multiple alleles were present at this QTL (Fig. 6D). While no single variant perfectly recapitulated the pattern of LOD scores (Fig. 6E), two SNPs in TOR1 and NTA1 scored the highest. Closer examination of the segregation pattern of these SNPs revealed that they were common to two caffeine sensitive strains, M22 and YJM981. Allele replacements of the TOR13875A variant in two strain backgrounds confirmed that it contributes to caffeine sensitivity (Fig. 6F). These results demonstrate how the Round-robin crossing scheme can be leveraged to focus the list of candidate causative variants.
Here, we have demonstrated the ability of a novel QTL mapping approach to query a large proportion of S. cerevisiae genetic diversity. We employed this approach to examine the genetic architectures of three MAPK-dependent traits: resistance to high salt or caffeine and growth differences due to mating type. The relationship of the mapped genetic variants to MAPK signaling pathways ranged from upstream modulators to core regulators and downstream targets.
Our modified X-QTL approach makes it possible to rapidly generate large mapping populations without extensive strain engineering. Quantitative trait mapping studies in yeast have uncovered the genetic architectures of traits central to biology, such as gene expression and protein-level variation [14], [21], [58], as well as traits of medical and industrial importance [15], [59]. Yet, these studies have only scratched the surface of the phenotypic variation of S. cerevisiae strains [10], [22], a space that is sure to expand with the continued discovery of additional yeast isolates [60], [61]. Our method promises to be of great utility in future yeast QTL mapping studies. Here, we have chiefly used our approach to isolate large pools of segregants for BSA-based QTL mapping, but we also illustrate its utility in isolating individual segregants that can be used in traditional linkage-based QTL mapping studies. In contrast to previous methods [20], [62], our method can be employed in nearly any strain background with minimal strain engineering and enables positive selection for haploids of either mating type. Furthermore, the fluorescent mating type markers are amenable to combination with fluorescent readouts of cell physiology or gene expression [20], [21]. Such reporter combinations will make it feasible to isolate mapping populations and select for phenotypes of interest simultaneously.
We have genetically dissected stress resistances mediated by two MAPK pathways, as well as interrogated the genetic contribution to the baseline cost of the mating pathway. We studied twelve diverse yeast strains that represent a large proportion of the genetic and phenotypic diversity of S. cerevisiae. Thus the QTL identified capture much of the species-wide genetic variation that influences the phenotypic consequences of these MAPK pathways. We identified QTL that encompass genes connected to MAPK signaling at multiple levels (S4 Table). Several QTL contained genes whose protein products are responsible for the first step of the MAPK signaling cascades: osmosensors (MSB2 & SHO1) and CWI sensors (WSC3 & MTL1) responsible for detecting the respective perturbations. Moving along the signaling cascades, we found several key regulators such as MSS4 and GPA1. Furthermore, one of the caffeine tolerance QTL contained SLT2, which encodes the MAPK at the center of the CWI response. Arriving at the output of the signaling pathways, we identified several genes transcriptionally activated in response to stress. In addition to the ENA genes, we found GPG1, which is activated by the HOG pathway to increase glycerol synthesis to counteract changes in osmolarity [63]. Finally, although the resistance phenotypes of the parent strains were not correlated, thirteen QTL occurred under both conditions. The overlap between QTL illustrates the crosstalk between the two stress responsive pathways [64], [65].
In addition to genes directly associated with MAPK pathways, we found genes that highlight the interconnectivity of the cellular signaling network. Variants in CYR1 and TOR2, core members of the cAMP-PKA and TOR pathways respectively, were tied to both stress responses. These pathways sense the internal availability of nutrients and repress the general stress response [11]. The TOR pathway in particular regulates a myriad of functions that mediate cell growth [66]. In contrast to higher eukaryotes, yeast carries two copies of the TOR gene. Our study links numerous alleles of TOR1 and TOR2 to differential growth under stress conditions. The two Tor proteins are redundant in regards to their participation in TORC1 complexes, but only Tor2p can partake in TORC2 complexes [41]. Despite this difference, the two TOR genes are diverging at the same rate as the majority of duplicated gene pairs [67]. However, these studies were based on individual strains. Our QTL results point towards functional variation in both Tor isoforms and raise questions about how they jointly regulate cell growth in different yeast isolates. Are the roles of the two proteins in TORC1 complexes truly equivalent, or are the TOR alleles indicative of functional tradeoffs? Similarly, it bears determination whether stress sensitive alleles of the TOR2 QTL, which were found in natural rather than ‘domesticated’ isolates, are beneficial under conditions approximating their source environment.
The identification of WHI2 further echoes the interplay between internal nutrient and external stress sensing pathways. While WHI2 is required for the activation of stress response pathways [46], its loss leads to prolonged growth under nutrient limited conditions [68]. Interestingly, WHI2 loss-of-function alleles were found to be driver mutations in experimental evolution experiments [69], [70]. Variants that abrogate Whi2 function were also discovered as recurrent secondary site mutations in yeast strains with specific gene deletions [71]. As such, whi2 mutants not only illustrate how laboratory evolution experiments can recapitulate the biology of wild isolates, but also suggest that the concomitant loss of stress responsiveness is adaptive under permissive growth conditions [69], [70]. Indeed, the WHI2 allele responsible for stress sensitivity was advantageous for growth under the permissive YPD condition.
The QTL results provide insights into the types of genes that carry variants shaping MAPK-associated phenotypes of different yeast isolates. The core elements of the HOG and CWI signaling cascades (HOG: SSK1, SSK2, STE11, PBS2, and HOG1; CWI: RHO1, PKC1, BCK1, MKK1, MKK2, and SLT2) are highly conserved across species. Analysis of the strain sequences revealed that despite their species-level conservation, all of these core proteins, with the exception of Rho1p, contain instances of non-synonymous coding variants. Yet, we only identified QTL at SLT2, indicating that the coding variants in the other core genes may be of little functional consequence.
In contrast to the central cascade, proteins responsible for sensing perturbations in osmolarity or cell wall integrity diverge rapidly between yeast and related fungal species [12], [72]. Divergent sensor proteins in other species are thought to enable the response to different habitats. For instance, other fungi use the HOG pathway to respond to a broader range of stresses than S. cerevisiae (reviewed in [11]). Consequently, we wondered if variation in sensor genes underlies the observed stress resistance variation among different S. cerevisiae isolates. We found stress sensors at two QTL for both salt and caffeine resistance, but these QTL had low LOD scores indicative of small effect variants. This is likely explained by the large degree of redundancy among the sensor proteins [73], [74]. Osmosensors are encoded by a set of five genes, and CWI sensors are encoded by six genes. While sensor redundancy allows for species-level fine-tuning of the stress responses, any single change is unlikely to have large phenotypic consequences. The same holds true for the large number of transcriptional targets of the MAPK pathways, with the exception of special effectors such as the introgressed ENA variants and GPG1 [75]. In contrast, many genes associated with large effects, such as CYR1, GPA1, MSS4 and TOR1/2, have functions essential to the MAPK and other signaling pathways. Together, these findings illustrate how genetic variants shape the phenotypic space of MAPK stress resistance traits within S. cerevisiae. The global picture that emerges from our results is that genes thought to underlie phenotypic differences between fungal species due to their high levels of divergence contribute little to trait variation among S. cerevisiae isolates.
The QTL we identified fall into several different classes in regards to their appearance among the twelve round-robin crosses. 33 of the 60 grouped QTL were ‘context dependent’, meaning they were detected in a single cross. The preponderance of ‘context dependent’ QTL was also notable in studies involving smaller panels of pairwise crosses [23], [34]. Context-dependency has been attributed to interactions between QTL and the genetic background. In particular, large effect variants can mask the presence of small effect variants [35], [36]. For salt resistance, the negative correlation between parent strain phenotype differences and the number of QTL identified was largely driven by the large effect of the ENA locus. For example, Cross 7 between two sensitive parent strains lacking the ENA introgression permitted the detection of a large number of small effect QTL. Of the twelve QTL detected in exactly two crosses, seven appeared in crosses involving the same parent strain, and as such can be classified as ‘strain dependent’ [34]. This set of QTL likely results from single alleles private to the shared parent strain [23]. Beyond this simple case, deducing the number of underlying alleles and their frequency becomes difficult. For instance, of the eleven QTL appearing in three crosses, nine were found in a pair of neighboring crosses plus one additional cross. Several models could explain this pattern, and we attempted to use the segregation pattern of the sequence variants within each QTL to determine which alleles best explained the observed QTL pattern. While this approach resulted in a significant reduction in the number of variants to consider, more than twelve crosses will be necessary to distinguish between the possible allele patterns with certainty. Going forward, our fluorescent markers allow facile creation of segregant pools from many additional crosses, as well as large mapping panels of individual segregants [17]. In conjunction with new genome engineering technologies, these approaches will enable even more complete genetic dissection of MAPK signaling and other complex traits.
Plasmids with drug markers were constructed in two steps. We first generated pRS416 and pRS426 Gateway expression vectors suitable for one and three fragment Gateway cloning. As described previously, the pRS vectors were digested with Sma1 and the Gateway cassettes introduced by ligation [76]. The drug markers were then introduced from a set of plasmids kindly provided by Mikko Taipale and Susan Lindquist (Whitehead Institute, Cambridge, MA). These Gateway drug markers plasmids were digested with BglII and XhoI in order to create fragments with the MX drug resistance markers alone. PCR was used to generate fragments of the pRS416 and pRS426 Gateway vector backbones with Xho1 and BglII cut sites and without the URA3 marker. PCR products were gel extracted and digest with Xho1 and BglII. The vector backbone with the two different Gatewat cassettes and the three different drug resistance markers where then combined by ligation.
The STE2 and STE3 promoters were cloned upstream of the genes encoding the fluorophores mCitrine and mCherry using overlap extension PCR. Promoter fragments based on Tong et al. [24] were PCR amplified from genomic DNA, while fluorophores were amplified from plasmids kindly provided by Gregory Lang (Lewis-Sigler Institute, Princeton University, Princeton, NJ). The resulting fusions, as well as the CYC1 terminator, were cloned into different Gateway entry vectors to facilitate their final combination in the three fragment Gateway plasmids.
Plasmids created for this study are listed in S9 Table, while primers used are listed in S10 Table.
The fluorescent mating type markers were tested in BY strains kindly provided by the Botstein lab (BY 12045 MATa Δura3, BY 12046 MATα Δura3). The BY/RM diploid was published previously (yLK1993 [17]). Other strains were previously described in Schacherer et al. [27]. To facilitate the generation of diploids for the round-robin cross, the KanMX cassette used to delete HO was replaced by the HphMX cassette in MATx strains [77]. MATa and MATx parent strains were mated in YPD and subsequently streaked to YPD+G418+Hygromycin to select for diploids. Putative diploids were tested for their ability to sporulate and to determine their ploidy [78]. The p41Nat plasmid carrying the fluorescent mating type markers was introduced by standard yeast transformation [79].
We used the NatMX cassette from pUG74 [80] to delete URA3 gene in the haploid round-robin parent strains in order to make them amenable to allele replacement protocols. Attempts to make allele replacements according to Erdeniz et al. were unsuccessful [81]. We modified the approach described by Stuckey et al. to make allele replacements [82]. For the WHI2 allele replacements, the entire WHI2 ORF was first deleted using the pCORE construct [83] and then replaced by transformation with PCR-amplified WHI2 ORFs plus approximately 250 bases up- and downstream from the different strains. For the TOR1 allele replacements, we first integrated K. lactis URA3 right at position 3875 of TOR1. We then removed the K. lactis URA3 integration according to Storici et al. using complementary oligos spanning the integration site and carrying the two different TOR1 alleles (TOR1 3875G and TOR1 3875A). Primers used in allele replacement experiments are listed in S10 Table.
We phenotyped a set of strains previously described by Schacherer et al. [27], as well as sets of random segregants stemming from the round-robin crosses. The strains were grown overnight at 30°C in 96-well plates containing YPD media according to standard yeast culture protocols [78]. The strains were then pinned in quadruplicate onto agar plates using a Singer RoToR HDA. Strains were phenotyped on YPD plates (the permissive control condition), as well as YPD plates containing 0.5 M sodium chloride, 1 M sodium chloride, 15 mM caffeine or 20 mM caffeine. Both sodium chloride and caffeine were autoclaved along with the other ingredients needed to make standard YPD plates. After strain pinning, plates were grown at 30°C in large plastic containers to avoid excessive evaporation. YPD plates were scanned after two days of growth, while stress condition plates were scanned once growth reached a level equivalent to that observed on the YPD plates [17]. Images of the plates were processed and analyzed to extract strain growth information as previously described [17]. Growth under the stress conditions was adjusted for growth under the permissive stress condition by calculating the ratio of the two growth measurements. We used a modified version of the epistasis test described by Lynch and Walsh [32] as detailed in Brem and Kruglyak [31].
We modified a published random spore prep protocol [78] in order to enrich for spores prior to FACS. Diploids were grown in 3 ml YPD overnight at 30°C. Cultures were spun down and washed with water. Cell pellets were resuspended in 1 ml sporulation media containing 25 mg/L ClonNAT or 200 mg/ml G418 to maintain the selection for fluorescent mating type marker plasmids. Half of the cell mixture was then added to 2.5 ml drug containing sporulation media. Sporulation cultures were incubated for four to six days at room temperature depending on the sporulation efficiency of a particular cross. 250 µl of the sporulation cultures was spun down for one minute in an eppendorf tube. After removal of supernatant the cell pellets were resuspended in 100 µl water. 17 µl of the cell suspensions was then added to new eppendorf tubes containing 3 µl of β-glucoronidase. After a brief vortex tubes were incubated in an Eppendorf Thermomixer at 30°C for two hours while shaking at 900 rpm. Digestion and opening of ascii after this two hour incubation was confirmed by light microscopy. Digestions were quenched by the addition of 100 µl water. In order to break apart digested ascii, tubes were vortexed for two minutes and then spun down for one minute. The supernatant was removed, another 100 µl was added and the tubes were vortexed for another two minutes. As hydrophobic spores stick to the sides of the eppendorf tubes all liquid is removed after the second vortex. Tubes were washed three times by addition of 1 ml water and inverting the tubes fully three times. After the washes, 1 ml 0.01% NP-40 (igepal) was added added to the tubes. Tubes were the sonicated using a tip sonicator for 1 min at power level 2.
In order to isolate individual spores or to assess the amount of spores isolated, dilutions of the spore prep (for most crosses 25 µl spore mix in 250 µl H2O was used) on the plates containing 25 mg/L ClonNAT or 200 mg/ml G418. Individual haploids were picked after two days of growth at 30°C using a fluorescence-equipped dissection microscope. For cell sorting 250 µl of the samples was plated on four plates containing the appropriate drug to select for the marker plasmid.
Spore preps were harvested after 2 days of growth at 30°C. Cells were washed off plates using 5 ml water and collected in 50 ml conical tubes. Cells were sonicated for 1 min at power level 2 and then diluted to an OD of 0.4 for FACS in 1xPBS. We used a BD FACSVantage SE to sort 500,000 green and red fluorescent cells each into vials containing 2 ml YPD with 100 µg/ml ampicilin to prevent bacterial contamination during the cell sorting. The sorted cells were added to 10 ml YPD containing 100 µg/ml ampicilin in glass tubes and cultured for 6 hours at 30°C. Cultures were spun down for 5 minutes at 3,000 rpm and after removal of supernatant cells were resuspended in 950 µl water. 200 µl of cells suspension were then plated on 5 different conditions: YPD, YPD with 0.5 M NaCl, YPD with 1 M NaCl, YPD with 15 mM caffeine, and YPD with 20 mM caffeine. Plates were grown for 2 days at 30°C. Plates were made by adding the appropriate amount of NaCl or caffeine to the standard YPD recipe prior to autoclaving.
We assessed to what extent segregant populations grew under the selective conditions and proceeded with conditions that resulted in a reduced yet adequate amount of growth compared to the permissive YPD condition [20]. Cells were harvested by washing plates with 5 ml water and collected in 15 ml conical tubes. Cells were spun down and stored at −80°C after removal of supernatant. DNA was isolated from the harvested cells according to Lee et al. [84]. Cell pellets of approximately 50 µl were used for each DNA prep. DNA was isolated from parent strains in the same manner except that parent strains were grown overnight in 3 ml YPD at 30°C.
Sequencing libraries were generated using Epicentre Nextera DNA Sample Prep kits or Illumina Nextera Sample Prepartation kits as described previously [17]. Residual salts and other contaminants can cause Nextera library preps to fail. We cleaned the input DNA using the Zymo Research Genomic DNA Clean & Concentrator kits if the initial library prep failed. Sample libraries were indexed using primers designed according to Meyer and Kircher [85] and sequenced in groups of 25 per lane on an Illumina HiSeq sequencer.
We created a set of genetic markers for each cross by aligning parent strain sequences to the sacCer3 reference genome and we then used custom scripts written in the programming language R to identify SNPs that differed between any set of parent strains. We then determined allele frequencies at these variant sites. We subtracted the YPD allele frequencies at each variant position from those obtained under the different growth conditions to determine the allele frequency skews that are due to a specific condition rather than permissive growth. To determine allele frequency differences due to mating type, allele frequencies from MATα samples were subtracted from those of corresponding MATa samples.
We converted allele frequency skews into LOD scores using the MULTIPOOL method [33] with the following parameters: -t, -n 1000, -m contrast, -c 2200, -r 100. MULTIPOOL outputs were used to generate genome-wide LOD score plots [21]. LOD plots were generated in R. All the subsequent analyses were conducted using the statistical programming language R. We called QTL at positions that exceeded a LOD threshold of 5 and QTL regions using LOD drop intervals of 2, as described previously [21]. In order to combine QTL across replicate experiments we determined which QTL had overlapping mapping intervals. We then used the mean of replicated QTL intervals to identify QTL that overlapped between different round-robin crosses. We combined QTL that were detected at different concentrations of the same conditions (0.5 M NaCl & 1 M NaCl, 15 mM Caffeine & 20 mM Caffeine) to provide a conservative estimate of the QTL LOD scores and regions.
We used a set of replicate BY/RM mapping experiments to determine the false discovery rate and reproducibility of our mapping approach. We generated allele frequency measurements for four independent experiments using two independent transformants of the BY/RM diploid with the mating type marker plasmid. We isolated mapping populations of either mating type and used the same selection conditions as for the round-robin crosses (YPD, +0.5 M NaCl, +1 M NaCl, +15 mM Caffeine, +20 mM Caffeine). The YPD 20 mM Caffeine condition resulted in very low growth of BY/RM segregants and we did not proceed further with this condition. Given the two mating type mapping populations and four independent experiments, we generated eight selections for each condition. We compared the allele frequencies resulting from technical replicate experiments stemming from one transformant (16 comparisons) and found no allele frequency fluctuations that resulted in LOD scores higher than 2.62. Analysis of biological replicates, i.e. comparing experiments based on different transformants, resulted in six peaks with LOD scores higher than our threshold of LOD 5. Yet, five of these peaks were actually the same region identified multiple times in different replicate pairs suggesting that they were due to a genetic difference between the two BY/RM transformants and not due to random allele fluctuations. The remaining peak tagged the mating type locus on chromosome 3 pointing towards a difference in the mating type selection rather than a random allele frequency fluctuation.
In order to asses the reproducibility of the mapping approach we determined condition-specific QTL for individual experiments and then asked to what extent QTL where found in pairs of replicates. To keep this assessment of reproducibility consistent with the round-robin mapping experiment, we did not ask whether QTL were identified multiple times across all replicates. Such a measure of reproducibility would be inflated in comparison to experiments with only two replicates, such as the experiments with the round-robin crosses. Rather we tested if QTL were reproducibly identified in designated pairs of experiments, either technical replicates performed with the same strain and the same condition or the experiments of opposite mating types of one replicate. At a LOD threshold of 5, 90.1% of QTL were reproducibly identified between technical replicates of the same mapping experiment. Reproducibility was 88.3% using mapping populations of opposite mating type. For the experiments with technical replicates, the directions of the allele effects of the QTL identified in only one experiment were consistent among replicates in 100% of the cases. This suggests that these QTL only narrowly missed the LOD threshold in replicate experiments. Indeed, their LOD scores were close to the threshold of LOD 5 (median LOD 5.9 vs. 9.84 for replicated QTL). For the round-robin experiments across all conditions, 74.4% of QTL (311 of 418 individually identified QTL, with one instance of one peak overlapping two in the second dataset, resulting in 155 jointly identified QTL) were identified in both of the mapping experiments. Since QTL that narrowly miss the LOD threshold are classified as not being reproducible, we asked to what extent allele frequency directions, in the control case BY versus RM, agreed over the identified region. We determined the mean allele frequencies for the region underlying a QTL and compared them across the pairs of replicate experiments. For the BY/RM experiments, the sign of the allele frequencies agreed for all eleven QTL that had missed replication. In the case of the round-robin crosses, allele frequency directions agreed for 93.5% of non-replicated QTL. This result was robust to permutation of the allele frequencies (p<0.001).
We used VEP [86] to identify genes with non-synonymous coding sequence variants within the combined QTL intervals. GO analyses were performed using the tools provided by the Saccharomyces Genome Database [87].
We used Cortex [57] to generate reference-assisted genome assemblies of the round-robin parent strains. We then used VCFtools [88] to subset the VCF file generated by Cortex to the individual regions of the QTL. We used VEP [86] to identify non-synonymous coding sequence variants within QTL intervals. We then generated a table for each QTL in the statistical programming language R that contained the replicate-averaged maximum LOD scores for each cross for the given QTL region and notations how each variant segregated among the crosses. For each variant, we then calculated the squared correlation between its segregation pattern and the pattern of LOD scores. We determined the distribution of for all the variants of a given region and used a p value of 0.01 to call significant variants. The coefficients of determination were visualized along QTL intervals using the UCSC Genome Browser [89].
|
10.1371/journal.pntd.0000194 | Albendazole versus Praziquantel in the Treatment of Neurocysticercosis: A Meta-analysis of Comparative Trials | Neurocysticercosis, infection of the brain with larvae of Taenia solium (pork tapeworm), is one of several forms of human cysticercosis caused by this organism. We investigated the role of albendazole and praziquantel in the treatment of patients with parenchymal neurocysticercosis by performing a meta-analysis of comparative trials of their effectiveness and safety.
We performed a search in the PubMed database, Cochrane Database of Controlled Trials, and in references of relevant articles. Six studies were included in the meta-analysis. Albendazole was associated with better control of seizures than praziquantel in the pooled data analysis, when the generic inverse variance method was used to combine the incidence of seizure control in the included trials (patients without seizures/[patients×years at risk]) (156 patients in 4 studies, point effect estimate [incidence rate ratio] = 4.94, 95% confidence interval 2.45–9.98). In addition, albendazole was associated with better effectiveness than praziquantel in the total disappearance of cysts (335 patients in 6 studies, random effects model, OR = 2.30, 95% CI 1.06–5.00). There was no difference between albendazole and praziquantel in reduction of cysts, proportion of patients with adverse events, and development of intracranial hypertension due to the administered therapy.
A critical review of the available data from comparative trials suggests that albendazole is more effective than praziquantel regarding clinically important outcomes in patients with neurocysticercosis. Nevertheless, given the relative scarcity of trials, more comparative interventional studies—especially randomized controlled trials—are required to draw a safe conclusion about the best regimen for the treatment of patients with parenchymal neurocysticercosis.
| Neurocysticercosis is a parasitic disease caused by the pork tapeworm, Taenia solium, when the larval form of the parasite lodges in the central nervous system. This disease is most commonly found among members of agricultural societies with poor sanitary conditions and economies based on breeding livestock (especially pigs) with low hygiene standards. It is a disease with long history in humans, and the usual therapeutic intervention was surgery until the development of antiparasitic cysticidal agents, the most common being praziquantel and albendazole. T. solium infection can take many different forms in humans, but we concentrated on parenchymal neurocysticercosis with viable cysts. A consensus statement by a panel of experts on the subject supports the use of antiparasitic treatment, but does not indicate either albendazole or praziquantel as the drug of choice for this type of neurocysticercosis, because data from single relevant clinical trials are not conclusive. We conducted a meta-analysis to further evaluate the comparative effectiveness and safety of albendazole and praziquantel for this particular type of neurocysticercosis. The outcomes of our meta-analysis suggest that albendazole is more effective than praziquantel in controlling seizures in affected patients and in leading to the total disappearance of cysts and subsequently cure of patients with neurocysticercosis.
| Neurocysticercosis is a parasitic disease caused by the larval form of Taenia solium, known as pork tapeworm, when the larvae lodge in the central nervous system (CNS). It happens when human ingests the eggs, acting as the intermediate host in the life cycle of T. solium. The eggs hatch in the intestine and the embrya penetrate the intestinal wall and are distributed via the blood, anchoring in the CNS as a larval form of the parasite [1]. With T. solium parasitosis, both self-reinfection and infection of household members are common.
Neurocysticercosis is mosst commonly found among members of agricultural societies with poor sanitary conditions and economies based on breeding livestock, especially pigs, with low hygiene standards [2]. However, it has also started to emerge in developed countries, as a result of immigration from endemic to nonendemic areas [3]. Its natural pool lies mainly in Latin America, sub-Saharan Africa, and Southeast Asia, and is an important cause of morbidity among local populations [2].
Neurocysticercosis is divided into four categories depending on the anatomical locus in which the larvae lodge—cerebral or parenchymal, subarachnoid or cisternal, intraventricular, and spinal [1]. The most common clinical sign of neurocysticercosis is epilepsy of any type, which is usually late-onset; this sign is typically found in parenchymal neurocysticercosis. Other common signs are focal neurological deficits, cerebellar or brainstem signs, signs of increased intracranial pressure, meningoencephalitic signs, dementia, or even death [4].
The standard therapeutic intervention was surgery until the development of cysticidal agents, the most common being praziquantel and albendazole [5]. Although there have been many clinical trials testing these drugs, controversy remains about their therapeutic value [5]. The reasons for this dispute include the severity of adverse effects, the actual reduction of cysts, and the subsequent control of seizures. This disagreement seems to have been resolved after the recent publication of a meta-analysis that shows the superiority of these agents compared to placebo [6].
We sought to investigate which of the two agents are preferable in the treatment of neurocysticercosis. Some studies have been published on this issue, although they mostly examine small numbers of patients. Specifically, we investigated the role of albendazole versus praziquantel in the treatment of patients with parenchymal neurocysticercosis by performing a meta-analysis of comparative trials [7] of their effectiveness and safety.
The studies for our meta-analysis were obtained from the PubMed database, Cochrane Database of Controlled Trials, and from references of relevant articles. Search terms included “albendazole”, “praziquantel”, “neurocysticercosis”, and “Taenia solium”. Although the search was performed without limitation on the language of publications, the evaluable studies were published in English, French, German, and Italian. There was no limitation on the year of publication.
Two independent reviewers (DKM and GP) performed the search and selected the studies that were relevant to the scope of our meta-analysis. Any discrepancy or disagreement between the reviewers was resolved by consensus in meetings involving all authors. A study was considered eligible if (1) it was a prospective trial, (2) it compared albendazole with praziquantel for the treatment of patients with neurocysticercosis, (3) it examined the partial or total disappearance of cysts and/or control of seizures, and (4) if it included patients infected with parasites in their cystic stage without perilesional inflammation. Studies using concomitant drugs such as corticosteroids, analgesics, and anticonvulsive drugs were not excluded.
The following data were extracted from each study: year of publication, study design, population of the study, therapeutic regimens used, concomitant drugs, number of patients, follow-up period, patients having control of seizures, proportion of cyst reduction, disappearance of cysts, total toxicity, and patients presenting intracranial hypertension as a side effect. A quality review of each randomized controlled trial (RCT) included in our analysis was performed by using the Jadad score, which examines whether there is randomization, blinding, and information on withdrawals in the study, and evaluates the appropriateness of randomization and blinding, if present. One point was awarded for the presence of each of the first 3 criteria, whereas the last 2 criteria could take the values of −1 (inappropriate), 0 (no data), and +1 (appropriate) [8],[9]. Thus, the maximum score for a study was 5, and a score more than 2 points denoted an adequate RCT according to the methodology. The reviewers calculated the score of each RCT independently. Any disagreement was resolved after consensus among all authors.
The primary outcome was the proportion of patients with controlled seizures. Secondary outcomes were the reduction of cysts in all of the examined patients, the proportion of patients with total disappearance of cysts, the proportion of patients with adverse events related with the administered antihelminthic drugs, and the proportion of patients with intracranial hypertension as a side effect caused by the administered drugs.
A patient was considered as having total control of seizures when there had been no seizures during the follow-up period. A patient was considered as having total disappearance of cysts when this outcome had been achieved after only one course of administered chemotherapy and without any surgical intervention at the follow-up CT scan, performed in a time frame of 3 to 6 months after the end of therapy. The reduction of cysts was defined as the proportion of the number of cysts that had resolved by the follow-up evaluation (numerator), which varied from 3 to 6 months post-therapy, divided by the number of cysts at baseline (denominator). Adverse events included any type of adverse event reported in the included studies.
Statistical analyses were performed using the “Review Manager 4.2” software and the SPSS 15.0 statistical software. The heterogeneity between studies was assessed by using the I2 test and χ2 test; for the χ2 test, p<0.10 was considered statistically significant in the analysis of heterogeneity [10]. Small-study bias was assessed by the funnel plot method [11]. Pooled odds ratios (ORs) and 95% confidence intervals (CIs) for all primary and secondary outcomes were calculated by using both the Mantel-Haenszel [12] fixed effect model and the DerSimonian-Laird random effects model [13]. For all analyses, results from the fixed effect model are presented only when there was no heterogeneity between studies; otherwise, results from the random effects model are presented. For the analyses of proportions of the reduction of cysts, we used a linear regression model in which the percentage of reduction of cysts for each treatment arm in the included studies was the dependent variable, and the administered drug was the independent variable. With this model, a beta (β) coefficient of the independent variable was calculated as well as the 95% confidence interval (CI) of the coefficient. For the analyses of seizure control for which the follow up period varied, we combined the logarithms of the rate ratios across the included trials (patients with outcome/[patients×years at risk]) using the generic inverse variance method.
Figure 1 is a flow diagram describing the process of study selection. We identified 103 potentially evaluable papers, 91 of which were excluded because they were reviews, case reports, letters or editorials, laboratory studies, small series of patients, retrospective studies, and meta-analyses that examined a different aspect of neurocysticercosis than the comparison between praziquantel and albendazole. Of the remaining 12 potentially evaluable papers, 2 studies were excluded because they included patients with neurocysticercosis that was not parenchymal, 1 because the majority of the enrolled patients had mixed living and calcified cysts, 1 because the enrolled patients were all put in the same group without providing separate data for each antiparasitic agent, and 2 because they were subsets of other larger trials. Thus, 6 trials were included in our meta-analysis [14]–[19].
The assessment of the evaluable studies according to the Jadad score was performed only for the 2 out of 6 studies [15],[19]. The rest of the studies were prospective [14], [16]–[18] but not RCTs. Thus, quality assessment of these trials using Jadad could not be done.
The studies differed in the administered dosing and duration of therapy for albendazole and praziquantel (Table 1). Most of the researchers administered 15 mg/kg/d of body weight of albendazole, but the duration of therapy varied from 8 days to a month [14]–[16],[18],[19]. In only one study albendazole was administered at a dosage of 20 mg/kg/d for 21 days [17]. We pooled these data, as the administration of albendazole for 7 days is as effective as for longer periods of therapy [20]. There was notable variation in the duration of praziquantel therapy, extending from a single day to 3 weeks. In all of the studies the dosage of praziquantel was 50 mg/kg/d, except one study in which praziquantel was administered at a dosage of 100 mg/kg in 3 divided doses at 2-hour intervals for a single day [14]. We pooled these data, as the administration of praziquantel for a single day is as effective as for longer periods of therapy [21]–[23].
Data on the complete control of seizures in patients with neurocysticercosis treated with albendazole or praziquantel were reported in 4 out of 6 studies (Table 2) [14]–[17]. One study reported a statistically significant effects in favor of albendazole, as reported in the crude data provided in the study [17]. To overcome the variation in the follow-up periods, we used the generic inverse variance method to combine the incidence of seizure control (patients without seizures/[patients×years at risk]) of the included trials (Table 2). Albendazole was associated with better control of seizures in comparison with praziquantel in the pooled data analysis (156 patients, random effects model [I2 = 51.2%], point effect estimate [incidence rate ratio] = 4.94 [seizure-free persons/person-years], 95% CI 2.45–9.98, Figure 2).
Data on the reduction of the total number of cysts from baseline to follow-up are reported in 5 out of 6 studies (Table 2) [14]–[17],[19]. A linear regression model of the proportion of reduction of cysts and the administration of albendazole or praziquantel yielded a beta coefficient (β) = 0.22 (standard error [SE] = 0.113) with 95% CI −0.05 to 0.48. The analysis included a total of 301 patients with 2565 cysts. Hence, there was no statistically significant difference in the proportion of the reduction of cysts between albendazole and praziquantel for the treatment of neurocysticercosis. In addition, in a sensitivity analysis excluding the data reported in the RCT by Sotelo et al [19] which comprised almost one half of the total number of cysts, there was no statistically significant difference in the proportion of the reduction of cysts between albendazole and praziquantel for the treatment of neurocysticercosis (β = 0.15 [SE = 0.18], 95% CI −0.30 to 0.59). The analysis included a total of 187 patients with 1342 cysts.
Data on the total disappearance of cysts are reported in all 6 studies (Table 2) [14]–[19]. Albendazole was associated with greater efficacy than praziquantel in the total disappearance of cysts (335 patients, random effects model (χ2-test p = 0.07, I2 = 50.3%), OR = 2.30, 95% CI 1.06–5.00, Figure 3). Since in the study by Cruz et al [18] it is not clear whether the patients with cystic lesions also had lesions involving other stages of the infection, we performed a sensitivity analysis without the aforementioned study, in which albendazole was more effective than praziquantel in inducing the total disappearance of cysts (301 patients, random effects model (χ2-test p = 0.05, I2 = 58.1%), OR = 2.62, 95% CI 1.09–6.32). We also performed a sensitivity analysis excluding data reported in the RCT by Sotelo et al [19], which included almost one-third of the total number of patients in this meta-analysis and showed statistical significance. There was no difference between the two regimens in inducing the total disappearance of cysts (221 patients, random effects model (χ2-test p = 0.08, I2 = 52.5%), OR = 2.20, 95% CI 0.79–6.13).
Data about mortality are reported in all 6 studies (Table 2) [14]–[19]. One death was reported in by Takayanagui et al [17] due to increased intracranial pressure. These data were not adequate to allow a meaningful analysis.
Data about patients with adverse events are reported in 5 out of 6 studies (Table 2) [14], [15], [17]–[19]. Albendazole and praziquantel did not differ in the proportion of patients with adverse events (388 patients, random effects model [χ2-test p = 0.06, I2 = 59.9%], OR = 0.67, 95% CI 0.26–1.69).
Data on intracranial hypertension developing as a consequence of the regimens administered are reported in 4 studies (Table 2) [14], [16]–[18]. There was no difference in the development of intracranial hypertension due to the administered therapy between albendazole and praziquantel (179 patients, fixed effect model [χ2-test p value = 0.58, I2 = 0%], OR = 0.31, 95% CI 0.05–2.09).
Neurocysticercosis is an endemic disease in many developing countries, and it may expand to the developed world, mainly as a result of immigration. Estimations report around 50 million new cases worldwide [24]. To our knowledge, until now the guidelines for the treatment of cysticercosis are the result of a consensus by a panel of experts in the subject [25]. Specifically, for viable parenchymal cysts the recommendations are based on evidence obtained from multiple case series with or without intervention, including dramatic results in uncontrolled experiments (level II-3 recommendation, which is considered a weak category of evidence), and on opinions of respected authorities, based on clinical experience, descriptive studies, and case reports or reports of expert committees (level III recommendation). Although these recommendations support the use of antiparasitic treatment, they do not point to either albendazole nor praziquantel as the drug of choice for this type of neurocysticercosis.
In a recent meta-analysis performed by Del Brutto et al. [6] it was suggested that, compared to placebo, cysticidal drug therapy results in better resolution of colloidal and vesicular cysticerci, lower risk for recurrence of seizures in patients with colloidal cysticerci, and a reduction in the rate of generalized seizures in patients with vesicular cysticerci. However, there has not yet been a meta-analysis comparing the effectiveness and safety of albendazole and praziquantel in patients with neurocysticercosis.
The outcomes in our meta-analysis suggest that albendazole is more effective than praziquantel in controlling seizures in the affected patients and in leading to the total disappearance of cysts and, subsequently, the cure of patients with neurocysticercosis. However, in the sensitivity analysis of the total disappearance of cysts, excluding the study by Sotelo et al [19], no significant difference was found between the drugs, although the odds ratio was rather similar to the analysis that included the study by Sotelo et al. [19]. This loss of statistical significance can be explained by the loss of power in the sensitivity analysis due to exclusion of the aforementioned study. Regarding other outcomes, there have been no statistically significant differences between albendazole and praziquantel in reduction of total number of cysts, mortality, total adverse events, and development of intracranial hypertension due to the administered therapeutic agents. Control of seizures and total disappearance of cysts were chosen as outcomes in our meta-analysis, because they are easily defined and quantitatively measured. In addition, new-onset seizures are among the most common symptoms that lead patients to seek medical care, and their resolution is one of the major goals of therapy.
In the analyses of outcomes we did not perform sensitivity analyses that excluded the study by Medina et al [16], in which patients did not receive corticosteroids. Since it is the only study with this characteristic, one may suggest that it could cause bias. It might be speculated that the absence of corticosteroids could interfere with the kinetics of the administered antihelminthics, and cause an increase in the rate of the adverse events. However, all the outcomes included in this study did not differ from the results of the other trials; adverse events are not reported in this study.
The reduced effectiveness of praziquantel could be explained by the interaction between praziquantel and corticosteroids, which results in decreased serum concentration of praziquantel [26]. Also, praziquantel interacts with anti-epileptic drugs [27],[28], thus altering its bioavailability. In contrast, corticosteroids interact with albendazole by decreasing the rate of elimination of albendazole sulfoxide, which is the active metabolite of albendazole, thus increasing serum concentrations of albendazole sulfoxide [29],[30].
Often, the first few days after the administration of antiparasitic agents to patients with neurocysticercosis there is a recrudescence of neurological symptoms, most importantly decompensation of intracranial pressure and the onset of seizures or worsening of pre-existing ones, owing to peri-lesional inflammation due to degeneration of the parasite; this condition can be life-threatening. The severity of inflammation is proportional to the parasitic burden, resulting in more severe manifestations in individuals with greater cyst loads [31]. A common approach to ameliorating this problem is the concomitant administration of corticosteroids to reduce edema, the inflammatory response, and intracranial hypertension [32]. Special attention should be paid to patients with high cyst loads, to whom the administered antiparasitic treatment causes an abrupt degeneration of cysts that may lead to severe inflammation and seizures [5]. In such cases corticosteroids should be administered before the antiparasitic agents. The single death reported in the study by Takayanagui et al [17] (the only death among patients of all trials included in this meta-analysis) was the result of increased intracranial pressure, which, however, pre-existed at the beginning of the trial. In 5 out of 6 studies included in our meta-analysis, corticosteroids were administered to patients [14], [15], [17]–[19]. Only in the study by Medina et al [16] were corticosteroids not administered; adverse events were not reported in this study.
It is believed by several experts that many cysts degenerate spontaneously over time, which may lead to the conclusion that the results of the evaluable studies may be biased [33]. Since it is not clear up to what extent this opinion is true, we analyzed studies that included patients with cystic lesions without perilesional enhancements or other evidence of surrounding inflammation, as evidence of a possible degenerative process, to rule out such a possibility. Antihelminthic drugs are effective against viable cysts, but not on remnants, granulomas, and calcifications of dead cysts. Thus, both outcomes we chose to study—the total disappearance of cysts and reduction of cysts—are useful indicators of the effectiveness of the administered therapy, because they estimate the effectiveness of the administered agents for lesions on which the agents are active.
There are some limitations in our meta-analysis that should be considered. First, one may claim that the number of the studies and the number of patients are too small to allow a definitive conclusion regarding the results of the compared therapies. This small sample size is important because it leads to large confidence intervals. In addition, publication bias cannot be appropriately assessed in a small set of studies. Also, among the studies selected there are only 2 RCTs [15],[19] in a total of 6 comparative trials, which prevents us from applying the usually applied methodology in obtaining an overall quality assessment of the included studies [8].
Second, there are discrepancies in the administered dosage and duration of therapy with the 2 antiparasitic agents used. Although there have been several studies aiming to establish an optimal dosage and duration of therapy, these important therapeutic parameters have not been standardized yet. We pooled all of the available data, since the dosage and the duration of therapy used in the trials included in this meta-analysis are generally accepted alternatives by the medical community.
Furthermore, there were differences in the length of follow-up for the control of seizures between the studies that varied from 6 to 24 months. This fact may give rise to methodological issues regarding the validity of combining these studies without considering the duration of follow-up. Thus, we performed an analysis using the generic inverse variance method combining the incidence of seizure control in the included trials (patients without seizures/[patients×years at risk]), in which the effect of different follow-up time is included. However, it should be noted that the caveat in this methodology is the assumption that the risk for seizures is constant, which is not proven. Despite the aforementioned limitations, the contribution of the meta-analysis in the literature sheds light in the subject given the scarcity of data.
In summary, neurocysticercosis is a disease with a long history in humans and with many different stages. We concentrated on parenchymal neurocysticercosis with viable cysts. The recommendations suggest the administration of antiparasitic treatment with concomitant use of steroids. This meta-analysis sought to provide more accurate estimates of the comparative effectiveness and safety of albendazole and praziquantel for this common parasitic infection. Nevertheless, more studies, especially randomized controlled trials, with homogeneous regimens and long follow-up periods, are required to draw a clear conclusion about the best regimen for the treatment of patients with parenchymal neurocysticercosis. |
10.1371/journal.pbio.1001444 | Cell Division in Apicomplexan Parasites Is Organized by a Homolog of the Striated Rootlet Fiber of Algal Flagella | Apicomplexa are intracellular parasites that cause important human diseases including malaria and toxoplasmosis. During host cell infection new parasites are formed through a budding process that parcels out nuclei and organelles into multiple daughters. Budding is remarkably flexible in output and can produce two to thousands of progeny cells. How genomes and daughters are counted and coordinated is unknown. Apicomplexa evolved from single celled flagellated algae, but with the exception of the gametes, lack flagella. Here we demonstrate that a structure that in the algal ancestor served as the rootlet of the flagellar basal bodies is required for parasite cell division. Parasite striated fiber assemblins (SFA) polymerize into a dynamic fiber that emerges from the centrosomes immediately after their duplication. The fiber grows in a polarized fashion and daughter cells form at its distal tip. As the daughter cell is further elaborated it remains physically tethered at its apical end, the conoid and polar ring. Genetic experiments in Toxoplasma gondii demonstrate two essential components of the fiber, TgSFA2 and 3. In the absence of either of these proteins cytokinesis is blocked at its earliest point, the initiation of the daughter microtubule organizing center (MTOC). Mitosis remains unimpeded and mutant cells accumulate numerous nuclei but fail to form daughter cells. The SFA fiber provides a robust spatial and temporal organizer of parasite cell division, a process that appears hard-wired to the centrosome by multiple tethers. Our findings have broader evolutionary implications. We propose that Apicomplexa abandoned flagella for most stages yet retained the organizing principle of the flagellar MTOC. Instead of ensuring appropriate numbers of flagella, the system now positions the apical invasion complexes. This suggests that elements of the invasion apparatus may be derived from flagella or flagellum associated structures.
| Malaria, toxoplasmosis, and related diseases are caused by infection with unicellular parasites called Apicomplexa. Their name refers to the elaborate invasion machinery that occupies the apical end of the parasite cell. This apparatus allows the parasite to force its way into the cells of its host, and to deliver factors that will manipulate host cell structure, gene expression, and metabolism. Once in the host cell the parasite will begin to grow. The parasite replicates its genome and organelles numerous times and then loads these various elements into numerous daughter cells that will further spread the infection. Here we report a fiber that coordinates the daughter cell budding process. The fiber links the centrosome, which controls the mitotic spindle, and the genome with the microtubule organizing center of the budding daughter. Parasite mutants lacking the proteins that build the fiber fail to form daughter cells at the earliest step. The fiber and its components are remarkably similar to fibers that coordinate flagella in algae. While Apicomplexa are not flagellated (with the exception of certain gamete stages) they evolved from flagellated algae. We propose that elements of the invasion apparatus evolved from the flagellum or flagellum associated structures.
| Apicomplexa are protozoan parasites responsible for numerous human and veterinary diseases. Human pathogens in this phylum include Plasmodium, the causative agent of malaria, Toxoplasma, an opportunistic pathogen that causes encephalitis in immunocompromised individuals and congenital toxoplasmosis, and Cryptosporidium one of the most important causes of severe early childhood diarrhea around the world. Apicomplexa are obligate intracellular parasites that follow a stereotypical propagation cycle. A motile zoite stage seeks out and invades a suitable host cell and in this process establishes a novel compartment, the parasitophorous vacuole, that houses the parasite during its intracellular development [1]. Parasites replicate and ultimately produce a new generation of zoites that destroys the host cell upon egress and fan out to infect new cells. Apicomplexans have adapted to tissue and host cell niches as varied as red blood cells, intestinal epithelial cells, macrophages and lymphocytes, or neurons.
The budding mechanism used by apicomplexans appears to be the key to their ability to scale their reproductive output to the size and biology of the specific host cell [2]. In this process, many species including the malaria parasite, deviate from the conventional cell cycle and pass through DNA synthesis and nuclear mitosis numerous times amassing a large number of genomes. Coinciding with the last round of mitosis, daughter buds are assembled and each nucleus is packaged into a new zoite. It is not understood how the parasites match the number of nuclear genomes with emergent daughter buds and how buds are placed and assembled correctly. The bud is scaffolded by microtubules that emanate from a newly formed apical microtubule organizing center (MTOC) (shown in blue in Figure 1A for T. gondii) [3]. These microtubules anchor the inner membrane complex (IMC, purple), an assemblage of membranous and cytoskeletal elements that establishes cell shape and is critical to the parasite's gliding motility [4]. The MTOC is thought to be a ring structure and can be further elaborated by additional elements including the conoid [3],[5],[6]. The MTOC also organizes the specialized apical secretory machinery that delivers proteins for host cell invasion and modification. Secretion of these organelles occurs at the extreme apex and through the ring [7],[8].
The centrosome has been demonstrated to organize parasite chromosomes and some organelles. Interestingly, both in T. gondii and P. falciparum chromosomal centromeres are constantly tethered to the centrosome [9],[10]. A similar physical association with the centrosome has been described for the apicoplast and the Golgi [11]–[13]. We hypothesized that a second tether linking the centrosome to the daughter bud MTOC and the associated invasion machine could provide a robust mechanism for cell assembly. In this study we use T. gondii, which divides using the simplest internal budding process in the phylum known as endodyogeny [14] as a model. We find that in T. gondii striated fiber assemblin (SFA) proteins, whose orthologs are found in the rootlet associated with flagellar basal bodies of single celled algae, assemble into a highly dynamic fiber during cell division. The SFA fiber links the centrosome and daughter MTOC, and ablation of SFA by conditional knock out results in multinucleated cells that fail to initiate the formation of daughter cells.
During cell division T. gondii assembles two daughter cells with a complex microtubular cytoskeleton and secretory apparatus within the mother cell (Figure 1A). This process has to be coordinated with mitosis and organelle segregation to ensure that the emerging daughter cells not only are competent to invade new host cells, but also carry the genetic and metabolic machinery required to propagate. It is not well understood how each daughter cell inherits a complete set of essential organelles. SFA is the main component of striated rootlets associated with basal bodies in green algae [15]–[17]. In previous work, Lechtreck and colleagues identified genes encoding homologs of SFA in Apicomplexa, including T. gondii [18]. This finding was surprising because, with the exception of the male gamete, Apicomplexa lack flagella. Nonetheless, antibodies raised against SFA from the green alga Spermatozopsis similis revealed a spot in proximity of the centrosome in T. gondii [18]. Transcription of the T. gondii SFA genes is cell cycle dependent with peak expression coinciding with DNA synthesis and mitosis (Figure 1B) [19],[20]. We therefore hypothesized that SFAs may function during division of apicomplexan parasites.
To define the function of SFA proteins in T. gondii, we first determined their localization. We focused on TgSFA2 and TgSFA3, two proteins that are expressed in the tachyzoite stage maintained in tissue culture. We engineered parasites in which the native TgSFA2 is tagged with a triple hemagglutinin (3×HA) at its C terminus. Southern blot analysis with a probe complementary to the 3′ end of TgSFA2 confirmed the insertion of the tag into the locus (Figure S1A). Western blot showed a single band of the mass predicted for TgSFA2-3HA (30 kDa) to be recognized by anti-HA antibodies (Figure 1D). To study TgSFA3, we expressed the gene in Escherichia coli, and raised antibodies against the recombinant protein. Independently, we also generated a strain in which TgSFA3 is endogenously tagged with yellow fluorescent protein (YFP) at its C terminus. A Western blot with anti-GFP antibodies showed a fusion protein of expected mass in the TgSFA3-YFP cell line, but not in the parental cell line (Figure 1D). This band is also recognized by the anti-SFA3 antibody, which in wild-type parasites recognizes native TgSFA3 (35 kDa) and the recombinant protein in bacterial lysates (rSFA3, Figure 1D). We next performed immunofluorescence assays (IFAs) on TgSFA2-HA parasites using HA and anti-SFA3 antibodies to detect both proteins simultaneously. TgSFA2 and TgSFA3 largely co-localize and both reveal two short fiber-like structures per parasite cell (Figure 1E).
While observing endogenously tagged TgSFA2 or the labeling by the TgSFA3 antibody by immunofluorescence we noticed that only a fraction of the parasites showed staining. At a given time 24% of the parasites express TgSFA2, while 76% do not (Figure 1C). These percentages closely match those previously reported for interphase and dividing parasites in asynchronous populations of T. gondii [21]. Next, we co-stained cells for TgSFA2-HA and with antibodies that detect markers of T. gondii cell cycle progression including centrin and IMC1. Centrin is a marker of the centrosome (Figure 2A–2C, red), IMC1 (blue) is part of the cytoskeleton of the IMC of the pellicle and outlines the mother cell as well as the forming daughters [22],[23]. Most parasites appear to be in G1 or early S phase, as determined by the presence of a single centrosome per parasite. Consistent with our prediction, no SFA labeling is discernible in these interphase cells. After centrosome duplication, SFA labeling becomes apparent as small punctuate structures that are very close to or overlap with centrosome labeling. In parasites that show anti-IMC1 stained daughter buds, TgSFA2 labels a long structure, extending away from the centrosome (Figure 2C). Notably, the SFA fiber is arched with a spiraling hook shape at its distal end. This pattern is also apparent in immuno-gold labeled cryosections of SFA2-HA parasites. Figure 2D shows the intra-nuclear mitotic spindle of T. gondii, an early step in the budding process [2],[14]. A short series of gold particles is visible at the bottom spindle pole. In parasites progressed further in division, gold particles form an arched line that climbs into the apical end of the daughter bud (Figure 2E). We conclude that SFA2 and 3 form a structure early in mitosis, that extends into a fiber during budding, but is absent in interphase.
To test whether SFA proteins have a functional role in parasite division we generated mutants in which their expression can be manipulated (Figure 3A). We constructed a strain, iΔSFA3, in which the native promoter of SFA3 is replaced by a tetracycline-regulatable promoter [24],[25]. The targeting construct was derived from a cosmid clone carrying the SFA3 locus by recombineering [24],[26]. A corresponding strain for SFA2 (iΔSFA2) was also made. In this case we used a plasmid that was designed to introduce both a regulatable promoter and a transactivator into the locus (both strategies and the specific parental strains used are described in detail in the Materials and Methods section and Figure S2). Disruptions of the targeted loci were confirmed by PCR (Figure S2B and S2D). Mutant parasites were cultured in the presence of anhydrotetracycline (ATc) and targeted protein levels were measured by Western blots using anti-SFA3 antibodies or by reverse transcription-PCR for the SFA2 mRNA. We noted a marked decrease in the levels of the targeted SFA proteins or mRNAs after 2 d of ATc treatment (Figure 3B). In both mutant strains parasite growth was severely impaired in the presence of ATc, as documented by their inability to form plaques in a fibroblast monolayer (Figure 3C, note that plaque formation of the parental strains is not affected by ATc). We conclude that SFA2 and SFA3 are non-redundant and both are required for parasite viability.
We hypothesized that the growth arrest of mutants deficient in SFAs is caused by defects in cell division. We cultured the mutant parasites for 24 and 48 h in ATc and stained using anti-IMC1 antibody to outline cells and DAPI to visualize nuclei. As shown in Figure 4A and 4B, for both mutants ATc treatment resulted in excessively large cells bearing multiple nuclei. We quantified this phenotype in the iΔSFA3 strain (Figure 4D), 59% of parasite cells are multinucleated (≥2 nuclei per cell) after 24 h of ATc treatment. Parasites with numerous apparently normal nuclei are also readily observed by electron microscopy (Figure 4C). To evaluate nuclear division and chromosome segregation more rigorously we stained mutants with a monoclonal antibody that we developed against the T. gondii centromeric histone variant 3 (CenH3). In Apicomplexans, centromeres are sequestered at the nuclear envelope in a centrosome-dependent manner and haploid and diploid nuclei have a single or duplicated CenH3 spot, respectively [9],[10]. We quantified the nuclear ploidy and note that ATc-treated mutants and controls are indistinguishable (Figure 5B). Moreover, we observed that every nucleus is associated with one or two centrosomes and that the centromere–centrosome association appears undisturbed (Figure 5A).
We next monitored daughter cell formation. Normally, IMC1 outlines the pellicle of both the mother and daughter cell (see Figure 2A and 2B). Strikingly, mutant parasites containing multiple nuclei showed aberrant or no daughter cells when stained with anti-IMC1 (Figure 4A). To test whether SFAs are required for the initiation or elaboration of daughters we stained mutants for the early marker of budding IMC subcompartment protein 1 (ISP1). ISP1 labels the apical cap of the IMC and can be detected prior to IMC1 [27]. In ATc-treated parasites ISP1 is visible in the mother cell pellicle but no daughter structures are discernible (Figure 5C). Taken together, these results suggest that loss of SFA proteins does not affect centrosome duplication and mitosis but severely impedes budding.
To better understand the mechanistic basis of the mutants' inability to bud, we monitored the localization of SFAs in relation to structures important for daughter cell assembly throughout division. Daughter cells are formed on a stereotypic scaffold of 22 sub-pellicular microtubules that arise from a circular apical organizing center, the apical polar ring [28]. In Toxoplasma this structure also includes the conoid, a motile structure thought to be involved in host cell invasion [5],[6]. We examined the localization of the SFA fiber relative to that of alpha-tubulin, Ring1 (RNG1), a component of the apical polar ring [29] and ISP1 [27]. We observed that the apical end of the SFA fiber consistently extends beyond the end of tubulin staining corresponding to the sub-pellicular microtubules (Figure 6A) and terminates at the apex of developing daughters, extending through the RNG1 staining and to the very tip of the ISP1 staining (Figure 6B and 6C). To unravel this complex architecture we imaged serial sections of dividing parasites by transmission electron microscopy (TEM). Figure 7A and 7B shows two consecutive sections through a daughter bud in which the conoid (Cn) and centrosome (Ce) are clearly identifiable. Spanning the area between them is an arching electron dense fiber (black arrowheads). The fiber terminates at a pair of microtubules that extend through the center of the conoid (arrow, [3]). Figure 7C–7E show a series of perpendicular sections through the conoid of a daughter cell. An electron dense fiber (arrowheads) curls up through the conoid coming into close contact with the apical rim of the structure and ending in the proximity of the central microtubule pair (arrow). A series of sections through a very early daughter bud shows the fiber to be already present at this stage (Figure 7F–7I). Note that it again makes contact with the apical rim of the conoid (Figure 7G, arrowhead) and that it emerges from in between the two centrioles of the centrosome (Figure 7F and 7G, arrow). Our light and electron microscopic observations suggest that the SFA fiber physically connects the centrosome to the tip of the forming daughter cell.
To visualize the dynamic development of the SFA fiber we inserted a YFP coding sequence into the genomic locus of SFA3 creating a C-terminal fusion protein. We time lapse imaged the SFA3-YFP strain and determined that SFA3 is visible for 2 h and 20 min (n>5), a time frame consistent with the duration of mitosis in T. gondii under imaging conditions [21]. Moreover, we observed that the SFA structure is dynamic and its morphology changes with time (Video S1). In order to have a spatial reference for the transition events of the SFA fiber we imaged SFA3-YFP in combination with Centrin1 fused to red fluorescent protein (RFP). This allowed us to concurrently monitor the position of the SFA fiber and the centrosome [30],[31]. In time-lapse imaging the YFP signal appears right on the centrosome (Figure 8A; Video S2). The fiber then elongates away from the centrosome. When the fiber reaches about half of its final length a “hook”-like shape at the tip away from the centrosome can be resolved (Figure 8A, 100′). The fiber reaches a maximum length of about 1 µm, at which point it appears to break close to its distal end (Figure 8A, 140′–160′). This leaves a small dot (presumably associated with the tip of the daughter cell). The longer centrosome associated portion shortens from the distal end and finally, SFA3-YFP is no longer detectable leaving only the centrosome visible (Figure 8A, 200′).
Polarized polymerization of subunits similar to microtubules or actin filaments could be a model for the growths of the SFA. To test this idea and to determine the direction of fiber extension we used fluorescence recovery after photo-bleaching. We chose SFA3-YFP parasites exhibiting two fibers of medium length (0.45 µm) and selectively bleached one of the two fibers using a diffraction limited 488 nm laser spot (Figure 8C). Figure 8B shows the target fiber prior to bleaching, note that the second fiber is not in the same focal plane as the target fiber and does not appear in this series of images. Figure 8C shows that after the laser pulse, the target fiber is no longer visible. We monitored the fluorescence of the bleached fiber after 1 h, and found reappearance of the YFP signal (Figure 8D). While the unbleached fiber had practically doubled in size to 0.85 µm, the YFP signal on the bleached fiber was only 0.35 µm (Figure 8D and 8E). For reference we also imaged the apicoplast labeled with ferredoxin/NADPH reductase-RFP (FNR-RFP). During division the apicoplast shows close apposition to the centrosome (see Figure 8F) [12]. When compared to the control fiber SFA3-YFP labeling of the bleached fiber appeared to be polar and proximal to the apicoplast. Thus, it appears that the fiber grows out by polymerization and that the new subunits are added at the end proximal to the centrosome which could be considered the “plus” end of the fiber. We note that we currently do not have a suitable probe to observe the bleached segment of the fiber, and thus cannot measure its entire length. We therefore cannot formally exclude proximal labeling due to laser induced stunting or breakage of the fiber.
Rootlet fibers are typically found in intimate contact with basal bodies or microtubules [15],[32]. Our observations are consistent with a polar SFA fiber that places and potentially governs the formation of the daughter cell and/or its MTOC. Alternatively, newly formed microtubules, for example, the central pair, could be recruiting the fiber to the MTOC, tethering the daughter in a secondary fashion. To distinguish between these two alternatives we tested whether daughter cell microtubules are required for SFA fiber formation or vice versa. We stained microtubules in the iΔSFA3 mutant using different tubulin antibodies. In Figure 9E, 9F. we show antibody 6–11B directed against acetylated alpha tubulin as this antibody recognized daughter cell microtubules particularly well [33]. In mutants treated with ATc for 48 h no daughter microtubules are detectable. Note that these particular cells (Figure 9F) are already multinucleated and undergoing another round of mitosis as indicated by the presence of two centrosomes per nucleus. Microtubules of the mother cell are readily detected. Conversely we treated SFA3-YFP or SFA2-HA parasites with 2.5 µM of the microtubule disrupting agent oryzalin [34]. As previously reported [35] oryzalin-treated parasites fail to produce daughter cells; note the lack of daughter IMC1 staining in Figure 9B. However, in these parasites SFA fibers are still detected (Figure 9A and 9B, green). In fact fibers are noticeably more abundant; 60% of vacuoles exhibit SFA fibers after 24 h of oryzalin treatment while only 25% of the control parasites do (Figure 9C). We further noticed that in oryzalin-treated parasites SFA fibers remain shorter and show a more uniform size distribution when compared with untreated parasites (Figure 9D). We conclude that daughter cell microtubules are not required for SFA fiber formation but that microtubule elongation may be required for the fiber to extend to its full length, break, and disappear. Alternatively, fiber elongation might require licensing by a checkpoint controlled by daughter cell growth.
Flagella provide motility and sensory functions to a large variety of single and multicellular eukaryotes. They are anchored in the cell by the basal body [36]. Flagellar basal bodies are embedded within a complex cytoskeletal system known as the flagellar rootlet system or basal body cage, which has been studied in most detail in the green alga Chlamydomonas reinhardtii [32]. There is evidence that these structures not only position the flagella but also define cellular axes of symmetry and asymmetry [37],[38]. The rootlet is composed of several types of biochemically and structurally distinct fibers some of which are made of microtubules. In Chlamydomonas, centrin-based fibers (also known as contractile fibers) interconnect basal bodies and connect the basal bodies to the nucleus [39]. Sinister fibers, first described in S. similis, connect the basal bodies to two of the four microtubules of the flagellar rootlet and to cytoplasmic microtubules [40]. Striated fibers are made up from a single SFA protein and run along microtubules, emerging close to basal bodies in post mitotic cells, and are thought to guide and stabilize their associated microtubules (Figure 10A) [17],[18]. It has been proposed that the mechanism by which SFA binds microtubules is related to the structure of a rod domain found in the protein which consists of 29 amino acid repeats. The periodicity of this repeat confers the characteristic striation pattern found in SMAFs, and also fits the spacing between tubulin subunits in microtubules [18]. Striated fibers are also found in association with basal bodies in mammalian cells (e.g., various receptor cells), but the proteins isolated from these fibers do not appear to be homologous to SFA proteins [41].
In this study we show that SFAs play a critical role in the cell division of the apicomplexan parasite T. gondii. We identified a fiber that is made up of at least two proteins, TgSFA2 and 3. This structure becomes apparent as soon as the centrosome is duplicated; it emerges from in between the two centrioles and grows away from the centrosome. Its distal end is intimately associated with the apical tip of the daughter cell (Figure 10D summarizes our current ultrastructural understanding). Interestingly, the SFA fiber is not only a tether between centrosome and the daughter; it is required for daughter assembly. In conditional mutants lacking the SFA fiber we do not detect daughter buds even using the earliest markers available. The “birth” of the daughter is the establishment of the apical MTOC. We propose that the distal end of the SFA fiber initiates and thus positions the daughter MTOC. Could the fiber itself be an MTOC? Several studies have demonstrated the ability of algal rootlet complexes to initiate microtubule assembly in vivo and in vitro [42],. It is tempting to note the peculiar shape of the end of the Toxoplasma fiber in the context of the circular MTOC found in these organisms and to speculate that the structure of the fiber may template the microtubule arrangement of the daughter pellicle. Further biochemical work is required to test this hypothesis in vitro. Chlamydomonas SFA has been demonstrated to have the intrinsic ability to self-assemble and self-organize; recombinant SFA forms striated fibers in vitro [44].
Why does a cell in a stage lacking flagella use a budding system that depends on elements of the flagellar rootlet? We believe this to be a consequence of the evolutionary history of Apicomplexa. The assembly of the flagellum and the mitotic spindle share deep evolutionary roots. The centrioles, which are at the core of many, but not all mitotic spindle poles, are homologous to the basal bodies [36]. In fact, in many cells the very same structure performs both functions. In Chlamydomonas, during interphase two closely apposed basal bodies organize the organisms' two flagella [32]. These flagella are resorbed upon entry into cell division and the basal bodies become associated with the poles of the mitotic spindle. Following division both daughters assemble a transition zone onto the centrioles and reform flagella. The rootlet (shown schematically in white in Figure 10B) appears to be important in both roles setting up division and symmetry planes and positioning the centrosome and the flagellum relative to each other as the cells move through their replicative cycle [37],[38]. Apicomplexa are believed to have an evolutionary past as photosynthetic aquatic algae. They are part of the Chromalveolata, a large branch of the eukaryotic tree of life that emerged from the endosymbiosis between a flagellated protist and a red alga [45]. The most conspicuous holdover of this past is a chloroplast-like organelle [46]. We hypothesize that the ancestors of Apicomplexa, as their present day kin likely depended on the rootlet to organize the relationship of their flagellar and mitotic MTOCs. As they adapted to intracellular parasitism they developed specialized cytoskeletal and secretory organelles that allow them to attack other cells [1]. Some precursors of these organelles are found in closely related flagellated protists that are fully or partially symbiotic, predatory or parasitic—in all these cases the organelles are found in close proximity to the flagellar basal body [47]–[49]. We propose that Apicomplexa subsequently abandoned the flagellum for most stages yet retained the organizing principle of the MTOC. Instead of ensuring that daughter cells have appropriate numbers of flagella the system now measures out and positions the apical invasion complexes (Figure 10C). Overall this suggests that elements of the invasion apparatus may be derived from flagella- or flagellum-associated structures.
In combination with the recently described tethering of the nuclear genome and other organelles [9],[10],[12], a remarkably hard-wired model for the assembly of infective parasite stages emerges. The role of the SFA fiber is crucial in this context, a self-organizing polar fiber that will initiate a daughter in the proximity of each centrosome once its components are expressed. The elegance of this mechanism is its scalability and independence of ploidy. It satisfyingly explains how T. gondii can form two daughters per round of budding, while P. falciparum forms ten to 20 in the red cell, and many thousands during liver cell infection. Direct evidence for the control of zoite formation by the flagellar rootlet is currently limited to the experiments with T. gondii presented in this study. However, we note that SFA homologs are encoded in the genomes of all apicomplexans for which sequence is available (and many other chromalveolates [50]). Furthermore electron dense structures comparable to those identified as SFA fibers in this study have been observed in previous ultrastructrual reports in Eimeria and Plasmodium [51]–[53]. Many mechanistic questions remain, some of them directed towards the relationship between the structure of the fiber and its function. There are also intriguing problems associated with the spatial and temporal control of initiation and breakdown of the structure, and how they are integrated into the parasites mechanisms of cell cycle control, which remain to be deciphered.
T. gondii RH strain parasites were maintained by serial passage in human foreskin fibroblast (HFF) cells and genetically manipulated as previously described [54]. To tag the genomic locus of TgSFA2 (GenBank accession XM_002367757) with a 3×HA tag, 585 bp of the open reading frame ending before the stop codon was amplified from the T. gondii genomic DNA. All primer sequences used are shown in Table S1. Similarly, 3,000 bp upstream of the stop codon of XM_002370621.1 were amplified to tag TgSFA3 with YFP. These amplicons were cloned via ligation-independent cloning (LIC) [55] into the pLIC-HA-CAT or pLIC-YFP-DHFR vector, respectively, to create in-frame fusions [56]. Transgenic clones were established by transfection of ΔKu80 parasites and chloramphenicol or pyrimethamine selection, respectively, as previously described [56]. Integration was confirmed by PCR or Southern blotting as previously described [57]. A probe complementary to the 3′ region of the SFA2 gene was amplified by PCR. Cen1-RFP [30] was introduced into SFA3-YFP parasites by transient transfection. FNR-RFP [57] was transfected into SFA3-YFP parasites and stable transgenics were isolated by fluorescence activated cell sorting [54].
To target SFA2, 1,500 bp immediately up- and downstream of the start codon were amplified and introduced into vector piKO in order to flank an HXGPRT selectable marker [58], a transactivator (TaTi) and the tetracycline regulatable T7S1 promoter (this plasmid was a kind gift of Dominique Soldati, University of Geneva). The final construct was linearized using NcoI/SpeI and transfected into ΔKu80 parasites. Clones were obtained after mycophenolic acid selection and screened for locus insertion by PCR screen (see Figure S2C and S2D). Mutants were grown in 0.5 µg/ml of ATc (Sigma-Aldrich) for 1–4 d, total RNA was isolated (RNAeasy, Qiagen), reverse transcribed (Invitrogen), and reverse transcription-PCR was performed using SFA2 specific and control primers (Table S1).
For SFA3 a cosmid (PSBLE51) was modified by recombineering [26] to replace the native promoter by a regulatable promoter. A suitable cassette was constructed by inserting a gentamycin marker [26] into the promoter replacement plasmid pDT7S4_087270 [24]. The resulting plasmid (pGDT7S4_087270) was used as template to amplify the modification cassette using 50-bp homology flanks for insertion into SFA3 (Figure S2A). The modified cosmid was isolated by double selection on gentamycin and kanamycin [26] and transfected into TATiΔKu80 parasites [24], clones were isolated after pyrimethamine selection and tested for promoter replacement by PCR (Figure S2A and S2B).
The complete coding region of TgSFA3 was amplified from the T. gondii genomic DNA and inserted into plasmid pAVA-421 6xHis [59]. Recombinant fusion protein was purified on Ni2−-NTA resin (Qiagen) [60]. Rabbits were immunized with 1 mg of purified protein, and serum was collected after 10 wk (Cocalico Biologicals). The sequence encoding for amino-acids 1–110 of TgCENH3 [9] was amplified and cloned into the same expression vector and purified in a similar fashion as SFA3. Mice were immunized with 0.4 mg of purified protein, and serum was collected after 10 wk.
For IFAs, host cells (HFF) were inoculated onto coverslips and infected with parasites. Coverslips were fixed 24 h after infection with 4% formaldehyde in PBS and permeabilized with 0.2% Triton X-100 in PBS/3% BSA. Coverslips were then blocked in 3% bovine serum albumin (BSA) in PBS as previously described [61]. Primary antibodies used were mouse anti-alpha tubulin at a dilution of 1∶1,000 (12G10, a gift of Jacek Gaertig, University of Georgia), rabbit anti-Centrin1 at 1∶1,000 (gift of Iain Cheeseman, Massachusetts Institute of Technology), mouse anti-GFP at 1∶1,000–1∶400 (Torry Pines Biolabs), rat anti-HA at 1∶1,000 (clone 3F10, Roche Applied Science), mouse anti-IMC1 mAb 45.15 [62] at 1∶1,000 (gift of Gary Ward, University of Vermont), rabbit anti-IMC3 [63] at 1∶500, mouse anti-ISP1 mAb 7E8 [27] at 1∶1,000 (gift of Peter Bradley, University of California, Los Angeles), rabbit anti-MORN1 [30] at 1∶250, anti-aceytlated tubulin (Sigma) at 1∶1,000, and rabbit anti-SFA3 at 1∶1000 (generated in this study). The secondary antibodies used were AlexaFluor 350, AlexaFluor 488, and AlexaFluor 546 (Invitrogen), at a dilution of 1∶2,000. Images were collected on an Applied Precision Delta Vision inverted epifluorescence microscope using a UPlans APO 100×/1.40 oil lens. Time-lapse imaging was performed on the same Delta Vision microscope in a climate controlled chamber at 37°C. Cells were grown and imaged on glass bottom Wilco culture dishes (Wilco Wells). Images were obtained every 10 min for 4 h, and processed to correct for cell drifting. Photobleaching of SFA3-YFP was performed on a Delta Vision microscope using a single 600-ms pulse with a 488-nm laser set at 20% power, on a specified, diffraction-limited, region. Images were subjected to deconvolution and contrast adjustment using Applied Precision software (Softworx). For quantitative image analysis (presence/absence of fibers and number of nuclei/cell or centromeric clusters/nucleus, as described in the Results) a minimum of 50 vacuoles were scored for each out of at least three repeats. Averages and standard deviations were calculated and plotted using Graph Pad Prism Version 5.0c. Fiber length measurements were done using Volocity (Perkin-Elmer) on images taken of SFA3-YFP parasites under oryzalin or DMSO control treatment. Each point represents the average fiber length in one imaging field. Measurements were done for at least 18 fibers (and up to 81) per image. Four images from three independent replicates were used.
Cells infected with SFA2-HA parasites were fixed in 4% para-formaldehyde/0.05% glutaraldehyde in 0.1 M sodium phosphate buffer pH 7.4, blocked with 1% FBS in PBS (all RT), followed by overnight infiltration in 2.3 M sucrose/20% polyvinyl pyrrolidone at 4°C. Samples were frozen in liquid nitrogen, and sectioned with a Leica UCT cryo-ultra microtome. Sections were blocked with 1% FBS and subsequently incubated with rat anti-HA (1∶100), followed by incubation with rabbit anti-rat (1∶400), and finally 10 nM colloidal gold conjugated protein A. Washed sections were stained with 0.3% uranyl acetate/2% methyl cellulose and viewed with a JEOL 1200 EX transmission electron microscope. Controls, omitting the primary antibody, were consistently negative at the concentration of colloidal gold conjugated protein A used. Infected cells were also fixed in 2% glutaraldehyde in sodium phosphate buffer 0.1 M, pH7.4, followed by post-fixation with 1% osmium tetroxide in sodium phosphate buffer, alcohol dehydration, and Epon resin embedding. Serial sections were obtained with a Leica UCT cryo-ultramicrotome and collected in carbon-coated single hole grids.
Western blotting was performed as previously described [57]. We used anti-HA antibodies at a dilution of 1∶1,000, anti-tubulin at 1∶1,000, anti-GFP at 1∶500, and anti-SFA3 antibodies at a dilution of 1∶1,000. Pre-immune sera for anti-SFA3 antibodies were used at a dilution of 1∶1,000. Horseradish peroxidase (HRP)-conjugated anti-rat or anti-rabbit antibody (Pierce) was used at a dilution of 1∶20,000.
|
10.1371/journal.pntd.0000993 | Enzymatic Shaving of the Tegument Surface of Live Schistosomes for Proteomic Analysis: A Rational Approach to Select Vaccine Candidates | The membrane-associated and membrane-spanning constituents of the Schistosoma mansoni tegument surface, the parasite's principal interface with the host bloodstream, have recently been characterized using proteomic techniques. Biotinylation of live worms using membrane-impermeant probes revealed that only a small subset of the proteins was accessible to the reagents. Their position within the multilayered architecture of the surface has not been ascertained.
An enzymatic shaving approach on live worms has now been used to release the most accessible components, for analysis by MS/MS. Treatment with trypsin, or phosphatidylinositol-specific phospholipase C (PiPLC), only minimally impaired membrane integrity. PiPLC-enriched proteins were distinguished from those released in parasite vomitus or by handling damage, using isobaric tagging. Trypsin released five membrane proteins, Sm200, Sm25 and three annexins, plus host CD44 and the complement factors C3 and C4. Nutrient transporters and ion channels were absent from the trypsin fraction, suggesting a deeper location in the surface complex; surprisingly, two BAR-domain containing proteins were released. Seven parasite and two host proteins were enriched by PiPLC treatment, the vaccine candidate Sm29 being the most prominent along with two orthologues of human CD59, potentially inhibitors of complement fixation. The enzymes carbonic anhydrase and APD-ribosyl cyclase were also enriched, plus Sm200 and alkaline phosphatase. Host GPI-anchored proteins CD48 and CD90, suggest ‘surface painting’ during worm peregrination in the portal system.
Our findings suggest that the membranocalyx secreted over the tegument surface is not the inert barrier previously proposed, some tegument proteins being externally accessible to enzymes and thus potentially located within it. Furthermore, the detection of C3 and C4 indicates that the complement cascade is initiated, while two CD59 orthologues suggest a potential mechanism for its inhibition. The detection of several host proteins is a testimonial to the acquisitive properties of the tegument surface. The exposed parasite proteins could represent novel vaccine candidates for combating this neglected disease.
| Adult schistosome parasites can reside in the host bloodstream for decades surrounded by components of the immune system. It was originally proposed that their survival depended on the secretion of an inert bilayer, the membranocalyx, to protect the underlying plasma membrane from attack. We have investigated whether any proteins were exposed on the surface of live worms using incubation with selected hydrolases, in combination with mass spectrometry to identify released proteins. We show that a small number of parasite proteins are accessible to the enzymes and so could represent constituents of the membranocalyx. We also identified several proteins acquired by the parasite on contact with host cells. In addition, components of the cytolytic complement pathway were detected, but these appeared not to harm the worm, indicating that some of its own surface proteins could inhibit the lytic pathway. We suggest that, collectively, the ‘superficial’ parasite proteins may provide good candidates for a schistosome vaccine.
| The persistence of adult schistosomes in the bloodstream for decades means they must deploy unique and effective immune evasion strategies at their interface with the host. The 1 cm-long worms are covered by a naked syncytial layer of cytoplasm, the tegument, connected by cytoplasmic tubules to underlying cell bodies that contain the machinery for protein synthesis, packaging and export. The tegument surface has a multilaminate appearance, interpreted as a plasma membrane overlain by a lamellate secretion, the membranocalyx [1]. This complex molecular architecture is maintained by export of the contents of multilaminate vesicles, which originate in the cell bodies of the syncytium. There is experimental evidence for slow turnover of the membranocalyx to the external environment [2] whilst recycling of the plasma membrane by internalisation has been anticipated but not conclusively demonstrated [3]. Initial observations suggested the membranocalyx was an amphipathic bilayer, probably composed of phospholipids, which served as a physical barrier to prevent antibody binding or host leukocyte attachment to the underlying plasma membrane. In addition, its supposed hydrophobic properties coincided with a demonstrable ability of worms in the bloodstream to acquire host molecules, particularly erythrocyte glycolipids (the so-called host antigens) [4]. Whether this acquisition is a deliberate process that benefits the parasite or an accidental consequence of the membranocalyx properties, and any relevance it has to immune evasion, remain unclear.
Building on techniques developed in 1980s to detach the tegument by freeze-thaw and enrich the surface membrane complex by differential centrifugation [5], we have characterized its composition using proteomic techniques [6], [7]. We developed a differential extraction scheme for the membrane preparation, with chaotropic agents of increasing strength, which enabled us to identify both membrane-associated and membrane-spanning constituents. These compositional findings demonstrated the importance of the tegument for nutrient uptake and maintenance of solute balance, as well as the presence of several hydrolases in the surface layers [6]. In a second study we incubated live worms with membrane-impermeant probes to biotinylate the most externally-accessible proteins and then recovered the tagged molecules by affinity chromatography for MS/MS identification [7]. This approach revealed that only a small subset of transporters, membrane structural proteins, enzymes, and schistosome-unique proteins were labelled, together with host immunoglobulins and complement C3. We concluded that these represented the most “exposed” surface constituents but we could not place them within the multilayered architecture of the surface with any certainty.
To add another dimension to our understanding of tegument surface organization we have now used an enzymatic shaving approach on live worms to release the components most accessible to the selected enzymes, for MS/MS analysis. By analogy with techniques for stripping adherent cells from culture flasks, we used trypsin to cleave exposed protein loops or domains without impairing membrane integrity. We also incubated worms with phosphatidylinositol-specific phospholipase C (PiPLC) to release any externally accessible GPI-anchored proteins. As a control for proteins released by the vomiting of gut contents during the incubation period, or damage due to handling, we compared protein release +/− PiPLC, using the iTRAQ technique. Finally we used a phospholipase A2 (PLA2) preparation purified from snake venom to erode the lipid bilayer complex to determine if proteins could be selectively detached. We report that both trypsin and PiPLC removed a small subset of proteins whilst inflicting minimal damage on the worms whereas PLA2 was more destructive. We show that the iTRAQ technique identified both parasite and host proteins enriched by PiPLC treatment whereas trypsin released a different subset, with only Sm200 common to both. The proteins we have identified in the adult are also present as transcripts in the lung schistosomulum. We suggest that collectively they may be candidates for a schistosome vaccine, especially if responses can be targeted to the lungs to interfere with intravascular migration of incoming larvae.
The procedures involving animals were carried out in accordance with the UK Animals (Scientific Procedures) Act 1986, and authorised on personal and project licences issued by the UK Home Office. The study protocol was approved by the Biology Department Ethical Review Committee at the University of York.
A Puerto Rican isolate of S. mansoni was maintained using albino Biomphalaria glabrata snails and NMRI strain mice as laboratory hosts. All animal experiments were approved by the Ethical Review Process Committee of the Department of Biology, University of York. Adult parasites were obtained by portal perfusion of mice seven weeks after exposure to 200 cercariae, using RPMI1640 medium (minus phenol red) buffered with 10 mM HEPES (both from Invitrogen, Paisley, UK). Parasites were extensively washed in the same medium and tissue debris and any damaged individuals removed under a dissecting microscope. No attempt was made to separate males from females.
Approximately 800 freshly perfused parasites from 20 mice were used in each of four replicate experiments with trypsin as the shaving enzyme. They were incubated in a 30 mL Corning flask (Corning, NY, USA) containing 5 mL of buffered RPMI, with trypsin MS (Promega, Southampton, UK) added at 10 µg/mL, for 30 min at room temperature (RT). The supernatant was recovered, transferred to a 15 mL Falcon tube and centrifuged at 500×g for 30 min to remove any insoluble material such as the haematin particles in gut vomitus. Streptomycin and penicillin were then added to a final concentration of 100 µg/mL to prevent microbial growth and tryptic digestion was continued overnight at 37°C, after which peptides were reduced and alkylated. Reduction was performed in the presence of 20 mM DTT for 30 min at 65°C in a water bath and, after cooling, alkylation was performed in the presence of 80 mM iodoacetamide for 1 h at RT in the dark. Trifluoroacetic acid (TFA) was then added to a final concentration of 0.1% before recovery of peptides by passage through a solid phase Strata C18-E extraction cartridge (55 µm, Phenomenex, Macclesfield, UK), followed by several column washes in 0.1% TFA and final elution in 750 µL of 50% acetonitrile/0.1% TFA. The eluted fraction was concentrated under vacuum to dryness and peptides resuspended in 20 µL 0.1% TFA.
A 3 µL aliquot of the tryptic peptide preparation was injected onto a reversed-phase PS-DVB monolith column (200 µm i.d.×5 cm, LC Packings, Amsterdam, Netherlands). Peptides were separated using a two-step linear gradient of 2–31.4% (v/v) acetonitrile in 0.1% aqueous heptafluorobutyric acid over 60 min, followed by 31.4–51% (v/v) in the same solvent over 5 min, at a flow rate of 3 µL/min; UV absorbance at 214 nm was monitored. Fractions were collected onto a MALDI target plate using a Probot (Dionex, Bannockburn, USA) with simultaneous addition of matrix solution (6 mg/mL α-cyano-4-hydroxycinamic acid (CHCA, Sigma, Poole, UK) in 60% (v/v) acetonitrile).
Positive-ion MALDI mass spectra (MS) were obtained using a 4700 Proteomics Analyzer with TOF-TOF Optics (Applied Biosystems, Framingham, USA) in reflector mode, over the m/z range 800–4000 and monoisotopic masses obtained from centroids of raw, unsmoothed data. The precursor mass window was set to a relative resolution of 50, and the metastable suppressor was enabled. The default calibration was used for MS/MS spectra, which were baseline-subtracted (peak width 50) and smoothed (Savitsky-Golay with three points across a peak and polynomial order 4); peak detection used a minimum S/N of 5, local noise window of 50 m/z, and minimum peak width of 2.9 bins. The twenty strongest peaks from each fraction, having a signal to noise (S/N) greater than 50 and a fraction-to-fraction precursor exclusion of ±0.2 Da, were selected for CID-MS/MS analysis. Singly-charged peptides were fragmented with Source 1 collision energy of 1 keV, and air as the collision gas. Peak lists from the MS/MS data, containing all m/z values from m/z 20 to the precursor m/z - 60, with a minimum S/N of 10, were provided by TS2 software (version 1.0.0, Matrix Science Ltd., London, UK). Each list, corresponding to one MALDI plate, was then submitted to a local copy of the Mascot program (version 2.1, Matrix Science) and searched against the SmGenesPlusESTs (260448 sequences; 71029272 residues), an in-house database derived from the publically available data in http://www.genedb.org/genedb/smansoni/), and the NCBInr Mus musculus database (139457 sequences) for host proteins. Search parameters specified only tryptic cleavages and allowed for up to one missed site, the variable carbamidomethylation of cysteines, and oxidation of methionines; precursor and product ion mass error tolerance was set to ±0.3 Da. A decoy database, generated by Mascot, was used with the significance threshold for protein identification set to achieve a false positive rate of 1 to 2% and peptide threshold set to ‘least identity’. A protein was considered positively identified if the ion score for a particular peptide had an expect value less than 0.05.
GPI-anchored proteins were recovered from live worms by in vitro incubation with PiPLC as the shaving enzyme, the experiment being performed twice to provide biological replicates. Downstream analysis required the worms from 40 mice, which were perfused and treated in two separate batches to minimise the time ex vivo; supernatants were then combined. Each batch was incubated at 37°C for 1 h in the presence of PiPLC (from Bacillus cereus, Sigma) at 1.25 Units/mL, with conditions as for trypsin. The supernatant was removed and concentrated at 4°C using a 5000 Da cut-off centrifugation device (Vivaspin 6, West Sussex, UK). The control for secretion, vomitus production and parasite damage due to handling comprised an identical experiment, minus PiPLC. This also served as a ‘background’ control for the other enzyme treatments. For one experiment, PiPLC-released proteins were obtained using PiPLC from a different source (from Bacillus thuringiensis, Europa Bioproducts, Wicken, Cambridigeshire, UK) employing the same conditions as above. On that occasion the GPI-released fraction was used for a 2-DE separation.
The composition of released material was evaluated by 1-DE using a pre-cast NUPAGE 4–12% Bis-Tris gel (Invitrogen) after a 45 min run at 200 V. The gel was then fixed in 40% methanol, 10% acetic acid for 30 min, stained with SYPRO Ruby (Invitrogen) for 2 h in the dark and imaged using a Molecular Imager FX (Bio-Rad, Bath, UK). Protein content in each lane of the gel was then estimated by densitometric analysis using Quantity One software (Bio-Rad). Fifty µg of the PiPLC-treated sample was also evaluated by mini 2-DE essentially as previously described [8], [9]. After electrophoresis the gel was first stained with SYPRO Ruby, imaged as above, and restained with Bio-Safe Coomassie (BioRad); all visible spots were selected for “in gel” digestion [8]. An aliquot of 1–2 µL of the digestion supernatant containing the peptides was spotted on a MALDI plate and dried before the addition of 0.6 µL of a saturated solution of CHCA matrix (in 50% acetonitrile/0.1% TFA). Peptide fragmentation data from each gel spot was processed by GPS Explorer Software (Applied Biosystems) underpinned by Mascot (settings as above), to provide a putative identity for the protein.
The relative composition of PiPLC-treated and control samples was characterized using isobaric tagging (the iTRAQ labelling technique) following the protocol provided by the manufacturer (Applied Biosystems). Prior to labelling, two aliquots of treated and control samples containing 10 µg protein were taken to provide technical replicates. Briefly, 10 µg of both control and PiPLC-treated samples were individually denatured, reduced, and alkylated with reagents supplied in the iTRAQ kit. Peptides were generated by trypsin digestion using a 1∶20 enzyme/protein ratio, at 37°C for 24 h, and labelled with iTRAQ reagents at lysine, terminal amine groups and partially at tyrosine residues. Test samples were labelled with tags 116 or 117 and control samples with tags 114 or 115, respectively. A peptide mixture was made by combining the four tagged samples and cleaned up using a strong cation-exchange cartridge to remove the detergents and excess iTRAQ reagents. The peptides were then affinity-purified in a Strata C18-E cartridge, eluted as for tryptic peptides, dried using a vacuum concentrator and resuspended in 10–20 µL of 0.1% TFA. LC-MS/MS was performed as described above. Protein identification and peptide quantification was achieved by submitting the TS2-generated MS/MS raw data files to Mascot, searching against SmGenesPlusESTs and NCBInr databases. Search parameters were tryptic peptides, with 0–1 missed cleavage; fixed modifications, β-methylthiolation of cysteines, iTRAQ tagging of lysines and N-terminal amine groups; variable modifications were oxidation of methionines and iTRAQ tagging of tyrosines. Precursor and product ion mass error tolerance was set to ±0.3 Da.
The 114-tagged sample (C1) was taken as the reference for calculating ratios. The Mascot software displays the median normalised geometric mean ratios and a factor from which the geometric standard deviation can be derived. Automatic outlier removal was performed by the Mascot software. A protein was considered enriched if the mean T1/C1 and/or T2/C1 ratios minus the 95% confidence limit exceeded the corresponding C2/C1 ratio plus its 95% confidence limit. For single peptide identifications, a protein was considered enriched if it appeared in both biological replicate experiments, and had a mean treated/control ratio >2.
The final approach for recovery of parasite surface molecules using enzymatic shaving, involved the treatment of live parasites with PLA2. This enzyme isolated from the venom of Crotalus durissus terrificus (deposited under accession number P24027 at NCBInr) was kindly provided by Prof. Andreimar Soares (Faculty of Pharmaceutical Sciences, University of Sao Paulo, Brazil). During the shaving experiment, approximately 800 freshly perfused parasites were incubated in a 30 mL Corning flask, containing 6 mL of buffered RPMI1640, with PLA2 added at 16 µg/mL for 1 h, at RT. Another batch of parasites was used in a control and parallel experiment in which no enzyme was added to the culture medium. The supernatants from control and treated samples were concentrated using a Vivaspin 6 filtration device (5000 Da cut off) at 500×g, at 4°C until the volume reached 500 µL. The samples were then transferred to 1.5 ml eppendorf tubes and centrifuged at 25,000×g for 20 min. After this step a membranous pellet, recovered from PLA2-treated parasites only, was extracted in 50 µL of 0.5% Triton-X100, yielding approximately 20 µg of protein. Reduction, alkylation, trypsin digestion and peptide clean-up were performed as described for the iTRAQ protocol (omitting the labelling steps). LC-MS/MS of the peptide mixture was performed essentially as described above.
Host plasma membrane proteins on the surface of adult parasites were investigated either on live worms, perfused from mice using RPMI medium, or by preparing OTC-embedded cryostat sections. Both were incubated with primary antibodies at 1∶100 dilution in PBS for parasite sections and in RPMI for live worms, containing 5% normal goat serum, for 1 h at RT. Monoclonal antibodies used were rat anti-mouse CD44 (558739), rat anti-mouse CD90 (Thy-1; 553016) and hamster anti-mouse CD48 (553682), all from BD Biosciences Pharmingen, New Jersey, USA. Labelling was detected by the use of goat anti-rat IgG conjugated to Alexa fluor 488 at 1∶500 dilution for 30 min in the same buffer, or goat anti-hamster IgG conjugated to Alexa fluor 568 under the same conditions.
A searchable database was created in 2005, comprising all S. mansoni transcripts then available from dbEST (http://www.ncbi.nlm.nih.gov/), the Sao Paulo Schistosoma mansoni EST genome project (http://bioinfo.iq.usp.br/schisto/), and the Wellcome Trust Sanger Institute ftp site (ftp://ftp.sanger.ac.uk/pub/pathogens/Schistosoma/mansoni/ESTs). All proteins of interest identified by the enzymatic shaving approach were searched against the compiled EST data to determine the number of transcripts detected in adults and lung stage schistosomula; this provided a very approximate guide to the relative levels of expression in the two life cycle stages.
Freshly perfused live parasites were subjected to trypsin digestion under controlled conditions, in order to release from the surface membrane complex of the tegument any proteins, or segments thereof, accessible to the enzyme. The vast majority of worms retained a normal appearance and activity over the 30 min incubation. The culture supernatant was recovered and released proteins/peptides allowed to digest further overnight. Tryptic peptides were then reduced and alkylated before their recovery and separation using reversed-phase chromatography (Figure S1A) for subsequent mass-spectrometric identification. The identities obtained by Mascot searching of the MS/MS data were then categorized according to their molecular function and, by inference, their potential cellular origin (Table 1).
The host complement proteins C3 and C4 and the leukocyte surface marker CD44 were released in two or more of the four replicate experiments. Host haemoglobin alpha and beta chains were also identified in one experiment, presumably an indication that regurgitation from the worm gut was occurring. A number of proteins, known to be associated with the tegument surface, were also released by the treatment. They included three phospholipid-binding annexins and the membrane protease calpain. Of the annexins, Smp_077720 was found in all four experiments and the others on two and one occasions, respectively; calpain was detected in only one experiment. The schistosome-unique proteins of unknown function, Sm200 and Sm25, were found on three and four occasions, respectively. The final membrane protein, not previously associated with the tegument, shows homology with cell surface proteoglycans. A pair of BAR-domain-containing endophilins was also released by the treatment in all four experiments while the putative potassium-channel inhibitor, SmKK7, was found on one occasion.
Although worm incubations were short-term, two known gut-derived proteases, asparaginyl endopeptidase and cathepsin B1 were detected. In addition, a total of 12 proteins with unknown function and localization were found (Tables S1 and S2), some containing domains, e.g. Ig-like and EGF-like, which may indicate a surface position. In spite of strenuous efforts to maintain worm viability, the trypsin treatment affected the integrity of the surface membranes to some extent, as evidenced by the appearance of seven cytoskeletal and 17 cytosolic proteins in the medium (Tables S1 and S2). Among the former, actin, fimbrin and severin have been proposed as constituents of the tegumental spines that reside immediately beneath the surface membrane complex. The more numerous cytosolic proteins, comprising glycolytic enzymes, chaperones, and antioxidants indicate the leakage of internal components in at least some of the parasites. Known tegumental surface transporters, ion channels, and enzymes (other than calpain) were conspicuous by their absence in the trypsin preparation.
The purpose of treatment with PiPLC was to release GPI-anchored proteins accessible to the externally applied enzyme in live worms; incubation with the enzyme for 1 h appeared to have no morphologically obvious deleterious effect. A 1-DE gel separation of material released by treated and control worms revealed a complex pattern of protein bands with Mr ranging from 10 to >250 kDa (Figure 1A). The two preparations displayed a strong similarity, with only three bands (arrowed) visibly enriched by the PiPLC treatment versus the control; two were of high molecular mass (approx. 200 kDa) while the other was located at the bottom of the gel (approx. 12 kDa). The PiPLC treatment released sufficient protein to permit a mini 2-DE separation (Figure 1B) for subsequent MS/MS analysis of gel spots. The analysis revealed the presence of proteins known to be GPI-anchored, such as ‘Surface protein’ (Sm200) and alkaline phosphatase. Protein orthologues of CD59 and carbonic anhydrase IV were novel features of this 2D map. The absence of gut proteases Sm31 and Sm32 was notable. The identities of other spots revealed the presence of cytosolic and cytoskeletal contaminants, including thioredoxin, fatty acid binding protein (Sm14), Sm22.6, enolase and triose phosphate isomerase (Table S3). The remaining spots on the SYPRO-stained gel (Figure 1B) were not detected by Coomassie staining so no identification could be assigned using single spot tryptic digestion.
As proteins were released into the culture medium during the 1 h incubation irrespective of whether PiPLC was present or not, we used the iTRAQ technique to determine the degree of enrichment due solely to the enzymatic shaving. Within an individual labelling protocol, splitting both the control and treatment samples provided technical replicates. After LC separation of the tagged peptide mixture (Figure S1B), fragmentation spectra of the most abundant peptides were generated, each containing signature peaks for the four reporter tags (Figure 2 and Figure S2). In most instances the fragmentation spectra yielded four peaks of approximately equal area (Figure 2A). The 115/114 (C2/C1) ratios approximated to unity, indicating approximately equal protein contributions of the two control samples to the iTRAQ mixture. However, the majority of 116/114 (T1/C1) and 117/114 (T2/C1) ratios were almost invariably less than one, sometimes significantly so (Figure 2A and Figure S2). We have no reason to believe that the control and treatment incubations differed except in the addition of enzyme, and so we must attribute the ratios of less than unity to a dilution effect on the background proteins in the treated samples. This dilution resulted from the addition of PiPLC plus the extra proteins released from the surface by its action. In a minority of fragmentation spectra the four peaks representing the reporter tags were not of approximately uniform area (Figure 2B and Figure S2). The reporter ion peaks from the treated samples were from 2 to 10 fold greater intensity indicating enrichment of the parent peptide by the PiPLC treatment. A total of 52 identities was obtained from fragmentation of the tagged peptides (Figure S2 and Tables S4 and S5) ranging from 1 to 15 per identity, primarily of cytosolic or cytoskeletal origin. The PiPLC-enriched proteins of parasite origin were, in descending order of abundance, Sm29, CD59a, Sm200, carbonic anhydrase, CD59b, alkaline phosphatase and ADP-ribosyl cyclase (Figure 3). In addition two host proteins, CD48 and CD90 (Thy1.2), were also enriched by PiPLC treatment. Proteins of known gut origin, α2-macroglobulin and saposin B, were present in control and treated samples in equivalent amounts indicating similar regurgitation of gut contents over the one hour incubation. Equivalent amounts of cytosolic (e.g, enolase, 14-3-3) and cytoskeletal proteins (e.g, actin, Sm20.8) in control and treated samples (Figure 3 and Table S4) indicated a certain degree of surface damage, but not inflicted by the PiPLC treatment. Although not enriched by PiPLC hydrolysis, because they lack the GPI-anchor, three other proteins are worth noting. Two of these, CD63/tetraspanin (TSP-2) and annexin IV (Smp_074140) were already known to be associated with the tegument surface and the third is an 8 kDa low molecular weight protein (LMWP). This last protein, present in a range of trematodes [10], has a signal peptide and so may be a true secreted protein released along with the membranocalyx; its status as a tegument surface protein needs to be confirmed.
Unlike the other two enzymatic treatments PLA2 had a dramatic effect on worm appearance and viability. Moderate concentrations of PLA2 produced visible worm damage and death, whereas greater dilutions had little selective effect in removing known tegumental surface components whilst still causing leakage of cytoskeletal and cytosolic components (Table S6). The known tegument surface proteins released were Sm29, LMWP and dysferlin. The PLA2 approach was therefore discontinued.
Confocal microscopy of adult worm sections revealed that host CD44 was confined entirely to the tegument with no staining of internal structures (Figure 4A). Examination of a Z-stack through the tegument of an intact male worm revealed that the pattern of staining was not uniform, being concentrated on numerous, parallel, transverse ridges and especially the spines on the dorsal tubercles (Figure 4B). On close inspection, the individual spines had a definite inverted V appearance the most intense staining being at the tip (Figure 4C). Antibodies to CD48 and CD90 failed to detect their respective targets on the worm surface.
Our in-house EST database was interrogated for the occurrence of transcripts in lung schistosomula encoding the proteins of interest released from adult worms by the enzyme treatments. Although it is difficult to make inferences about abundance as some data were obtained from normalised libraries, nearly all transcripts were represented in both larvae and adults in roughly similar numbers (Table 2). The exceptions were the proteoglycan (5L and 0A) and ADP-ribosyl cyclase (0L and 1A). Transcripts for calpain and Sm200 were particularly abundant, whilst the largest number recorded (63) was for one of the CD59 orthologues in the lung schistosomulum. Based on the transcript evidence we tentatively conclude that the most exposed tegument surface proteins of adults are also likely to be present on the surface of the migrating lung schistosomulum.
Only one tegument surface protein, Sm200, was released by both the trypsin and PiPLC treatments of live worms. The difference between the two was that PiPLC released the entire GPI-anchored molecule for subsequent trypsinisation as a separate step. Conversely the trypsin treatment released peptide fragments from Sm200 molecules only during the incubation period, which were then recovered using reversed-phase chromatography. Mapping of the peptide hits onto the primary amino acid sequence revealed a very different pattern of distribution (Figure S3). A total of 15 peptides, with a significant Mascot score, was identified for the PiPLC-released protein, distributed throughout the entire molecule. In contrast only six peptides were identified for the trypsin-released protein and four of these were clustered towards the C-terminus of the protein, i.e. the region located nearest to the GPI anchor. This may indicate that only part of the Sm200 protein is accessible to trypsin in the live worm. In this context, the programs NetNGlyc and NetOGlyc both at http://www.cbs.dtu.dk/services/ predicted five N (scores >0.6) but no O-glycosylation sites on the amino acid sequence coding for Sm200. It is notable that four of the five potential N-linked sites are located furthest from the GPI anchor site in the N-terminal region where few peptides were identified following trypsin treatment of the live worms. Thus it is plausible that attached N-glycans protect the native Sm200 polypeptide chain in situ on the worm surface from trypsin attack.
Our previous biotinylation studies have provided insights into the disposition of proteins in the complex molecular structure of the schistosome tegument surface. In the present study we have taken an alternative approach to obtain information about the location of tegument components. This involved shaving the surface of live adult worms with selected enzymes to release accessible proteins, and their identification by proteomics. It is axiomatic in this approach that the integrity of the worm surface is not compromised by the treatment. Our results indicate that trypsin and PiPLC largely fulfilled this criterion but PLA2 did not. We must assume that this last enzyme, via its attack on the lipid bilayers, rapidly caused generalized erosion of the surface membranes and loss of integrity. We shall therefore consider only the results of the trypsin and PiPLC treatments to make inferences about tegument surface organization. In addition, the male parasite has at least two times the surface area of the female, assuming all surfaces are accessible to the enzymes. If the gynaecophoric canal is effectively sealed, then the male dorsal surface would have contributed the bulk of released protein.
It was inevitable that the in vitro culture of live worms resulted in the contamination of the enzyme-released proteins by gut content. Thus the detection of proteins such as α2-macroglobulin, saposin B, and hemoglobinase (asparaginyl endopeptidase) in the two enzymatic shaving experiments was to be expected. Conversely, even with the most careful handling of parasites including their recovery from mice by perfusion with culture medium, the presence of cytosolic and cytoskeletal proteins must be attributed to some degree of worm damage. In the present study, we used the iTRAQ labelling technique on test and control samples to discriminate enzymatic enrichment from normal secretion or protein leakage due to invisible damage. This approach successfully overcame the dominance of abundant peptides from contaminating cytosolic and cytoskeletal proteins to highlight the few proteins released by PiPLC treatment. Nevertheless, one must bear in mind that as enzyme accessibility has not been addressed in this investigation, caution should be exercised when considering the fold enrichment found for GPI-anchored molecules. In this regard, it is possible that a higher fold enrichment for a given protein may only indicate a more exposed/accessible location at the parasite surface rather than imply protein abundance.
A major finding of our study is that seven parasite proteins can be detached from the surface by treatment with PiPLC and enriched compared to control incubations. From this we infer that each is inserted into the parasite surface by a GPI-anchor. As these anchors are invariably located at the C-terminus of such proteins we conclude that the extraneous 30 kDa PiPLC enzyme is able to access the anchor in proximity to a lipid bilayer. Whether this bilayer is the secreted membranocalyx or the underlying plasma membrane is problematic. Access to the latter would mean that the digesting enzyme had to pass through the protective membranocalyx. One possible inference is that the seven GPI-anchored proteins are inserted into the membranocalyx. For carbonic anhydrase this seems unlikely because of its assumed function in regulating acid-base balance at the parasite surface. It would be best placed to do this if located between the two lipid bilayers so that, by analogy with the erythrocyte, CO2 generated inside the worm could be converted to HCO3− for diffusion into the bloodstream. Indeed, the reverse diffusion of a chloride ion through the membranocalyx would be required to balance the charge, bringing it into close proximity with the anion transporters, which our previous studies have shown are present in the tegument plasma membrane [6]. The gene model for carbonic anhydrase is incomplete, lacking both the N-terminal exon encoding the signal peptide and the C-terminal exon(s) encoding the site for the attachment of a GPI-anchor. However, its enrichment by PiPLC treatment of live worms attests to the presence of a GPI-anchor. Two isoforms of CD59 orthologues, identified by possession of the CCxxDxCN motif, were also highly enriched by the PiPLC treatment, implying their outer location at the surface. These proteins are potentially significant because the human CD59 protein protects self-cells against complement fixation by blocking formation of the C5 to C9 membrane attack complex [11]. A similar role for the two schistosome molecules is very attractive, as a component of the parasite's mechanisms of immune evasion. It is of note that in the S. mansoni genome there are four additional gene models encoding CD59 orthologues, which could also be tegument-associated [12]. These CD59 orthologues in the tegument surface are better candidates for inhibition of complement fixation than the SCIP-1 protein [13], subsequently identified as paramyosin [14], [15].
The protein most highly enriched by PiPLC treatment was Sm29 that has no known function. It was originally selected by a bioinformatic search as a protein with a single membrane spanning domain [16]. It was identified in the final pellet after differential extraction in a compositional analysis of the tegument membranes [6] and was accessible to biotinylation in live worms [7]. Subsequent confocal microscopy revealed its association with the tegument although technical issues do not allow a firm conclusion about its precise location [17]. Our PiPLC shaving results demonstrate both its peripheral location and the fact that it is GPI-anchored, the latter prediction also made by big-PI Prediction server (http://mendel.imp.ac.at/gpi/gpi_server.html). The peripheral location of Sm200, again with no known function, was also revealed by PiPLC shaving. This protein first cloned by Hall et al. [18] was already known to be GPI-anchored and located in the tegument surface [19], [20]. Surprisingly therefore it was not found in compositional analysis of the tegument membrane [6], but was accessible to biotinylation in live worms [7]. These observations suggest that whilst definitely surface-attached it is readily lost during the processing of tegument membranes by differential extraction for MS analysis, a feature that may be linked both to its size and GPI-anchor. Moreover, Sm200 has recently been identified in circulating lipoprotein particles from the blood of schistosome-infected humans, confirming its turnover into the vascular environment [21].
The two remaining GPI-anchored proteins shown to be enriched by iTRAQ labelling after PiPLC treatment were alkaline phosphatase and ADP-ribosyl cyclase. The former is well characterized and was previously shown to be GPI-anchored in schistosomula [22]. It has been used as a membrane marker in the development of methods for tegument surface isolation [5]. It was identified by compositional analysis of the tegument surface membranes in the urea/thiourea/CHAPS/sulfobetaine (UTCS) fraction and insoluble pellet [6], and is accessible to biotinylation in live worms [7]. Its presence in the UTCS fraction indicates that it is, at least partially, loosely associated with the surface membranes i.e not membrane spanning. The recently characterised ADP-ribosyl cyclase was also shown to be GPI-anchored and localized to the outer tegument of the adult schistosome [23]. Both schistosome enzymes must be able to access their substrates in live worms but their precise function at the tegument surface remains to be established. In the case of ADP-ribosyl cyclase, a role in calcium mobilization has been proposed [23] but alternatively it could function in immune evasion by regulating ecto-NAD+ levels, thereby reducing substrate availability for CD38- and CD157-mediated effector functions of lymphocytes. Corroborating the identification of GPI-anchored molecules using the iTRAQ technique, four out of the seven proteins assigned as GPI-anchored were also detected by our 2-DE analysis. These were Sm200, alkaline phosphatase, two isoforms of CD59 orthologues and carbonic anhydrase. A definite proof of their enrichment due to PiPLC's activity is the fact that these molecules are underrepresented and are not easily identifiable in 2-DE maps produced using soluble worm preparations [8] or membrane-extracted proteins from crude or differentially-extracted S. mansoni tegument [6].
Use of the iTRAQ technique to identify proteins enriched by trypsin treatment was not possible because the released material was already partially digested. This prevents accurate quantification of protein for labelling and therefore the trypsin shaving fraction was subsequently processed as a peptide digest. We have taken as a reference of cytosolic/cytoskeletal contaminants in the trypsin shaving, the number and diversity of those proteins released during 1 h incubation at 37°C in the absence of added enzyme. When we compared the number of proteins that could indicate leakage or vomitus with the number of proteins in the trypsin-treated parasites, we observed we had less contamination, most likely due to a shorter incubation time, 30 min. In addition, we are assuming that as trypsin should not be able to permeate the parasite tissues, the peptides originating from membrane proteins are likely to represent the most exposed/accessible domains at the parasite surface. We therefore focus only on (1) membrane and membrane-associated, (2) vesicular pathway and secreted, and (3) host proteins. The peripheral location of Sm200 was again confirmed but it was the only one of the seven PiPLC-released proteins that was also released by trypsin. This suggests that the rest may be protected from proteolysis, potentially by their N and/or O-linked glycans or by a sequestered location. The release of three annexins by the trypsin treatment indicates their superficial location. One (Smp_077720) was previously detected by both biotinylation and compositional analysis and two are new to this study (Smp_074140 and Smp_074150). The known phospholipid-binding properties of this group of proteins means that they could have a role in promoting adhesion of the membranocalyx to the plasma membrane, acting like a molecular ‘velcro’ via the four binding domains that each possesses. More recently, certain annexin isoforms have been implicated in immunomodulatory functions such as the resolution of inflammation [24]. They are able to interact with receptors on the surface of leukocytes to control apoptosis and their clearance by macrophages. The existence of such a process at the surface of the tegument, mediated by schistosome annexins, accords with the absence of leukocyte binding to adult worms in vivo [25]. However, a comparison of the orthologies of the schistosome and human annexins is difficult because of the evolutionary distance. Thus, BLAST searching of the schistosome sequences against the NCBInr database reveals the closest homologues of Smps 077720, 075150 and 075140 are human annexins A13, A7 and A8 respectively, not the A1 isoform that has been most implicated as an anti-inflammatory agent. The conjecture will only be resolved by expression of the schistosome annexins for assays of function.
The presence of Sm25 as a tegument surface protein has a chequered history. It was proposed as a vaccine candidate because anti-Sm25 antibody levels correlated with protection in mice vaccinated with a crude tegument membrane preparation [26]. It was then cloned and designated as an N-glycosylated integral membrane protein [27] but later characterized as a palmitoylated protein with the implication that it was on the cytosolic leaflet of the plasma membrane [28]. Further immunocytochemical studies suggested it was distributed throughout the tegument syncytium but not associated with the surface membranes [29] and it could only be biotinylated when parasites were permeabilized by Triton X-100 [30]. It was not found in either our compositional or biotinylation studies on the tegument surface. The removal of Sm25 from live worms by trypsin, when so few other membrane proteins are released, suggests a unique accessibility. The proteolytic enzyme calpain was previously identified at the tegument surface by both compositional analysis [6] and biotinylation [7]; its potential as a vaccine candidate has already been exploited [31], [32]. The final member of this group is a heparin sulphate-like proteoglycan which has not been identified in any previous studies and merits further investigation. In the secreted protein category, this is the first report of SmKK7 release from the tegument surface; it has previously only been reported from cercarial secretions [33]. The detection of a pair of BAR domain proteins in all four experiments after trypsin treatment is intriguing since these banana-shaped molecules located on the cytosolic surface of plasma membranes function to promote membrane curvature in exo- and endocytosis [34]. This suggests that the trypsin may be entering the fine tubules located at the base of tegument pits where multilaminate vesicle fusion with the tegument plasma membrane occurs [35], [36]. That conclusion would place Sm25 and the proteoglycan in a similar location where constituent proteins are poorly protected by overlying membranocalyx, perhaps loosely analogous to the relationship of the flagellar pocket of trypanosomes [37] to variant surface glycoprotein export.
The release of complement C3 from the parasite surface by trypsin treatment is consistent with our previous biotinylation study [7]. We can now add C4, its precursor in the complement cascade, but not the C5 to C9 components of the membrane attack complex, suggesting complement fixation but then inhibition of the cascade. The failure to detect any immunoglobulins, when murine IgM, IgG1 and IgG3 heavy chains were biotinylated [7], is puzzling as they would be expected to initiate complement fixation. However, there is evidence that these proteins are resistant to trypsin degradation, particularly under the low concentration conditions employed in our experiments [38]. In fact, aiming to preserve worm viability, trypsin was used at a concentration approximately 100 times lower than that typically employed for stripping adherent cells from culture flasks. A positive consequence of such a gentle shaving treatment on live worms is justified by the reasonable number of genuine membrane proteins that were identified. In contrast, a previous study on S. bovis with trypsin treatment, used methanol-fixed parasites [39] and found an overall higher number of protein identifications. However, membrane-associated and/or integral proteins were poorly represented, and the results are not comparable with our approach using live worms.
Our enzyme shaving experiments extend the list of host proteins known to be firmly associated to the parasite surface. The release of CD48 and CD90 (Thy1.2) by PiPLC confirms their possession of GPI-anchors and most likely accounts for their transference from the host (leukocytes) in the same manner as host glycolipids transfer from erythrocytes through a process currently termed cell-painting [40]. Supporting this finding is the demonstrable ability of purified Thy-1to reincorporate into the plasma membrane of murine Thy-1- cells directly from aqueous suspension and without the use of detergents [41]. However, failure to detect the presence of both CD48 and CD90 by immunocytochemistry indicates their relative paucity. The detection of host CD44 is more surprising since it is an extracellular protein anchored in the membrane of a range of cell types by a single transmembrane domain [42]. However, the intense staining of the tips of spines on the dorsal tubercles, the point of contact between the worm and the vascular endothelium during peregrination around the portal vasculature, suggests that the transfer occurs when the apposing membranes of tegument and endothelium are pressed together. The observations on CD44 staining are a testament to the acquisitive properties of the schistosome surface.
Our studies on the surface accessibility of tegument proteins are relevant to the development of schistosome vaccines. In mice immunized with attenuated cercariae challenge schistosomula are eliminated in the lungs by inflammatory foci that block their onward migration [43]. Both priming of the immune response [44] in skin-draining lymph nodes and the pulmonary effector responses appear to be mediated by proteins on the schistosomular tegument surface [45]. Although we have characterised surface proteins on the adult worm, the presence of the encoding mRNAs in the lung schistosomulum provides support for the suggestion that those same proteins are exposed on the larval tegument surface. In this context Sm29 has already been used successfully as a vaccine candidate [17]. We therefore suggest that the other six PiPLC-anchored proteins would also repay investigation as putative vaccine candidates, especially Sm200 because it was also removed by trypsin. Indeed, the presence of a GPI-anchor may make proteins in this group more prone to detachment from the surface for transcytosis across the pulmonary capillary endothelium, processing by accessory cells and presentation to reactive CD4+ Th1 cells in the lung parenchyma. To this list we can add the three annexins released by trypsin and LMWP that may be a tegument secretion, noting also that calpain is a proposed vaccine candidate [31]. It is conceivable that a cocktail of these proteins would elicit strong protection against intravascular migrating schistosomula, especially if targeted to the lungs, whereas immunisation with a single protein has only modest success [46].
|
10.1371/journal.pntd.0002263 | Bacillus thuringiensis-derived Cry5B Has Potent Anthelmintic Activity against Ascaris suum | Ascaris suum and Ascaris lumbricoides are two closely related geo-helminth parasites that ubiquitously infect pigs and humans, respectively. Ascaris suum infection in pigs is considered a good model for A. lumbricoides infection in humans because of a similar biology and tissue migration to the intestines. Ascaris lumbricoides infections in children are associated with malnutrition, growth and cognitive stunting, immune defects, and, in extreme cases, life-threatening blockage of the digestive tract and aberrant migration into the bile duct and peritoneum. Similar effects can be seen with A. suum infections in pigs related to poor feed efficiency and performance. New strategies to control Ascaris infections are needed largely due to reduced treatment efficacies of current anthelmintics in the field, the threat of resistance development, and the general lack of new drug development for intestinal soil-transmitted helminths for humans and animals. Here we demonstrate for the first time that A. suum expresses the receptors for Bacillus thuringiensis crystal protein and novel anthelmintic Cry5B, which has been previously shown to intoxicate hookworms and which belongs to a class of proteins considered non-toxic to vertebrates. Cry5B is able to intoxicate A. suum larvae and adults and triggers the activation of the p38 mitogen-activated protein kinase pathway similar to that observed with other nematodes. Most importantly, two moderate doses of 20 mg/kg body weight (143 nM/kg) of Cry5B resulted in a near complete cure of intestinal A. suum infections in pigs. Taken together, these results demonstrate the excellent potential of Cry5B to treat Ascaris infections in pigs and in humans and for Cry5B to work effectively in the human gastrointestinal tract.
| Ascaris suum is an intestinal parasitic nematode of pigs that is very closely related to Ascaris lumbricoides, a major intestinal parasitic nematode of humans that infects more than one billion people worldwide. Because of reduced efficacy and the threat of resistance to the current small set of approved drugs to treat Ascaris infections, new treatments are needed. Here we test against A. suum infections the effectiveness of Cry5B, a nematode-killing protein made by the natural soil bacterium Bacillus thuringiensis and representing a promising new class of anthelmintics. We demonstrate for the first time that A. suum possesses the receptors to bind Cry5B and that Cry5B can kill A. suum larvae and adults in culture. Most importantly, we demonstrate that oral administration of Cry5B to pigs infected with A. suum larvae results in a near complete elimination of the infection. Given the similarities between A. suum and A. lumbricoides and the similarity between the pig and human gastrointestinal tracts, our data indicate that Cry5B has excellent potential to treat Ascaris infections in veterinary animals and in humans.
| Ascaris lumbricoides, the large roundworm, is the most common parasitic nematode infection of humans with estimates of more than one billion people infected worldwide in mostly lesser developed countries [1]. Chronic infections in humans are associated with growth and cognitive stunting, impaired nutritional status, and dysfunctional immune regulation related to a polarized Th2 immunity and altered protective antibody responses to bacterial vaccination [1], [2], [3], [4]. More acute symptoms associated with larval migration through lung tissue involve inflammation, breathing difficulties, and fever [4], [5]. Adult Ascaris infections can lead to abdominal pain, nausea and diarrhea, and potentially life-threatening blockage of the intestinal or biliary tracts [4], [6]. Although treatable with approved anthelmintics, the arsenal of available drugs against nematodes is largely limited to classes of benzimidazoles, macrocyclic lactones, and imidazothiazoles, and Ascaris drug resistance in human therapy is a rising concern [7], [8], [9], [10].
Ascaris suum is a common infection in pigs and a significant and costly management problem in confinement and free-range facilities because of high worm fecundity and resistance of infective eggs to environmental stressors [11]. Adverse effects in pigs are similar to those in humans and include migrating larvae-induced liver scaring (white spots), pulmonary inflammation, reduced nutritional absorption and feed efficiency, altered responses to bacterial vaccination, and a polarized Th2 immunity [4], [12], [13], [14], [15]. Ascaris suum and A. lumbricoides are genetically similar and there is currently some evidence that they are the same species [16].
Crystal (Cry) proteins made by the soil-bacterium Bacillus thuringiensis (Bt) are vertebrate-safe proteins used extensively and successfully on organic and conventional agriculture, in transgenic food and non-food crops, and in vector-control programs to kill insect pests [17], [18], [19]. Cry proteins act as ingested pore-forming proteins that target invertebrate-restricted intestinal receptors [20]. The Cry protein Cry5B has been shown to be effective against two intestinal parasitic nematode infections in rodents when administered orally via gavage [21], [22], [23]. Specifically, a single dose of 100 mg/kg body weight (BW) of Cry5B is able to eliminate 90% of Ancylostoma ceylanicum hookworm parasites from infected hamsters (A. ceylanicum is a zoonotic hookworm species capable of infecting humans and causing hookworm disease [24]) and eliminate 70% of Heligmosomoides bakeri (previously designated as H. polygyrus) parasites from the intestine of mice (H. bakeri is a murine-specific parasitic nematode [25]). The receptors for Cry5B in nematodes are invertebrate-specific glycolipids present on the nematode intestinal mucosa [21], [23], [26], [27], [28], [29], [30]. Both H. bakeri and A. ceylanicum are phylogenetically related Clade V nematodes [31]. Here we test for the first time if A. suum, a Clade III nematode [31], possesses Cry5B-binding receptors and if Cry5B is able to intoxicate A. suum in vitro. We further test Cry5B therapeutic efficacy in vivo against an experimental A. suum infection in pigs, a comparable in vivo model for both A. lumbricoides infections in humans and for the efficacy of Cry5B protein in the human gastrointestinal (GI) tract.
All animal experiment was carried out under protocols approved by the Beltsville Area Animal Care and Use Committee (BAACUC), protocol number 10-012. All housing and care of laboratory animals used in this study conform to the NIH Guide for the Care and Use of Laboratory Animals in Research (see 18-F22) and all requirements and all regulations issued by the USDA, including regulations implementing the Animal Welfare Act (P.L. 89-544) as amended (see 18-F23).
Adult A. suum female worms were obtained from naturally infected sows collected at the Beltsville Agricultural Research Center Abattoir. The proximal 2 cm of uteri from each female was dissected and aseptically transferred to 100 mm culture dishes where the pooled uteri were macerated with a syringe plunger and the mixture passed through a sterile 100 um nylon sieve (BD Falcon #352360, Bedford, MA) to provide a single suspension of eggs. The egg suspension was washed with PBS containing 200 µg/ml penicillin, 200 units/ml streptomycin and 10 µg/ml amphotericin B (antibiotics) three times by centrifugation at 200×g after removal of intervening supernatant washes by suction. The final suspension was placed into 100 mm culture dishes and maintained at ambient room temperature for at least 60 days (with occasional washing of the eggs) to embryonate the eggs to an infective stage. The suspension was counted for eggs with well-formed and ensheathed larvae and stored at 4°C until used.
Experimental pigs were obtained from a facility at the Beltsville Agricultural Research Center, Beltsville, MD. Pigs were derived from boars from a four-way crossbred composite BX line (Duroc X maternal Landrace X terminal Landrace X Yorkshire) designed by scientists at the USDA/ARS/US Meat Animal Research Center, Clay Center, NE to be genetically similar to genetics in the commercial swine industry at the time they were born; the genetics of the gilts are predominantly of the BX composite line. Pigs were from a herd screened yearly for porcine reproductive and respiratory syndrome virus (PRRSV), influenza (H1N1 and H3N2), pseudorabies, and brucellosis by the Veterinary Services Group at the Beltsville Agricultural Research Center and have been negative for these infections. They were individually housed in stalls with a non-absorptive concrete floor surface with ad libitum access to water and fed a nutritionally adequate corn/soybean-based feed once per day.
Pigs used to collect larval and adult stages of A. suum were at least eight weeks of age when inoculated per os with a suspension of approximately 10,000 to 30,000 infective eggs. To collect the parasitic and adult stages of A. suum for in vitro testing with Cry5B, infected pigs were euthanized using Euthasol (Virbac AH, Inc., Fort Worth, TX) with doses described by the manufacturer at various times after inoculation and the entire small intestine was removed. The intestinal tract was opened with a scissor and the contents and mucosa rubbed between two figures into a large vessel over a slow stream of warm tap water. Larvae were isolated by an agar-gel technique [32]. Briefly, the intestinal suspension was mixed with an equal volume of molten 2% agar and poured into pans containing a cheese cloth to form a solid gel which was then placed into containers of 0.85% NaCl (saline) warmed to 37°C and incubated for 2 hrs. After removal of the gel from the holding container, larvae were isolated by decanting the fluid and washing the larvae with several volumes of warm saline and settling in conical glasses for several repetitions. The larval suspension was then transferred to sterile 50 ml culture tubes and aseptically washed (repeated settling of larvae and decanting of supernatants) with warm RPMI1640 media containing antibiotics five times followed by a one hr incubation at 37°C and then five additional washes. The larval suspensions were shipped overnight from Beltsville, MD to La Jolla, CA in sealed 50 ml tubes with media containing 10% FBS. Adult worms were collected from intestinal contents passed over copper separation screens, washed and shipped similar to the larval stages.
For glycolipid experiments, Caenorhabditis elegans was grown and prepared as described [30]. Early fourth-stage A. suum larvae (L4) collected from the intestine at 14 days post-inoculation (PI) were washed in water and resuspended in three pellet volumes of water. The A. suum pellet was shock frozen with liquid nitrogen in a porcelain mortar, and the frozen pellet was ground with a pestle until it thawed. After repeating the freezing and grinding steps four times, the suspension was sonicated four times. Complete membrane disruption was verified by microscopy. Upper phase glycolipids were purified based on the Svennerholm partitioning method [33], except that upper phase A. suum glycolipids were applied on a tC18 silica cartridge (Millipore) four times (instead of two). The thin-layer chromatography (TLC) assay was carried out as previously described in [30]. Briefly, purified upper phase glycolipids were separated on a HPTLC plates in a chamber filled with 4∶4∶1 chloroform∶methanol∶water. Developed plates were either stained with a non-specific orcinol stain to ensure comparable glycolipid loading of different nematodes or probed with biotinylated Cry5B [29]. The TLC Cry5B competition assays were undertaken in the presence of 100 mM glucose or galactose. Cry5B binding was visualized with the help of the avidin/alkaline phosphatase (Reagent A/B) Vectastain ABC-AP kit. The experiment was performed three times, with a representative trial shown here.
For phospho-p38 experiments, day 14 L4 were exposed to either no protein or 100 µg/mL Cry5B (10 larvae per condition) for 2 hr (37°C, 5% CO2) in RPMI 1640 plus 5% fetal bovine serum plus antibiotics (100 U/mL penicillin, 100 µg/mL streptomycin; 0.25 µg/mL fungazole). Harvesting and processing of larvae and Western blotting were carried out as described [23]. The experiment was independently performed a total of four times with similar results; one representative result is shown. The intensities of the bands were quantitated using NIH Image J (v1.41). The intensities of the α-tubulin bands were very similar and were used to normalize the intensities of the phospho-p38 bands.
The medium used for in vitro assays was the same as for p38 phosphorylation experiments. In vitro assays with L4s were carried out in a 500 µl volume in a 24-well plate. There were 5–6 L4 of mixed genders per well because of the difficulty to discern larval gender at this stage of development. In each independent trial, there were two wells/dose (10–12 L4s total). For day 14 PI larvae, three independent trials were performed; for day 19 PI larvae, two independent trials were performed. For control wells, 10 µL 20 mM Hepes pH 8.0 buffer was added in place of Cry5B (always added as 10 µL in 20 mM Hepes pH 8.0; final concentrations = 0.1 µg/mL, 1.0 µg/mL, 10.0 µg/mL, 100.0 µg/mL, and 1000 µg/mL). For day 14 PI larvae, the effective dose 50% (ED50) value was calculated at day 4 (half way through the experiment) and day 7 (completion of experiment) using PROBIT [34]. For day 19 PI larvae, we did not calculate ED50 values as all the doses, by and large, behaved similarly.
Adult worm assays were carried out in 125 mL or 250 mL flasks (buffer volume 10 mL or 20 mL respectively) with one worm/flask (2 worms/dose/trial with two independent trials). For control flasks, 20 mM Hepes was added in place of Cry5B (one dose, 100 µg/mL). The medium in the flasks was refreshed/changed every two days. Plates or flasks were put at 37°C in a 5% CO2 in air incubator. The assays were scored daily over the period of one week. Larvae that moved in the absence or presence of gentle touch with an eyelash were scored as alive; larvae that did not move were scored as dead. A motility index score on a scale from 3-to-0 was used to determine adult worm viability: 3 represents a parasite with vigorous movement similar to control no drug at the start of the experiment; 2 represents a parasite with whole-body movements (seen without external stimulus) significantly slower than control no drug at the start; 1 represents a parasite that was not moving on its own but moved when prodded with a wooden inoculating stick (tested at three different body locations); and 0 represents a worm that did not move even when prodded.
To determine the in vivo efficacy of Cry5B against A. suum, a total of ten recently weaned four week old pigs from two different litters were split into two groups of five pigs each. All ten pigs were inoculated per os with 5,000 infective A. suum eggs and at ten and 12 days PI five pigs (placebo group) were given per os a suspension of spore lysate (HD1 parent strain) and five pigs (Cry5B-treated group) were given spore crystal lysate (HD1 parent strain with Cry5B-expressing vector) at a final dose of 20 mg/kg BW of Cry5B. HD1 Bt lysates with and without Cry5B were prepared as described [22]. Spore counts were checked and were found similar for both groups. Pigs were euthanized on day 15 PI with infective eggs and L4 isolated from the intestinal contents by the gel-agar method. The total number of L4 from each pig was counted under a dissecting microscope and average number of L4 recovered from each group were calculated and compared for a statistical difference using Student's t-test assuming unequal variances. The recovered L4 were fixed with 10% buffered-formalin and the length of each larva was subsequently measured microscopically using the Olympus MicroSuite – B3SV system at 40× magnification (Olympus America, Melville, NY). Briefly, aliquots of the larval suspension were transferred to a microscope slide and covered with a glass slip. The slide was scanned and at least 10 larvae from each pig were manually traced using a computer mouse. The length of each larva was calculated in µm using calibrated software and the average ± the standard error of the mean recorded for the two treatment groups.
To determine if A. suum expresses Cry5B-binding glycolipid receptors, glycosphingolipids were isolated from A. suum L4 and separated by thin-layer chromatography (Fig. 1A). These glycolipids were then probed with protease-activated, biotinylated Cry5B. Ascaris suum L4 contain several Cry5B-binding glycolipids that show a profile somewhat different than that of C. elegans (Figure 1B). The specificity of the Cry5B binding was determined by competition experiments. As previously found with C. elegans, A. ceylanicum, and P. pacificus [21], [27], [30], galactose, more than glucose, is able to compete binding of Cry5B to glycolipids from A. suum (Figure 1B).
After determining that A. suum expresses Cry5B receptors, we asked whether parasite viability is susceptible to Cry5B in vitro. Larvae and adults were isolated from the pig small intestine at 14 days PI (early stage L4), 19 days PI (mid stage L4), and 54 or 89 days PI (adult stage) [35]. Worms were incubated in culture media with various doses of Cry5B and scored daily for intoxication based on motility after gentle physical touching. Conditions were repeated either twice (day 19 PI L4 or adults) or three times (day 14 PI L4). At all parasitic stages, measurable intoxication was seen with time in culture (Figure 2A, B, C). Day 14 L4s displayed a nice dose-dependent response to Cry5B. For day 14 L4s, we calculated an ED50 value at day 4 of 1.1 µg/mL (95% confidence interval 0.30–3.1) and at day 7 of 0.094 µg/mL (95% confidence interval 0.011–0.30). Day 19 L4s were even more sensitive to Cry5B, with quicker intoxication even at the lowest doses used.
Adult parasites exposed to Cry5B were largely immotile by day 4 (index score 1), while controls were still vibrant (index score 3). The difference is most striking by day 6. On day 6, control adult parasites were mostly very healthy – three out of four were fully motile, all scoring as a 3, and one out of four moved only when touched, scoring as a 1. Conversely, on day 6 all four of the adult parasites in Cry5B were apparently dead (did not move even when touched multiple times). The two groups are statistically different on day 6 (P = 0.018, 2-sample Kruskal-Wallis Test). The data demonstrate that A. suum is biologically susceptible to Cry5B as L4 and adult stages from the intestine.
A molecular hallmark of nematode intoxication by Cry5B is activation of the p38 mitogen-activated protein kinase (MAPK) pathway upon exposure to the pore-forming protein [23], [36], [37]. To ascertain whether intoxication of A. suum by Cry5B results in a similar response, A. suum L4 were exposed to Cry5B and probed for phosphorylated (activated) p38 levels by immunoblotting. As before, total protein levels were controlled using an α-tubulin antibody [23], [36], [37]. As shown in Figure 3, treatment of A. suum with Cry5B results in significant up-regulation of phospho-p38. Densitometric analyses indicate that this up-regulation is 4.3 fold.
The therapeutic efficacy of Cry5B in vivo was tested in pigs infected with A. suum. Ten pigs were each inoculated with approximately 5,000 infective A. suum eggs. Five infected pigs (Cry5B-treated group) received two 20 mg/kg BW (143 nM/kg BW) doses of Cry5B via gavage, one dose each on days 10 and 12 PI (Cry5B was administered as a spore-crystal lysate; [22]); the other five infected pigs (placebo group) received one dose of a spore lysate without Cry5B on those same days. On day 15 PI, the pigs were euthanized and worm burdens (L4) in the small intestine were assessed. The two-dose Cry5B treatment resulted in a near complete (97%) elimination of A. suum L4 (Figure 4). All pigs appeared clinically normal after egg inoculation and spore lysate treatments. In addition, microscopic tracings of the length of individual A. suum L4 recovered from Cry5B-treated pigs averaged 3937±86 µm and was significantly different from an average length of 5796±205 µm from the placebo-treated pigs, suggesting that the growth of the few L4 remaining in the intestine was severely affected by Cry5B. These results indicate Cry5B has great potential to treat Ascaris infections in humans.
Here we demonstrated for the first time that a Bacillus thuringiensis crystal protein (Cry5B) can intoxicate Ascaris suum parasites in vitro and can reduce growth and facilitate expulsion of parasites from the intestine of pigs in vivo. Intoxication in vitro was observed for the fourth-stage larvae (L4) and adults isolated from the pig intestine. Consistent with its ability to intoxicate A. suum, Cry5B glycolipid receptors were isolated from A. suum larvae. Furthermore, as seen previously with C. elegans, A. ceylanicum, and P. pacificus, receptor binding was inhibited by galactose competition. Similar to results with C. elegans and A. ceylanicum, treatment with Cry5B also results in significant up-regulation of phospho-p38 MAPK. Taken together, these data suggest that Cry5B intoxicates A. suum in a manner similar to other nematodes.
That A. suum expresses as least one Cry5B-binding glycolipid species could be predicted from the literature. The phosphorylcholine substituted glycolipid species in C. elegans [Gal(β1-3)Gal(α1-3)GalNAc(β1-4)[PC-6]GlcNAc(β1-3)Man(β1-4)Glc(β1-1)ceramide] that binds Cry5B [27] is perfectly conserved in A. suum [38]. This species is apparent as a Cry5B-binding glycolipid “E” in our overlay (Figure 1B; [27]) and is also found in the human parasite A. lumbricoides [39].
Strikingly, two 143 nM/kg BW doses of Cry5B were able to nearly completely (97%) eliminate A. suum L4 from the intestines of infected pigs (for comparison 143 nM/kg BW of albendazole would equate to 0.04 mg/kg BW). The dose Cry5B used is within the range of other commercially available swine anthelmintics (dichlorvos = 6–9 mg/kg BW; pyrantel pamoate = 22 mg/kg BW; levamisole = 8 mg/kg BW), some also used to eliminate parasitic larval stages of A. suum (fenbendazole = 5–10 mg/kg BW for three days) [40]. The timing of Cry5B treatment of pigs was targeted because the early L4 is metabolically similar to the adult [41] and worm accumulation is stable in the intestine between 10 and 17 days after inoculation, a period that precedes a self-cure response by the pig that significantly reduces the number of L4 in the intestine [42]. It is currently impractical to evaluate the efficacy of Cry5B against adult A. suum in vivo because of limited current capacities to produce Cry5B in quantities to dose the large number of pigs needed to establish adult worm infections after experimental inoculation and the weight of rapidly growing pigs needed for a minimum of 49 days after inoculation when the worms would become patent and produce eggs. Future studies of the minimum dose and frequency of treatment and the evaluation of tolerability are planned as Cry5B delivery and production systems are optimized and become more cost effective.
This in vivo result has several major implications. First, our results importantly show that Cry5B can affect parasitic nematode expulsion in pigs, which possess an intestinal tract functionally similar to that of humans [43], [44]. This finding further sustains the potential of Cry protein anthelmintics for human therapy.
Second, this current study and earlier experiments with infection of hamsters with A. ceylanicum [21], [23] demonstrate that Cry5B can significantly bind to and facilitate expulsion from the host of two of the three major intestinal parasitic nematode infections in humans (hookworms and the large roundworm Ascaris). This result is important because of the limited arsenal of anthelmintics available to treat human parasitic nematode infection and newer anthelmintics like monepantel, developed to treat gastrointestinal nematode parasites in sheep, have limited efficacy against Ascaris [45].
Third, A. suum is in a phylogenetic group (Clade III) distinct from the two other intestinal nematodes tested in vivo for Cry5B anthelmintic activity to date, A. ceylanicum and H. bakeri (both Clade V) [31]. It has also been previously shown that Cry5B intoxicates the plant-parasitic nematodes of the genus Meloidogyne, which belong to Clade IV [46], [47]. The data presented here are the first known describing susceptibility of a Clade III nematode to a Bt crystal protein and, together with previous results, suggest that Cry5B can broadly impact parasitic nematode infections.
Fourth, our results demonstrated that Cry5B has excellent utility in treating nematodes of veterinary importance, e.g., A. suum infection in pigs. Because Cry5B works mechanistically different from other approved anthelmintics [34], Cry5B could be used in places where there is resistance to current anthelmintics. Of particular interest is the observation that that a Bt strain expressing Cry5B intoxicated Haemonchus contortus parasites in vitro [48]. This is an important because there is world-wide resistance of H. contortus to conventional anthelmintics used in sheep and goats and new anthelmintics are needed to treat economically important ruminants [49]. In addition, anthelmintic resistance in the equine ascarid parasite Parascaris equorum is a growing concern [50] and the utility of testing for Cry protein efficacy against an important parasitic ascarid of horses is more economically feasible in swine as a physiologically similar non-ruminant host species.
In summary, these data demonstrate the excellent potential of Bt Cry proteins as a major new class of anthelmintics to treat mammals of veterinary importance and, most importantly, for the two billion humans infected with intestinal nematodes.
|
10.1371/journal.pgen.1007707 | A survey of inter-individual variation in DNA methylation identifies environmentally responsive co-regulated networks of epigenetic variation in the human genome | While population studies have resulted in detailed maps of genetic variation in humans, to date there are few robust maps of epigenetic variation. We identified sites containing clusters of CpGs with high inter-individual epigenetic variation, termed Variably Methylated Regions (VMRs) in five purified cell types. We observed that VMRs occur preferentially at enhancers and 3’ UTRs. While the majority of VMRs have high heritability, a subset of VMRs within the genome show highly correlated variation in trans, forming co-regulated networks that have low heritability, differ between cell types and are enriched for specific transcription factor binding sites and biological pathways of functional relevance to each tissue. For example, in T cells we defined a network of 95 co-regulated VMRs enriched for genes with roles in T-cell activation; in fibroblasts a network of 34 co-regulated VMRs comprising all four HOX gene clusters enriched for control of tissue growth; and in neurons a network of 18 VMRs enriched for roles in synaptic signaling. By culturing genetically-identical fibroblasts under varying environmental conditions, we experimentally demonstrated that some VMR networks are responsive to the environment, with methylation levels at these loci changing in a coordinated fashion in trans dependent on cellular growth. Intriguingly these environmentally-responsive VMRs showed a strong enrichment for imprinted loci (p<10−80), suggesting that these are particularly sensitive to environmental conditions. Our study provides a detailed map of common epigenetic variation in the human genome, showing that both genetic and environmental causes underlie this variation.
| Multiple published studies have demonstrated that epigenetic variation can contribute to phenotypic variation. In the present study, we identified regions of common methylation variation in five cell types, observing that these show enrichments for functional genomic features. Surprisingly, we found that these epigenetic variations can form biologically relevant networks that are specific to each cell type, often occurring near genes that have functional relevance to the cell type. Further these regions show reduced heritability, suggesting they may be responsive to environmental cues. We confirmed this by subjecting isogenic fibroblast cultures to different environmental stress. Our study provides insight into patterns of normal epigenetic variation in the human population.
| Understanding the causes and consequences of genomic variation among humans is one of the major goals in the field of genetics. Over the past decade, studies such as the Hapmap and 1000 Genomes Projects have resulted in detailed maps of genetic variation in diverse human populations, identifying millions of single nucleotide polymorphisms, copy number variants and other types of sequence variation [1–6]. These maps have acted as the catalysts for thousands of genome-wide association studies [7], and have provided insights into diverse processes such as mechanisms of human disease, mutation, evolution, migration, selection and recombination [8–11].
However, alterations of the primary DNA sequence are not the only type of genomic variations that occur among humans. In particular there are now well-documented examples of epigenetic marks, such as DNA methylation and histone modifications, that show significant inter-individual variation [12–14]. However, in contrast to sequence polymorphism, relatively few studies have examined the distribution of epigenetic variation across the genome, and as a result our understanding of the causes and consequences of epigenetic polymorphism remains limited.
Familial and twin studies in human and mice [12,13,15–20] have shown that a substantial fraction of sites showing variable DNA methylation levels are highly heritable, and for some loci this epigenetic polymorphism has been linked with nearby genetic variation [21–24]. However, these same studies have also demonstrated that a subset of methylation variation exhibits low heritability [12,16–18,25]. While stochastic variation or technical variability could explain reduced heritability levels, differing environmental exposures such as smoking [26–28], diet/in-utero environment [29–31] and stress [32–35] have all been shown to modify the epigenome. In addition, other natural processes such as aging and X chromosome inactivation apparently underlie epigenetic variation of some sites [36–38]. Whatever the root cause of epigenetic polymorphism, several studies have demonstrated that a subset of these variations are functionally significant and associate with the expression levels of nearby genes [23,39]. Accordingly there is now substantial interest in elucidating the role of epigenetic variation in a variety of disease phenotypes [40–48], indicating that the study of epigenetic polymorphism holds significant promise for understanding the molecular etiology of disease.
In this study, we have performed a screen to identify regions of common epigenetic variation using population data derived from five different human cell types. By searching for clusters of probes with high inter-individual variability, we uncover hundreds of loci in the human genome that exhibit highly polymorphic DNA methylation levels that we term variably methylated regions (VMRs). We show that VMRs co-localize with other functional genomic features, are enriched for CpGs that influence gene expression, and provide evidence that epigenetic variability at some of these loci is influenced by both genetic and environmental factors. We also show that VMRs form cis and trans co-regulated networks enriched for transcription factor binding sites and genes with cell-type relevant functions. Finally, consistent with the notion that the epigenome represents a dynamic link between our genome and the environment[49,50], we experimentally demonstrate environmental effects on methylation at VMRs using cultured fibroblasts, revealing signatures that overlap those observed in our population-level datasets. Together, our results provide novel insights into the biology of variable methylation across the human genome.
We performed an analysis of inter-individual variation of DNA methylation in five isolated cell types from two human cohorts (Fig 1A): 1) Primary fibroblasts, EBV-immortalized lymphoblastoid cells, and phytohemagglutinin stimulated primary T cells taken from umbilical cords of 204 newborns [23]; and 2) sorted glia and neurons from prefrontal cortical tissue from 58 deceased donors [51]. Genome-wide methylation profiles were previously generated for all samples using the Illumina Infinium HumanMethylation450 BeadChip (450k array) (Illumina, San Diego, CA, USA). After filtering (see Methods), we analyzed methylation profiles for 293,782 filtered autosomal CpGs in each of the five cell types. We utilized a sliding window approach (Fig 1B) to characterize VMRs composed of three or more neighboring CpGs with variation ≥95th percentile of standard deviation of β-values in that cell type among all samples of each cell type. To avoid the confounder of gender [52], identification of VMRs was performed separately on males and females, and then the resulting set of VMRs in each gender were combined together for further analysis (see Methods).
In total, we identified 699 VMRs in fibroblasts, 1,423 VMRs in T cells, 699 VMRs in B cells, 1,137 VMRs in neuronal cells and 1,104 VMRs in glial cells. Hereafter, these VMRs are abbreviated as FVMRs, TVMRs, BVMRs, NVMRs and GVMRs, respectively. Genomic positions and relevant annotations for VMRs partitioned by cell type are provided in S1 Table. VMRs had a mean size of 863bp, and contained a mean of 6.4 CpGs (S1 Fig).
While many characterized VMRs were specific to a given cell type, others were common across cell types and tissues. Examples of cell-type specific and shared VMRs are displayed in Fig 2A. The extent of VMR sharing between different tissues was related to their relative developmental origin. For example, approximately one third of VMRs identified in glia were also found in neurons, and ~68% of VMRs found in B cells were observed in T cells. In contrast only 23% of VMRs found in fibroblasts were also seen in B cells (Fig 2B). Between fibroblasts, blood, and brain cells, there were 149 shared VMRs (Fig 2B). In addition, by performing pairwise correlation of methylation levels at CpGs within VMRs shared in different cell types taken from the same individual, we observed much higher correlations between closely related cell types, suggesting that observed population variation is plausibly established in precursors of these cell types and maintained, or influenced by common factors and regulatory mechanisms. For example, methylation levels within VMRs shared between T cells and B cells had a mean correlation coefficient of r = 0.79 (S2 Fig). Likewise for neurons and glia, shared VMR-CpGs were highly correlated (mean r = 0.78, S2 Fig). However, the same degree of correlation was not observed for comparisons between fibroblasts and either T cells (mean r = 0.56) or B cells (mean r = 0.43, S2 Fig).
Before extending our analysis of VMRs, we first replicated our approach in two additional populations. We applied our sliding window approach for identifying VMRs to (i) an additional cohort of 62 fibroblast cell lines [42,53], and (ii) a cohort of whole-blood methylomes from 2,680 individuals sampled from the general population [54]. This led to the identification of (i) 986 VMRs in fibroblasts, 230 of which were also observed in the original fibroblast population, and (ii) 1,368 VMRs in whole blood. Because this latter dataset was composed of methylation profiles generated from peripheral blood, rather than purified cell types, we compared VMRs identified in these controls with shared VMRs that were identified in both B cells and T cells in the Gencord cohort: 390 of the 477 shared B and T cell VMRs and were also found in the replication cohort, yielding a 28-fold enrichment over that expected by chance (p<2.5x10-321).
Differentially methylated CpGs have been shown to often be enriched in specific regions of the genome and to co-localize with other functional epigenetic signatures [55–57]. In order to gain insight into the genomic context of CpGs in VMRs, we tested the enrichment of these CpGs in relation to various genomic features compared to a background set of CpGs assayed on the array (S2 Table).
We first performed enrichment analysis using Refseq gene and CpG island (CGI) annotations, observing consistent trends across datasets (S2 Table). Specifically, we noted that in all five of the cell types tested, VMRs were significantly enriched in 3’ UTRs and depleted in 5' UTRs (enrichments in 3’ UTRs ranging from 1.1- to 1.4-fold across the different cell types, p = 4.6x10-2 to p = 2.5x10-11). Likewise, the depletion of VMRs within 5’ UTRs ranged from 1.3- to 1.6-fold (p = 2.4x10-8 to p = 4.7x10-37) (S2 Table). The depletion in 5’ UTRs was also reflected in enrichment tests conducted using CGI annotations, which revealed significant depletions in CGIs and concomitant enrichments in CpG shores, shelves, and sea categories (S2 Table).
To further explore the co-localization of VMRs with functional genomic regions, we assessed the overlap of FVMRs and BVMRs with Chromatin State Segmentation annotations from a normal human lung fibroblast (NHLF) cell line and an EBV-immortalized lymphoblastoid cell line (GM12878), respectively; these data were previously generated by the ENCODE project [58], and included genome-wide annotations for 15 chromatin states characterized using combined epigenetic signatures from various datasets. Consistent with observed depletions in gene 5’ UTRs and CpG islands, which both tend to occur within or adjacent to gene promoters and transcriptional start sites, we also noted significant depletions of both FVMRs and BVMRs in regions defined by “Active Promoter” chromatin states in respective cell types (S2 Table). The strongest VMR enrichments in both cell types occurred in chromatin states associated with enhancer activity (S2 Table).
We also examined various other categories of genomic features in relation to VMRs, and observed the following: (i) housekeeping genes [59] were strongly under-represented in VMRs in each of the five cell types tested, (ii) loci from the GWAS catalog [60] were enriched in VMRs found in T cells, glia and neurons, (iii) loci showing human-specific methylation levels from a multi-primate analysis [61] were enriched in four of the five cell types, and (iv) loci showing parent-of-origin specific methylation associated with imprinted regions were enriched in neuronal VMRs.
We next sought to investigate the positional relationships of co-regulated VMRs. In each cell type we constructed pair-wise correlation matrices of all VMRs based on the β-values of the probe with the highest population variance within each VMR. The resulting heat maps of pairwise correlations revealed the presence of strongly co-methylated blocks of CpGs, whose methylation levels varied together in both cis and trans, and that these patterns were distinct to each cell type (Fig 3; S3 Fig). For example, as shown in Fig 3, FVMRs exhibit strong cis correlations within several chromosomal regions. Significantly, evidence of strong co-regulation in trans can also be seen, with several regions located on multiple different chromosomes also exhibiting strong co-variation in epigenetic state. Visual inspection of the strongest trans correlations in fibroblasts located on chromosomes 2, 7, 12 and 17 showed that each of these co-regulated clusters of VMRs corresponded to different members of the HOX gene superfamily, suggesting that such VMRs might correspond to coordinately regulated loci with shared biological functions.
Based on this observation, we sought to formally identify signatures of co-regulation among different VMRs. We used weighted gene co-expression network analysis (WGCNA; see Methods) [62,63], to identify co-methylated networks of VMRs within each cell type. This identified seven co-regulated modules in fibroblasts, four in T cells, two in B cells, seven in neurons, and five in glia, with each module composed of between 11 and 467 distinct co-regulated VMRs (median module size, n = 41) (S4 Table, S4 Fig). Consistent with our initial visual observations, WGCNA identified several co-regulated modules within the set of fibroblast VMRs that included all four human HOX gene clusters (Fig 4A).
In order to assess the biological relevance of these co-regulated VMR networks, we performed Gene Ontology (GO) enrichment analysis on the set of genes linked to the VMRs within each module (Fig 4C, S4 Table, S5 Table, and S5 Fig). Although for many networks the number of associated genes was too small to reach significance at 10% FDR, in four of the five cell types tested we identified enrichments for GO terms that were of direct functional relevance to the specific cell type. The five most significant GO enrichments and associated modules for each cell type are presented in Table 1. For example, in fibroblasts, the most significant functional categories were within the blue module that included multiple HOX gene clusters, including terms associated with the basic control of tissue growth and morphogenesis, such as “anterior/posterior pattern specification” (GO:0009952; 52-fold enrichment, FDR q = 3.09x10-14) and “embryonic organ morphogenesis” (GO:0048598; 52-fold enrichment, FDR q = 6.4x10-12). In T cells, the most significant GO enrichments were found for the blue module, made up of 95 co-regulated VMRs enriched for genes involved in T cell function, including the terms “T cell aggregation” (GO:0070489; 11-fold enrichment, FDR q = 1.06x10-6) and “T cell receptor signaling pathway” (GO:0050852; 12.8-fold enrichment, FDR q = 9.5x10-7). In glial cells, significantly enriched terms included a module consisting of 467 VMRs linked to genes associated with “negative regulation of neurogenesis” (GO:0050768; 3.7-fold enrichment, FDR q = 2.4x10-4). Finally, in neurons, the most strongly associated functional categories were with a module comprised of 18 VMRs including the GO term “synapse assembly” (GO:0007416; 61-fold enrichment, FDR q = 1.1 x10-5). Complete lists of enriched GO terms and modules are provided in S5 Table.
Based on the trans nature of these co-regulated VMR networks, we hypothesized that coordinated epigenetic regulation of these sites might be based on the binding of specific trans-acting factors to the members of each VMR network. We therefore analyzed the overlap of each VMR WGCNA module with validated transcription factor binding sites (TFBS) for 161 different transcription factors (TFs) studied by the ENCODE project [64]. We observed significant enrichments for TFBS in several VMR modules that were specific to each cell type (S6 Table). The top three enriched TFBS per cell type are provided in Table 2. In several instances, the most significant TFBS enrichments converged on modules highlighted by GO analyses. For example, EBF1 and RUNX3, which are both involved in lymphocyte differentiation and proliferation [65], were significantly enriched TFs in the blue module in T cells (RUNX3, 2.1-fold enrichment, p = 6.1x10-7; EBF1, 2-fold enrichment, p = 2.7x10-5). Similarly, in fibroblasts, TFBS for SUZ12 (3.4-fold enrichment, Fisher’s. p = 5.3x10-11) and EZH2 (2.3-fold enrichment, p = 1.5x10-9), were the most significantly enriched among VMRs of the module that included multiple HOX-genes (Fig 4C). Prior studies have shown that as part of the polycomb complex, SUZ12 and EZH2 have roles in the establishment of epigenetic modifications, and specifically in the regulation of HOX genes [66].
Motivated by the signatures of co-methylation observed in our VMRs, we next sought to broadly explore the potential underlying factors associated with the regulation of VMR methylation variability. To do this, we first assessed the relationships between CpGs within VMRs, genetic variation, and gene expression. We tested for enrichment of FVMRs, BVMRs and TVMRs with previously described CpG methylation:gene expression associations (eQTMs) and CpG methylation:SNP associations (cis mQTLs) in fibroblasts, B cells and T cells [23]. We observed significant enrichments for VMRs in all three cell types for both CpGs that function as eQTMs and those linked with mQTLs, with enrichments of 16-, 3.4-, and 2.8-fold in eQTMs, and 4.7-, 4.8-, and 6-fold for association with mQTLs in FVMRs, BVMRs, and TVMRs, respectively (all p-values <10−45, S2 Table). To further investigate the relationship of VMRs with underlying genetic variation we used methylation heritability estimates characterized in peripheral blood leukocytes from a cohort of 117 families [19]. Overlaying heritability estimates onto VMR-CpGs across the five cell types revealed that methylation levels for CpGs within VMRs showed significantly increased heritability compared to non-VMR CpGs (Fig 5A). Thus, epigenetic variation at VMRs is often associated with nearby gene expression, and methylation levels at many VMRs shows strong evidence of being under local genetic control.
However, despite this evidence for genetic influences underlying a large fraction of epigenetic variability, the existence of co-regulated modules of VMRs in trans led us to hypothesize that a subset of epigenetic variation might be linked to non-genetic influences, such as differing environmental exposures. To further explore the influence of non-heritable factors on the epigenetic state of VMRs, we analyzed methylation profiles derived from whole blood samples from 426 monozygotic (MZ) twin pairs [35]. Previous studies have shown increased discordance of DNA methylation levels between MZ twins with age, presumably due to differing environmental exposures and/or stochastic processes [67]. We first identified a total of 1,289 VMRs (8,251 CpGs) in these twins (S7 Table), which showed strong overlap with VMRs identified in both B cells and T cells (64% and 59%, respectively). Based on the premise that epigenetic differences between MZ twin pairs provides a measure of the non-genetic component of epigenetic variability, at each CpG we calculated the mean absolute methylation discordance for all autosomal CpGs within each MZ twin pair. We observed a highly significant increase in MZ twin discordance for CpGs within VMRs versus non-VMR CpGs (p<10−300) (Fig 5B). While it is possible that this increased twin-twin discordance might be related to the inherent variability of CpGs, these observations are also consistent with the influence of environmental effects on methylation variability at a subset of VMRs.
To further investigate potential links of our VMRs with known environmental effects on DNA methylation, we utilized a published data set of Illumina 450K data generated from 128 children conceived in the rural Gambia in either the rainy or dry seasons. Here, maternal nutrition at conception shows substantial seasonal variation and is known to be associated with epigenetic differences in children conceived in each season [34]. We applied a Student’s t-test to compare methylation values in children conceived in the rainy versus dry season, calculating the resulting p-value for seasonal difference as a measure of environmental influence at each CpG. Comparing CpGs within VMRs to those in the background set of all non-VMR probes, we observed a strong enrichment for CpGs influenced by season of conception in VMRs from all five cell types when compared to the background set of p-values obtained from all 293,782 non-VMR probes (p = 8.4x10-5 in Fibroblasts to 1.1x10-70 in T cells) (S6 Fig).
To experimentally verify whether methylation levels at some VMRs are responsive to environmental cues, we performed cell culture experiments in which we grew genetically identical fibroblasts under different environmental conditions, varying the rate of culture media replenishment and cell density with time (summarized in Fig 6A). Skin fibroblasts from a single normal male (GM05420) were seeded in parallel from a single master culture into eight separate flasks, and allowed to grow under normal or low-nutrient conditions, achieving varying levels of cell density at each time point. Every 48 hours one flask was harvested from each media replenishment regime, DNA extracted and profiled on the 450k array, resulting in DNA methylation profiles for nine samples (see Methods).
We reasoned that as these fibroblast cultures shared an identical genetic background, any epigenetic variations observed among them would be attributable to non-genetic factors, such as varying culture environment. We applied the same sliding-window method to identify VMRs in methylation data from these cultured isogenic fibroblasts, identifying 135 putatively “environmentally responsive” VMRs. This included many of the same VMRs identified previously in our population-based analysis of umbilical cord-derived fibroblasts (S8 Table), with a 5.3-fold enrichment for overlap between these two sets of VMRs compared to the background set of probes on the array (p = 2.9x10-29). Examples of VMRs showing changes in methylation level with culture conditions are shown in Fig 6B.
Concordant with our population analysis, GO analysis of the 135 VMRs from cultured isogenic fibroblasts revealed enrichments for HOX genes, as well as several of the same GO terms associated with the co-regulated FVMR modules (S9 Table). Strikingly, these environmentally responsive VMRs were also enriched 35-fold for CpGs within known imprinted loci versus the null (p = 5.6x10-79). This included overlaps with differentially methylated regions associated with the imprinted genes MKRN3, IGFIR, ZNF331, PEG3, L3MBTL, GNAS and MEST (Fig 6C) [68].
Here we surveyed variation in DNA methylation patterns in five purified human cell types, identifying hundreds of genomic loci that exhibit a high degree of epigenetic polymorphism in the human population: we term these ‘Variably Methylated Regions’ or VMRs. We observed that VMRs are enriched for various functional genomic features, most notably enhancers, suggesting a potential role in regulating gene expression patterns. Unexpectedly, we found that many VMRs form co-regulated networks both in cis and in trans, with multiple VMRs spread across different chromosomes at which methylation levels vary in a coordinated fashion. These co-regulated networks were specific to each cell type, had reduced heritability, and were also enriched for gene sets with cell-type relevant functions. For example, we observed VMR networks associated with genes enriched for synaptic transmission in neurons, regulation of nervous system development in glia, and T cell activation in T cells. These observations suggest that some VMRs represent loci that form co-regulated pathways that are implicated in the regulation of genes with cell-type specific functions. The dispersed nature of these co-regulated VMR networks indicates that they are potentially regulated by trans acting factors, and consistent with this we found significant enrichments for relevant transcription factor binding sites associated with some networks.
While many VMRs are influenced by local genotypes, our analyses of monozygotic twins, a cohort of African samples conceived in divergent nutritional environments, and in-vitro culture of genetically identical fibroblasts cell lines indicates that epigenetic variation at some VMRs is linked to environmental factors. Indeed, using isogenic fibroblast cultures derived from a single individual that were grown under different environmental conditions, we were able to replicate many of the same VMRs found in our original population analysis, thus showing that epigenetic variation at these loci is an environmentally inducible trait. Intriguingly these environmentally-responsive VMRs showed a strong enrichment for imprinted loci (p<10−94), suggesting that these genes are particularly sensitive to environmental conditions. This observation that varying cell culture conditions result in epigenetic alterations across the genome, presumably accompanied by changes in gene expression, highlights that the use of cultured cells for investigating epigenetic phenomena should be approached with caution. We suggest that unless carefully controlled, variations in cell culture conditions could easily introduce significant epigenetic and transcriptional changes that could confound many in vitro studies.
VMRs in fibroblasts comprised co-regulated modules that included all four HOX gene clusters that are each located on different chromosomes. While a previous study has reported that the HOX genes exhibit variable methylation that correlates with their expression levels [53], our analysis builds on these observations by showing that methylation across multiple HOX gene clusters is correlated in both cis and trans. Furthermore, using validated transcription factor binding sites, we found a significant enrichment for transcription factors EZH2 and SUZ12 at these VMR sites associated with HOX genes. These two transcription factors are components of the Polycomb Repressive Complex 2 (PRC2), which functions as a histone H3K27-specific methyltransferase and regulates both epigenetics and expression of HOX genes [66]. Thus, we propose a model where coordinated variation of DNA methylation at multiple loci in trans, corresponding to a network of co-regulated genes, is under the control of transcription factor binding in response to physiological and/or environmental cues. In the case of the HOX gene network in cultured fibroblast cell lines, such cues could be the availability of nutrients, local cell density and other growth conditions, allowing the cells to modify their growth trajectories in response to the prevailing environment. Consistent with this model, recent observations were made in macrophages, a type of immune cell that has a variety of roles in different tissues around the body, which mirror our findings. Two prior studies showed that the epigenetic state of enhancer elements in these cells responds to the tissue microenvironment in which they reside, and is regulated by networks of tissue- and lineage-specific transcription factors that drive divergent programs of gene expression [69,70]. Studies of chromatin accessibility have also shown that manipulating the presence of specific transcription factors can lead to global modification of epigenetic state at multiple loci in trans [71].
We defined VMRs as clusters of probes with high variance, as the use of single probes to determine epigenetic variability is inherently unreliable. This is because other phenomena unrelated to epigenetic variance can influence the β-value reported by a probe, independent of DNA methylation levels. These include, for example, underlying genetic variants that alter probe binding, random technical effects such as hybridization or wash artifacts on the array, or the simple fact that some probes might simply perform poorly and yield inherently more variable results. By considering groups of multiple closely spaced probes that all show high variability makes it much less likely that our results would be influenced by the effects listed above, thus improving specificity for the detection of true epigenetic variability, and reducing artifacts.
One of the strengths of this study is that we specifically utilized purified cell types for our analysis, some of which were also of homogeneous age. This has the advantage of removing the confounder of both cellular heterogeneity and age effects, both of which are known to influence DNA methylation [36,37,72]. Such differences would otherwise result in many false positive VMRs due to underlying differences in cell fractions or age among individuals.
One of the limitations of this analysis is that we used methylation profiles from the Illumina 450k array, which targets only a small subset (~3%) of CpGs in the human genome, and has coverage that is biased towards gene promoters and CpG islands. As such, the maps of VMRs we provide here are far from comprehensive, and future work that utilizes more comprehensive approaches (e.g. whole genome bisulfite sequencing) will undoubtedly provide more complete genomic maps of epigenetic variation. However, to our knowledge currently no such datasets on a population-scale are available. One other potential caveat is that the methylation profiles for B cells, fibroblasts and T cells were all generated from cells that had been cultured in vitro, and furthermore the B cells were also immortalized by Epstein-Barr virus infection, a process which is known to induce widespread epigenetic changes [73]. However, we observed good replication of the VMRs identified from cultured/immortalized B cells and T cells in an independent cohort where DNA was extracted from uncultured blood, indicating that many of these same VMRs observed even in immortalized B cells are also present natively.
One confounder that deserves mention is that there is potential that some of the enrichments we identify between VMR probes and factors such as mQTLs and eQTMs could be driven by the inherent variability of VMRs. This is because statistical power to identify an association with a locus is heavily influenced by the underlying variability of that site. Thus, it is possible that some of enrichments we observe are due either partly or wholly to this effect. Similarly, it is also possible that the increased discordance we observed in MZ twin pairs at VMRs might be driven simply by the higher variability of methylation at these loci.
In conclusion our study of DNA methylation polymorphism provides novel insights into the nature and function of epigenetic variation. The coordinated phenomenon we observed where methylation levels at networks of multiple genomic regions varies in response to the local environment is consistent with popular theories that the epigenome can indeed act as an interface between the genome and environment [39,50,74].
We obtained DNA methylation data generated using the Illumina 450k HumanMethylation BeadChip from two published studies. We utilized data from the Gencord cohort from the EMBL-EBI European Genome-Phenome Archive (https://www.ebi.ac.uk/ega/) under accession number EGAS00001000446, representing 90 fibroblast cultures, 61 T-cell cultures and 111 immortalized B-cell cultures derived from a cohort of newborns [23]. We also utilized methylation data representing FAC-sorted glial and neuronal cells from 58 deceased donors downloaded from GEO (http://www.ncbi.nlm.nih.gov/geo/) under accession number GSE41826 [51]. Prior to analysis for methylation variation, each dataset underwent several filtering and normalization steps, as follows. In each individual, probes with a detection p>0.01 (mean n = 348 per sample) or mapping to the X or Y chromosomes were removed. 482,421 probe sequences (50-mer oligonucleotides) were remapped to the reference human genome hg19 (NCBI37) using BSMAP, allowing up to 2 mismatches and 3 gaps, retaining those 470,576 autosomal probes with unique genomic matches. Probe coordinates were converted to hg19 using liftover. Probes that overlapped SNPs identified by the 1000 Genomes Project (minor allele frequency ≥0.05) either including or within 5bp upstream of the targeted CpG (n = 13,376 autosomal probes) were discarded, as such variants can introduce biases in probe performance. We also removed probes overlapping copy number variants of ≥5% frequency in CEU HapMap samples [75].
After filtering, we retained 457,200 autosomal and 11,021 chrX probes on which we performed a two color channel signal adjustment and quantile normalization on the pooled signals from both channels and recalculation of average β-values as implemented in the lumi package of R [76]. The Illumina Infinium HumanMethylation450 BeadChip contains two assay types (Infinium type I and type II) which utilize different probe designs. As the data produced by these two assay types shows distinct profiles, to correct this problem we performed a beta mixture quantile normalization method utilizing BMIQ [77] on the normalized data. β-values were quantile normalized using the normalizeQuantile function in the aroma.light R package. One pair of neuronal/glial samples was excluded on the basis that they showed discrepant gender, as determined by PCA analysis of β-values on the sex chromosomes. We further removed 22,267 probes where any of the cell types had >5% probes with detection p-value >0.01. For sliding window analysis, we sub-selected 293,782 autosomal 1kb windows containing 3 or more probes. CpGs in these regions were then annotated based on their position relative to RefSeq genes using BEDTools v2.17 [78].
To compare similarity of VMRs across cell types, we utilized the fact that >1 cell type was available from each individual (Fibroblasts, T cells and B cells in one cohort, and Neurons and Glia in the other). We compared the similarity of VMRs between two cell types by first selecting probes from VMRs found in both cell types, and then computing pairwise Pearson correlation coefficients on the β-values from the two cell types from each individual. The number of individuals and probes in pairwise VMRs were:
To identify regions of common highly variable methylation that should be robust to fluctuations in single probes, we chose an approach to identify loci containing multiple independent probes showing high population variance. To avoid gender effects creating false positives in our analysis, either biological or technical due to cross-hybridization artifacts [52], we first divided males and females and analyzed each gender separately. For each probe, we calculated the standard deviation (SD) of the β-value separately in each cell type. We then utilized a 1kb sliding window based on the start coordinate of each probe, beginning at the most proximal probe on each chromosome and moving down consecutively to the last probe on each chromosome. We defined VMRs as those 1kb regions containing at least 3 probes ≥95th percentile of SD in that cell type, with an additional criterion that at least 50% of the probes in that window were also ≥95th percentile of SD. The relevant scripts used for this paper can be found at Github (https://github.com/AndyMSSMLab/VariableMethyl). VMRs that were found in the same cell type in either males or females were then combined, and used in all downstream analysis.
To identify potential co-regulation relationships among VMRs, we applied Weighted Gene Correlation Network Analysis (WGCNA) to each set of VMRs identified per cell type [63]. Input values for each VMR were β-values for the variable probes within VMR which had standard deviation ≥95th percentile. As suggested in the WGCNA user manual, we plotted scale-free topology fit indices and mean connectivity plots with varying soft-thresholding power. Based on these plots in all five cell types, we chose power of 6 as our soft-thresholding power and scale free topology fit index>0.8 in all cases. We generated adjacency matrices by raising the correlation matrix to the power of 6, which was then transformed into topological overlap matrix (TOM). VMRs were then classified into modules using hybrid dynamic tree cutting with a minimum cluster size of 10 VMRs. VMRs in each module were selected at Module Membership value ≥0.7. We associated VMRs with gene annotations based on either their localization within ±2kb of Refseq transcription start sites, overlap with DNAseI hypersensitive sites that showed significant association in cis with gene expression levels within ENCODE cell lines [79], and significant associations between methylation and gene expression levels (eQTMs) in T cells, B cells, fibroblasts [23]. The fraction of VMRs that were linked to a gene varied from a low of 78.4% in T cells, to a high of 81.8% in Fibroblasts. If two VMRs were members of the same module, but located on different chromosomes, then this was considered a trans association, cis otherwise. For each module we performed Gene Ontology enrichment analysis using in house scripts. Each VMR was annotated with Refseq genes which either overlapped gene body or promoter region as described above. Refseq genes associated with 293,782 probes tested where used as background for gene ontology analysis. P-values for each GO term were generated using the hypergeometric distribution and incorporated 5% FDR correction.
We downloaded the track of Uniform transcription factor binding sites (TFBS) from the UCSC Genome browser [80], containing experimentally determined binding sites for 162 transcription factors. As the precise boundaries of some VMRs were not well defined, we extended TFBS coordinates by ±500bp prior to overlap with the set of VMRs identified in each cell type. Enrichment analysis for TFBS to occur within each module of co-regulated VMRs identified by WCGNA versus the background was performed using a Fisher's exact test. The 2x2 table for Fisher’s exact test contained whether the probe is in a specific module or not and whether they overlap TFBS or not.
Methylation dataset for Replication study (2,680 samples) was downloaded from GEO (GSE55763) and was normalized the same way as described above [54]. VMRs were called with same criteria. A VMR in discovery cohort was considered successfully replicated if the VMR coordinates in discovery cohort overlap (minimum 1 bp) with VMRs found in replication cohort.
We downloaded dataset for CpG features, Refseq gene annotations (3’ UTR, 5’UTR, CDS, Intron & Intergenic), GWAS catalog, Segmental Duplications, Simple Repeats and ChromHMM features from the UCSC table browser [80]. eQTMs and mQTLs were obtained from Gutierrez-Arcelus M et al. [23]. Sites showing human-specific methylation patterns from an analysis of multiple primate species was obtained from Hernando-Herraez et al. [61]. 647 genes thought to be linked with environmental response were obtained from NIEHS website (https://www.niehs.nih.gov), and 1,847 genes categorized under the GO term “response to environmental stimulus” were obtained from Amigo [81]. Cell-count corrected heritability estimates for CpG methylation were obtained from McRae et al. [19], and annotations of CpG islands showing evolutionary constraint and biased gene conversion (BGC) from Cohen et al. [82].
All enrichment analyses were performed by overlapping probe coordinates in VMRs with respective feature and using all 293,782 probes as background. The enrichment p-value was generated using Hypergeometric distribution (phyper function in R). The fold enrichment was calculated using following formula: (probes in VMR overlapping Feature/probes in VMR)/(probes in Background overlapping Feature/probes in Background).
We downloaded published data set of Illumina 450K data generated from 430 monozygotic (MZ) (GEO dataset GSE105018) [35], and 128 children conceived in the rural Gambia in either the rainy or dry seasons (GSE99863) [34]. After normalization using the same methods as described above, we applied a Student’s t-test to compare methylation values in children conceived in the rainy versus dry season, and calculated the resulting p-value for seasonal difference, a measure of this environmental effect. A Wilcoxon Rank-Sum test was performed to compare the distribution of this p-value between VMRs and non-VMRs.
A growing culture of human skin fibroblasts from a normal male individual (GM05420) was obtained from Coriell Institute for Medical Research (Camden, NJ). Cells were grown in RPMI1640 media supplemented with 1mM L-glutamine, 10% FBS and 100u/L each of penicillin and streptomycin. A single vial of fibroblasts was initially grown in a 2ml culture plate, with media changed every 24 hours. Once the cells attained 80% confluency they were trypsinized and split equally into two T25 flasks. Each flask was treated identically, with media changed every 24 hours until the cells achieved 80% confluency (approximately 7 days after seeding). Both cultures were then trypsinized, mixed, and the cells seeded equally into a total of nine T25 flasks, which were then harvested at set time points (TP) under different culture regimes, as follows:
At each time point, cells were harvested by trypsinization, pelleted by centrifugation, and frozen at -20 Celsius. Once all cultures were harvested, DNA was extracted in a single batch using the Qiagen DNeasy blood and tissue kit and these samples processed together on a single chip using the Illumina 450k HumanMethylation BeadChip according to manufacturer’s instructions. The resulting data were then processed and normalized as described above. Given small sample size, we excluded the step where we performed quantile normalization on β-values (aroma.light), and VMRs across these nine samples defined as 1kb regions containing at least 3 probes ≥95th percentile of SD, with an additional criterion that at least 50% of the probes in that window were also ≥95th percentile of SD. Methylation array data from cell line GM05420 have been deposited in the NCBI Gene Expression Omnibus (GEO) (http://www.ncbi.nlm.nih.gov/geo) under accession number GSE76836.
|
10.1371/journal.pgen.1005040 | Extreme-Depth Re-sequencing of Mitochondrial DNA Finds No Evidence of Paternal Transmission in Humans | Recent reports have questioned the accepted dogma that mammalian mitochondrial DNA (mtDNA) is strictly maternally inherited. In humans, the argument hinges on detecting a signature of inter-molecular recombination in mtDNA sequences sampled at the population level, inferring a paternal source for the mixed haplotypes. However, interpreting these data is fraught with difficulty, and direct experimental evidence is lacking. Using extreme-high depth mtDNA re-sequencing up to ~1.2 million-fold coverage, we find no evidence that paternal mtDNA haplotypes are transmitted to offspring in humans, thus excluding a simple dilution mechanism for uniparental transmission of mtDNA present in all healthy individuals. Our findings indicate that an active mechanism eliminates paternal mtDNA which likely acts at the molecular level.
| Emerging evidence raises the possibility that human mitochondrial DNA (mtDNA) is not strictly maternally inherited, but it has not been technically possible to test this hypothesis directly. We identified trios with discordant mtDNA haplotypes, parent-offspring trios were validated using polymorphic microsatellites, and then used extreme-high depth mtDNA re-sequencing to look for paternally transmitted mtDNA. Despite having up to ~1.2 million-fold coverage of mtDNA, we find no evidence that paternal mtDNA haplotypes are transmitted to offspring in humans. Our findings exclude a simple dilution mechanism for uniparental transmission of mtDNA present in all healthy individuals.
| In eukaryotes, cytoplasmic genes are generally inherited from the mother. The mechanisms responsible for this appear to differ between organisms. The elimination of mitochondrial DNA (mtDNA) during spermatogenesis prevents paternal transmission in Drosophila melanogaster [1]. Conversely, in the Japanese medaka fish Oryzias latipes, sperm mtDNA is lost after fertilization [2]. In cows and humans, sperm mtDNA is eliminated from two or four cell embryos [3,4], and sperm loss may also occur throughout embryogenesis in mice [5].
Although sperm mitochondria are tagged with ubiquitin and actively destroyed through P62 and LC3-mediated autophagy [6,7], there is no direct evidence showing the destruction of sperm mitochondrial DNA (mtDNA) [8], and there are several examples where paternal mtDNA has escaped this process. Extensive paternal transmission of mtDNA has been observed in the marine mussel (Mytillus sp) [9], and occasionally in the fruit fly (Drosophila melanogaster)[10], Lepidopteran insects and the honey bee (Apis mellifera) [11]. For the most part, the “leakage” of paternal mtDNA during transmission in mammals has only been observed in atypical situations such as inter-strain breeding in mice (Mus musculus) [12], or following in vitro embryo manipulation in cattle (Bos taurus) [13]. However, the description of a paternally-derived 2 base pair pathogenic deletion in the mtDNA MTDN2 in a 28-year-old man with a mitochondrial myopathy [14], the persistence of human sperm-derived mtDNA when introduced into somatic cells [15] and in abnormal fertilised human embryos [16], coupled with evidence of paternally transmitted mtDNA in healthy sheep (Ovis aries) [17], and the great tit (Parus major) in its natural habitat [18], questions the accepted dogma of exclusive maternal transmission.
In humans, the debate hinges on the analysis of mtDNA sequences at the population level. By studying partial or complete mtDNA sequences from individuals across the globe, some have argued that the co-occurrence of phylogenetically unrelated genetic variants indicates inter-molecular recombination between paternal and maternal mitochondrial genomes [19], but others have argued that the high mtDNA mutation rate confounds this analysis through the generation of homoplasy [20], which can reach ~20%. Recent experimental data showed no evidence of the active removal of sperm mtDNA from developing mammalian embryos [8], pointing towards a passive dilution process based on differences in the amount of mtDNA between human gametes. Here we test this hypothesis directly.
First we determined the proportion of paternal haplotypes transmitted to the offspring assuming no preferential destruction of sperm-derived mtDNA. We measured the amount of mtDNA in healthy human sperm and pre-fertilization oocytes on the same assay plate. This gave a mean ratio of 1:15,860, in keeping with previous reports [8,21–23]. Based on these observations, the 95% prediction interval for the proportion of paternal haplotypes at fertilization is 10–5 to 1.8x10-4.
Next we developed an extreme-depth mtDNA re-sequencing assay to detect very low levels of paternal haplotypes. Given previous work showing a background level of ~1% heteroplasmy for single variants using deep mtDNA re-sequencing [24], we set out to identify trios where the father and the child had two or more variant differences within a <200bp stretch of mtDNA, thus allowing the detection of extremely rare paternal haplotypes at a much lower heteroplasmy level captured within the same sequencing read. Sanger sequencing of the mtDNA from 99 mother-father-child trios from the Avon Longitudinal Survey of Parents and Children identified four different trios with >2 discordant alleles (subsequently referred to as discordant variants, S1 & S2 Tables, Fig 1A & S1 Fig). Parent-offspring trios were confirmed with >99.9% accuracy using 16 short microsatellites in all members of each trio. Next we designed PCR amplicons spanning the four discordant regions. Each amplicon was >2Kb to eliminate nuclear pseudogene amplification, and BLAST analysis predicted exclusive annealing to mtDNA (Fig 1A), which was confirmed by the failure to generate a template from rho-0 cells devoid of mtDNA. All four amplicons were amplified in all four trios using ultra-high fidelity polymerase (PrimeSTAR GXL, TaKaRa Bio Europe, fidelity = 1 error/~16,230 base pairs), allowing each trio to act as a control for the others. Extreme high-depth re-sequencing (Illumina MiSeq, 250bp paired-end reads) yielded stable coverage spanning the region of interest in the offspring (Fig 1B).
We then compared the frequency of minor alleles and haplotypes both within and between the four trios. The frequency of isolated single variants was similar to that observed previously at lower depth at ~ 0.5% [24]. As expected, the number of minor haplotypes containing two or three variants was substantially less (Table 1). Overall, the frequency of minor haplotypes containing two variants approximated the square root of the single variant frequency, and the frequency of the three-allele haplotypes was approximately equal to the cube root of the single variant frequency. These observations were in keeping with a random background mutation frequency affecting single base-pairs of ~0.5% [24]. The observed alleles contributing to the ultra-rare haplotypes did not correspond to commonly co-occurring population variants, making external sample contamination unlikely (Fig 1D and S2 Table). The frequency of unexpected maternal or paternal haplotypes was greater in other members of the same family trio than in other trios. Given that all of the trios were analysed in the laboratory simultaneously, these rare shared haplotypes probably arose through very low-level contamination at the time the original samples were taken, in the order of <10–5 molecules (Table 1). We incorporated these observations in a significance test of our findings, and determined whether we had adequate statistical power to detect paternally inherited mtDNA.
The power to detect paternally inherited mtDNA was estimated by determining the difference between mismatched haplotype frequencies in mothers and children. These are simply modelled by Poisson distributions with a without paternal contributions, and the difference is described by a Skellam distribution [25]. The power was investigated for a range of theoretical paternal mtDNA contributions, and the observed background heteroplasmy levels in mothers (Fig 2). For the observed background heteroplasmy levels in the mothers (4.3 x 10–5, 6.0 x 10–5, 5.7 x 10–5, and 7.6 x 10–5, Table 1), sequencing at >300,000-fold depth in the relevant offspring had >80% power to detect the predicted level of transmitted paternal mtDNA.
A bootstrap test was used to determine whether the observed discordant haplotypes in the children were consistent with the predicted paternal contribution, incorporating the 95% confidence intervals for proportion of sperm haplotypes in the oocytes (10–5 to 1.8x10-4). A paternal contribution was not compatible for trios A and B (p = 0.004, p = 0.016), but there was insufficient evidence to reject the other two trios (p = 0.21, p = 0.12) that had a lower total coverage for the full range of possible transmitted levels of sperm mtDNA. The power calculation suggests that the uncertainty in two trios likely reflects the confidence interval for our estimate of sperm haplotypes in the fertilized oocyte, and not directly reflect the likelihood of any paternal transmission of mtDNA.
The full data set and relevant code is available from the authors and at: https://www.staff.ncl.ac.uk/i.j.wilson/PaternalTransmission
When taken together these findings indicate that the extremely rare variant haplotypes seen in the offspring are highly unlikely to have arisen through the passive ‘leakage’ of paternal mtDNA within the trio. Lower levels of paternal mtDNA transmission are also unlikely because we did not see common mtDNA population haplotypes in any of the individuals studied (Fig 1D), which would be expected if there were frequent paternal leakage in the human population. The lack of common population haplotypes at extreme high depth also shows that our experimental approach was not subject to significant cross-contamination. Finally, we show that there is little to be gained by sequencing at >5000-fold if single variants are of interest, because the ‘noise’ level will prevent further resolution of low-level heteroplasmy.
The ‘background noise’ level that we observed has several potential origins. First, the near exponential decrease in frequency of single variants, two variant, and three variant haplotypes is consistent with a random background mutation frequency introduced by DNA replication errors, either within the biological system or through PCR amplification. We did not observe patterns of nucleotide changes that would imply a direct insult to the DNA templates (such as C>U deamination damage). Finally, even at extreme high depth (>1 million reads in some trios) we saw negligible or no haplotypes observed in other trios. Given that the trios were analysed together, this means that our laboratory procedures were robust with negligible cross contamination. It is therefore likely that the rare paternal haplotypes seen in maternal samples (Table 1) were introduced when the samples were collected.
Our observations were made on DNA extracted from buccal swabs. It is thus theoretically possible that paternal mtDNA was originally transmitted to the offspring and subsequently lost from buccal cells before the DNA samples were acquired. The loss of mutated mtDNA has been observed in patients with mtDNA diseases, most typically m.3243A>G, probably through selection against a deleterious allele at the stem cell level [26]. However, for most inherited heteroplasmic mtDNA mutations (eg m.8993T>C), the percentage level of the mutation is the same in a wide range of different tissues [27], including buccal swabs. It is therefore likely that if there were paternal transmission, this would be detected in all tissues. Moreover, since variants we studied here are haplogroup markers and are unlikely to have significant biochemical consequences [28], it is highly unlikely that a paternal haplotype allele would be lost from a particular tissue through selection. Absolute reassurance would be provided by studying other tissues, but this approach has its own difficulties because somatic point mutations accumulate in non-dividing tissues, increasing the background ‘noise’ levels to a point that would preclude the detection of passively transmitted paternal mtDNA.
Given that human paternal mtDNA can be detected in very early embryos [4], what is the mechanism underpinning exclusive maternal inheritance of mtDNA in humans? Recent observations that ubiquitination and P63/LC3 tagging of sperm mitochondria does not lead to mitophagy in mice, the mechanism is likely to act at the level of the mitochondrial genome [8]. This is plausible, given recent observations in heteroplasmic mice where a similar mechanism was proposed for heteroplasmy segregation during inheritance [29,30], and within different tissues and organs during life [31], where selection against specific mtDNA molecules occurs at levels well below the level required to cause a biochemical defect affecting oxidative phosphorylation within the cell. A similar mechanism was proposed following the introduction of sperm mtDNA into somatic cells [15]. Although it remains to be determined how this selection process occurs at the molecular level, it is of fundamental importance in preventing the accumulation of deleterious mutations in the human population, effectively taking Muller’s ratchet [32] ‘down a gear’.
Ethical approval for the study was obtained from the ALSPAC Ethics and Law Committee and the Local Research Ethics Committees. The ALSPAC study web site contains details of all the data that is available on the cohort through a fully searchable dictionary (http:www.bris.ac.uk/alspac/researchers/data-access/data-dictionary/).
Excess oocytes were collected after follicular reduction from healthy donors, lysed for 16 hours in 50mM Tris-HCl, pH 8.5, with 0.5% Tween 20 and 200ng/ml proteinase K, at 55°C, followed by heat inactivation at 95°C for 10 minutes. Three single oocytes and 43 sperm DNA samples were analysed using quantitative real-time PCR (qPCR). A multiplex qPCR assay, using probes and primers targeting a region of MT-ND1 and the nuclear housekeeping gene β-2 microglobulin (β2M), were used to measure mtDNA copy number on a CFX96 Touch™ Real-Time PCR Detection System (BioRad, USA).
Buccal DNA samples were extracted in a different laboratory from 100 mother-father-child trios obtained from the Avon Longitudinal Study of Parents and Children Cohort (ALSPAC) [33,34]. Mitochondrial DNA (mtDNA) haplogroup defining single nucleotide polymorphisms (SNPs, www.phylotree.org) were determined in each mother and father by Sanger sequencing specific regions of the mitochondrial DNA genome. (Big Dye v3.1 kit and capillary electrophoresed on an ABI3130xl Genetic Analyzer, Life Technologies, Warrington, UK). Alignment and variant calling was performed using SeqScape software (v2.1.1, Applied Biosystems) reference to the GenBank sequence NC_012920.1. Four trios were identified with >2 discordant variants falling within a ~250 bp read-length (S1 Table). Parent-offspring trios were confirmed using the Promega Powerplex 16 HS system (Promega, Southampton, UK, performed by NorthGene Ltd.).
2x 2Kb mtDNA amplicons were designed to span the discordant regions in each trio (Fig 1A) and to avoid nuclear pseudogene co-amplification with a high-fidelity polymerase (PrimeSTAR GXL DNA Polymerase, TaKaRa Bio Europe, France). Primer sequences were: set 1 = TATCCAGTGAACCACTATCAC-F (m.11010-11030) and GGGAGGTTGAAGTGAGAGG-R (m.13453-13435); set 2 = ATTCATCGACCTCCCCACC-F (m.14797-14815) and CTGGTTAGGCTGGTGTTAGG-R (m.389-370). Initially, primer efficiency and specificity was assessed as successful after no amplification of DNA from rho0 cell lines, which contain no mtDNA. Amplified products were assessed by gel electrophoresis against DNA+ve and DNA-ve controls, and quantified using a Qubit 2.0 fluorimeter (Life Technologies, Paisley, UK). Each amplicon was individually purified using Agencourt AMPure XP beads (Beckman-Coulter, USA), pooled in equimolar concentrations and re-quantified.
Amplicons were ‘tagmented’, amplified, cleaned, normalised and pooled into a multiplex using the Illumina Nextera XT DNA sample preparation kit (Illumina, USA). Multiplex plex pools were sequenced using MiSeq Reagent Kit v3.0 (Illumina, USA) in paired-end, 250 bp reads on the same flow cell.
Post-run FASTQ files were analysed using an in-house developed bioinformatic pipeline. Reads were aligned to the rCRS (NC_012920) using BWA v0.7.10 [35], invoking—mem [35]. Aligned reads were sorted and indexed using Samtools v0.1.18 [36]. Variant calling was performed in tandem using VarScan v2.3.13 [37,38], (minimum depth = 1500, supporting reads = 10 and variant threshold = 1.0%) and LoFreq v0.6.1 [39]. Concordance calling between VarScan and LoFreq was >99.5%. Concordant variants were annotated using ANNOVAR v529 [40]. In-house Perl scripts were used to extract base/read quality data and coverage data.
Potential paternal haplotypes were identified from the pool of analysed reads (*.sam files) using command line scripting, generating motif-specific counts for each trio. For example, [grep—c 'CCTCACTGCCCAAGAACTATCAAACTCCTGAGCCAACAACTTAATATGACTAGCTTACACAATAGCTTTTATAGTAAAGATACCTCTTTACGGACTCCACTTATGACTCCCTAAAGCCCATGTCGAAGCCCCCATCGCTGGGTCAATAGTACTTGCCGCAGTACTCTTG\|CAAGAGTACTGCGGCAAGTACTATTGACCCAGCGATGGGGGCTTCGACATGGGCTTTAGGGAGTCATAAGTGGAGTCCGTAAAGAGGTATCTTTACTATAAAAGCTATTGTGTAAGCTAGTCATATTAAGTTGTTGGCTCAGGAGTTTGATAGTTCTTGGGCAGTGAGG'. /*.sam] was used to identify whole reads containing m.11299C and m.11467G in forward and complement orientations (corresponding bases underlined). Counts were generated for all possible permutations of motifs, in all available samples.
The mtDNA counts for healthy human sperm had mean 77.2 (SD 53.9, n = 43) and oocytes had mean 1220000 (SD 183000, n = 3). An empirical distribution for the ratio of sperm to oocyte mtDNA levels immediately after fertilization was estimated by bootstrapping. 100,000 individual mtDNA counts were sampled from the raw sperm mtDNA counts and 105 corresponding oocyte mtDNA counts were drawn from a normal distribution with mean and standard deviation as above for the oocytes. A 95% prediction interval was then calculated from the 2.5th and 97.5th percentiles of this ratio of sperm mtDNA to oocyte count.
A hypothesis test for paternal transmission was performed by bootstrapping under the null hypothesis that the discordant haplotypes were due to paternal inheritance at the ratio predicted from the direct measurements on individual gametes. For each of the four discordant haplotypes, 105 sampled ratios were simulated in the same way as for the prediction interval, and from this 105 replicate discordant child haplotypes were generated by binomial sampling, with p equal to these ratios and n equal to the total coverage over all haplotypes for this motif and trio. Observed background noise was added to this count at a Poisson rate equal to the discordant maternal haplotype rate. The simulated maternal count was pure Poisson noise, generated at the same rate as the child noise. The bootstrapped differences in proportion were then compared to the observed data, and extreme values are evidence to reject the null hypothesis. The p-value is estimated percentile rank of the observed data in the bootstrapped distribution.
|
10.1371/journal.ppat.0030122 | Limits on Replenishment of the Resting CD4+ T Cell Reservoir for HIV in Patients on HAART | Whereas cells productively infected with human immunodeficiency virus type 1 (HIV-1) decay rapidly in the setting of highly active antiretroviral therapy (HAART), latently infected resting CD4+ T cells decay very slowly, persisting for the lifetime of the patient and thus forming a stable reservoir for HIV-1. It has been suggested that the stability of the latent reservoir is due to low-level viral replication that continuously replenishes the reservoir despite HAART. Here, we offer the first quantitative study to our knowledge of inflow of newly infected cells into the latent reservoir due to viral replication in the setting of HAART. We make use of a previous observation that in some patients on HAART, the residual viremia is dominated by a predominant plasma clone (PPC) of HIV-1 not found in the latent reservoir. The unique sequence of the PPC serves as a functional label for new entries into the reservoir. We employ a simple mathematical model for the dynamics of the latent reservoir to constrain the inflow rate to between 0 and as few as 70 cells per day. The magnitude of the maximum daily inflow rate is small compared to the size of the latent reservoir, and therefore any inflow that occurs in patients on HAART is unlikely to significantly influence the decay rate of the reservoir. These results suggest that the stability of the latent reservoir is unlikely to arise from ongoing replication during HAART. Thus, intensification of standard HAART regimens should have minimal effects on the decay of the latent reservoir.
| Latently infected resting CD4+ T cells represent a stable reservoir for human immunodeficiency virus (HIV). When HIV-infected individuals are treated with highly active antiretroviral therapy (HAART), this latent reservoir decays slowly, with a half-life of up to 44 months. As a result, latently infected resting CD4+ T cells represent the major known barrier to eradication of HIV infection. Two factors are believed to contribute to the stability of the latent reservoir in the setting of HAART: replenishment by low-level viral replication and the intrinsic stability of resting memory CD4+ T cells. Unfortunately, it has not been possible to measure replenishment of this latent reservoir. In this study, we take advantage of a cohort of patients on HAART whose plasma virus consists largely of one (patient-specific) predominant plasma clone (PPC) that is grossly underrepresented in resting CD4+ T cells. We use the PPC as a label for ongoing viral replication by observing the accumulation of the PPC in resting CD4+ T cells over time in each patient. Analysis of the rate at which the PPC accumulates in resting CD4+ T cells allows us to quantitatively infer the maximum inflow of cells into the latent reservoir for HIV. Thus, we are able to provide the first quantitative constraint to our knowledge on the replenishment rate of the latent reservoir in the setting of HAART. Our results indicate that the rate of replenishment is very small and likely not a source of stability in the setting of HAART. These results have important implications regarding therapeutic options for purging the resting CD4+ T cell reservoir and curing HIV infection. Specifically, these results suggest that the intrinsic stability of latently infected resting CD4+ T cells, and not low-level viral replication, must be targeted therapeutically in order to achieve eradication of the latent reservoir.
| The discovery of a stable latent reservoir for human immunodeficiency virus type 1 (HIV-1) [1–5] in resting CD4+ T cells uncovered a major obstacle to curing HIV-1 infection and revealed limitations of previous analytical predictions concerning eradication [6]. This reservoir persists despite years of highly active antiretroviral therapy (HAART) [7–9]. The observation that suspension of treatment leads to rapid rebound in viral load [10] may reflect the persistence of latently infected CD4+ T cells and possibly other viral reservoirs, as well as some degree of active viral replication that continues despite HAART [11–17].
The mechanism underlying the stability of the latent reservoir remains unclear. Some investigators have argued that residual viral replication continuously reseeds the latent reservoir [11,18,19], thereby providing long-term stability. With extremely sensitive methods, a low level of free virus can be detected in the plasma of patients on HAART who have suppression of viremia to below the limit of detection of ultrasensitive clinical assays [12,20–24]. In addition, many patients on HAART have transient episodes of detectable viremia termed blips [25–27]. These findings suggest that patients on HAART have a low level of viremia that may replenish the latent reservoir in resting CD4+ T cells through de novo infection of cells that then enter the reservoir. The other major explanation for the persistence of the latent reservoir is that the stability arises from the intrinsic dynamic properties of the latently infected cells. Because the reservoir consists of resting memory T cells [28], which form the basis of life-long immunity to previously experienced pathogens, the cells that harbor latent HIV-1 are fundamentally long-lived. Latently infected cells are protected from host immune responses because there is little or no transcription of viral genes in these cells [29,30]. The fact that some patients do not develop drug resistance despite long periods of HAART supports the idea that the virus can persist through mechanisms that do not involve continuous cycles of replication. Thus, the intrinsic stability of latently infected cells provides a plausible alternative explanation for the stability of the reservoir [21,31].
Curing HIV-1 infection will require elimination of the latent reservoir. It is therefore critical to understand which of these potential mechanisms are responsible for its stability. As a step in this direction, we have used mathematical modeling to understand the dynamics of the reservoir. Mathematical models have proven useful for analysis of several aspects of HIV-1 infection, including measurement of the turnover of different T cell subsets [32–35] and the response to therapy [6,36–44]. In this study, we take advantage of (and experimentally extend) a data set consisting of HIV-1 sequences from patients on HAART who maintain a distinctive pool of plasma virus [45]. In these patients, most of the residual viremia is comprised of a single predominant plasma clone (PPC) that is specific in sequence to each patient. Using this PPC as a label and a simple mathematical model, we take a maximum likelihood approach to quantitatively constrain the rate at which de novo infection replenishes the latent reservoir.
The procedures for obtaining and analyzing the sequences used in this study have been described in detail elsewhere [45]. Briefly, we studied asymptomatic HIV-1-infected adults who had achieved suppression of viremia to <50 copies/ml on a stable HAART regimen for ≥6 mo and were willing to make frequent study visits. We previously described five patients who each had a PPC. For this present study, we exclude two of these patients (pt. 113 and pt. 139) due to a lack of sufficient follow-up sequence information beyond the intensive sampling period. The characteristics and treatment histories of the patients included in our analysis (pts. 135, 148, and 154) are representative of many HAART patients (exhibiting frequent and infrequent blips, on different HAART regimens with viral suppression from roughly 1.5 y to over 6 y) as previously described [46]. One of these patients, pt. 154, experienced (previous to this study) sequential failure of AZT monotherapy followed by failure of a three-drug HAART regimen and also exhibited multiple blips during the course of the study. Of all study participants, then, pt. 154 best represents the potential for ongoing viral replication (and therefore replenishment of the latent reservoir) in the setting of HAART.
To allow consistent amplification and sequencing of the small number of viral genomes present in the plasma of patients with viral loads below 50 copies/ml, plasma virus was first pelleted by ultracentrifugation, and then analyzed by limiting dilution reverse transcriptase (RT)-PCR, cloning, and sequencing using a previously described ultrasensitive genotyping method [45]. Viruses persisting in the resting CD4+ T cell reservoir were analyzed by a novel limiting dilution PCR assay [45]. Resting CD4+ T cells were purified from peripheral blood mononuclear cells by magnetic bead depletion as previously described [45]. As we have previously shown [3], these cells do not produce virus without stimulation and therefore by definition harbor latent virus. A segment of the pol gene encompassing all of protease and the first 219 amino acids of RT was amplified with nested PCR under limiting dilution conditions that ensure that each positive reaction has a ∼90% probability of being clonal. Products of positive PCR reactions were directly sequenced. Reactions containing more than one distinct template were identified by examination of chromatograms and excluded from the analysis.
As described previously, sequence analysis was carried out using techniques designed to avoid PCR resampling and PCR error [45]. Independent sequences that were identical to one another throughout this region of RT were identified using Varplot (kindly provided by Stuart Ray, Johns Hopkins University). Care was taken to avoid PCR errors in the sequence analysis. Proviral DNA samples were analyzed by limiting dilution PCR and direct sequencing. This approach has the advantage of eliminating PCR errors except for those that occur in the first or second cycle. For plasma virus, RT-PCR reactions were set up at limiting dilution, and positive reactions were cloned. Multiple clones were sequenced from each reaction. This allowed ready recognition of PCR errors as mutations appearing in only one clone from a set of clones obtained from a limiting dilution reaction. PCR errors were detected at a frequency that was no greater than the frequency expected based on a formal error analysis carried out on viral RNA from a cloned laboratory isolate of HIV-1 that was amplified under the same conditions [45]. These errors were reverted to patient consensus. Phylogenetic analysis was carried out on a segment of the RT coding region representing amino acids 38–219 as previously described [45]. For each time point, pie charts were constructed in which distinct taxa present at that time point were represented as separate slices, with the size of each slice being proportional to the number of independent clones isolated with that sequence.
The number of latently infected cells carrying replication-competent virus was quantified as previously described [47]. The number of HIV DNA-containing resting CD4+ T cells was determined by quantitative real-time PCR [45].
To evaluate replenishment quantitatively, we used a simple mathematical model to represent the dynamics of the latent reservoir in patients on HAART who had suppression of viremia to <50 copies/ml and whose residual viremia was largely composed of a PPC. Because only a small number of plasma virus sequences can be obtained from a given blood sample when the viral load is below 50 copies/ml, patients underwent intensive (every other day) plasma sampling over a 3-mo period, and data from this period of intensive sampling were pooled. Consistent with our sequencing data, we assumed that there were no latently infected cells containing PPC at the beginning of the 3-mo sampling period. Otherwise, we did not make any assumptions about the origin of the PPC. We make the conservative assumption that the PPC first appears at the beginning of the observation period (time t = 0), despite the fact that the PPC could have been present before we detected it. Sequencing and fitness studies (including direct examination of env, RT, and protease genes) have revealed no significant functional differences between the PPC and other sequences from the relevant patient [45]. Furthermore, we have previously detected the PPC in resting CD4+ T cells [45] (Table 1), albeit at a very low frequency, suggesting that the PPC is replication competent. Therefore, we assume that the PPC is infectious and that once the PPC appears in the plasma, it should begin entering the latent reservoir if there is any inflow into the reservoir. We assume that the PPC has permanently disappeared from the plasma (and thus can no longer enter the reservoir) after a period of time, te. In order to be maximally conservative, we allow te to be at most the period of time that we experimentally observed the PPC in each patient's plasma (although it is not certain that the PPC has entirely disappeared at later time points). Because we use the PPC as a label for new entrants into the latent reservoir, we use a previously described mathematical model used for tracking labeled cells in studies of T cell dynamics in the setting of HIV infection [48,49]. Our model consists of two state variables representing the fraction of latently infected cells containing PPC proviruses (L1) and latently infected cells containing all other proviruses (L2) where
where kin is the rate constant for the entry of free viruses into the latent pool and kout is the decay rate of latently infected cells. Equation 1 has solution
where we introduce
We extend this simple model of reservoir dynamics so that we consider L1(t) and L2(t) separately. We assume that these two populations of latently infected cells have the same kinetic properties and that the rate of replenishment of the latent reservoir is proportional to the fraction of the plasma virus of a given type. This model is described by a system of two ordinary differential equations,
where f is the fraction of plasma virus that consists of the PPC. Equations 3 and 4 may be solved explicitly:
Using Equations 5 and 6, we can predict the dynamics of each latent reservoir population and therefore the fraction of latently infected cells in each pool at various time points while the PPC was present. At time points t after the PPC had disappeared from the plasma (at time te), the two latent reservoir pools are described by the equations
which have explicit solutions:
The fraction of latently infected cells at time t that should contain the predominant plasma clone, ΦPPC, can be calculated from Equations 9 and 10 as
To be conservative, the initial number of latently infected cells containing PPC DNA (L1(0)) was set to zero for each patient. Because blood samples from time t = 0 were unavailable, the initial number of latently infected cells containing all other viral DNA sequences (L2(0)) was determined by extrapolation from experimental measurements by limiting dilution PCR at various time points for each patient (Table 2). We assume that the probability of finding k PPC cellular sequences out of n total latent reservoir sequences follows a binomial distribution, with the probability of success at time t set to the solution of Equation 11 at time t:
We used a maximum likelihood approach in order to find the value of kin most consistent with data collected from each patient [50]. Briefly, for specific values of kout, likelihood estimates, P(k), were calculated for a large range of kin values from Equations 12 using data collected from patients 135, 148, and 154. The kin value corresponding to the maximum likelihood estimate was designated as the most likely kin value.
We report analyses of patients that yielded informative results for kin (i.e., kin ≤ koutL).
All simulations and calculations were performed with MATLAB version 7.2.0.232 (http://www.mathworks.com/).
In a previous report, we described a population of HIV-1-infected individuals on HAART who had suppression of viremia to less than 50 copies/ml for an average of 34 mo [46]. Sequences of the residual plasma virus were obtained from these individuals by intensive sampling (three times per week) over a 3-mo period as well as at various intervals afterward for more than 1 y. Sequences from proviruses in resting CD4+ T cells (from our previous work, we know that these sequences are a reasonable surrogate for rescuable virus in the same population of cells [45]) were obtained at the beginning and end of the period of intensive sampling. In half of these patients, a single, homogenous but distinct viral sequence constituted a large fraction of the residual viremia but was profoundly underrepresented within sequences from resting CD4+ T cells at baseline (Figure 1). Linkage analysis suggested that this predominant plasma sequence actually represented a single PPC [45]. Because the PPC sequence was easily and specifically distinguishable from other sequences as a single sequence that was repeatedly detected in the plasma, this situation provided the ideal setting for determining whether the residual viremia could replenish the latent reservoir. Since most of the residual viremia was comprised of a unique genotype rarely found in resting CD4+ T cells, entry of a substantial number of these plasma viruses into the latent reservoir at later time points could be readily detected. Therefore, we were able to employ this unique plasma virus population as a functional label for measuring the rate of replenishment for the latent reservoir in the setting of HAART.
In order to find the kin value most consistent with each patient's data, we take a maximum likelihood approach. To calculate a likelihood estimate for each kin tested, however, we must approximate kout. The dynamics of the latent reservoir may behave according to one of three regimes (described in Text S1), depending on the magnitude of koutL(t) compared to kin (Figure 2). In order to cover the most likely kin values for all possible levels of replication in the setting of HAART, we approximate kout for when the latent reservoir decays exponentially (
, where
is an experimentally determined half-life of the latent reservoir) and for when the reservoir does not decay (kout = kin/L(0)) (regimes 1 and 2, respectively).
For each time point when cellular sequences were obtained, the most likely kin was chosen as the one with the maximum likelihood estimate, given the data (Table 1). We also calculated a 95% confidence interval around this value and use the upper bound of the 95% confidence interval as a conservative constraint on the maximum value of kin. Experimentally, at all but one time point, we found no PPC sequences in the latent reservoir. Consequently, the most likely value of kin calculated by the maximum likelihood approach was equal to zero cells per day. However, the fact that we found no PPC sequences in the latent reservoirs of our patients at these time points does not definitively indicate that there were no PPC sequences in the reservoir. In order to assign a more conservative frequency for the PPC sequence within the latent reservoir, we repeated all likelihood analyses with an add-one estimator [51], where 1 was added to the number of cellular PPC sequences obtained at each time point and 2 was added to the total number of cellular sequences obtained at each time point. Because the add-one approach already overestimates the presence of the PPC in the latent reservoir, the most likely kin calculated with the add-one estimator (mean value and not upper bound of the 95% confidence interval) is used as a conservative constraint on the reservoir inflow rate. To calculate the flow of replication-competent viruses into the latent reservoir, each patient's calculated kin values were multiplied by α, the experimentally determined ratio of cells carrying replication-competent virus to cells carrying HIV-1 DNA (Table 2). From this point forward, kin will refer to the flow of replication-competent virus into the latent reservoir.
For the case when kout dominates reservoir decay, we set kout = 0.000525 day−1, which reflects the previously reported 44-mo half-life of the latent reservoir. Initially, we assumed that the PPC was present for only the 90-d intensive sampling period despite the fact that two of the three patients maintained the PPC beyond this time. Under these assumptions, our maximum likelihood analysis found a most likely reservoir inflow rate of 0 cells/day, with an upper 95% confidence bound of 151 cells per day in pt. 135 (Table 3). The add-one estimator predicted a most likely kin of 94 cells per day in this patient. Similar rates were observed for pt. 148, and slightly higher rates were observed for pt. 154. Pts. 135 and 154 both maintained the PPC in the plasma at levels comparable to or above f beyond the intensive sampling period (until study day 174 for pt. 135 and day 922 for pt. 154) (unpublished data). To more accurately reflect the persistence of the PPC, we repeated the maximum likelihood analysis and assumed the PPC was present from time zero until the last study day when the PPC was observed in the plasma. Under this assumption, we again found a most likely inflow rate of 0 cells/day for pt. 135. Using the upper bound of the 95% confidence interval, we found that the data are consistent with kin up to 111 cells per day (Table 3). The add-one estimate predicts the data to be consistent with kin up to 70 cells per day in this patient. Similar or slightly higher rates were calculated for the other patients.
Because the true decay rate of the latent reservoir remains controversial, we repeated the above analyses for kout = 0.0039 day−1 (Table 3), which reflects a 6-mo half-life for the latent reservoir, the fastest reservoir decay rate reported [19,52]. By performing our analysis for the slowest (44-mo half-life) and fastest (6-mo half-life) reported reservoir decay rates, we cover the entire range of possible kin values. With a 6-mo reservoir half-life and the assumption of the 90-d step input of the PPC, kin was constrained to up to 335 cells per day by the data for pt. 135, with the add-one estimate constraining kin to 209 cells per day. Repeating the analysis for experimentally determined patient-specific PPC inputs, we found kin to be constrained to up to 232 cells per day by the data and 147 cells per day with the add-one likelihood estimate for this patient.
In our analysis above, we used an approximation of kout based on an assumed exponential decay of the latent reservoir (regime 1). The true range of kin values for each patient is bounded by the range of kin values calculated for regime 1 (latent reservoir decays exponentially) and regime 2 (latent reservoir is at steady state). We therefore calculated kout for the latent reservoir at steady state and repeated the analyses described above (Table 3). We initially assumed a 90-d step input of the PPC into the plasma of each patient and found that our analysis found a most likely reservoir inflow rate of 0 cells/day, with an upper 95% confidence bound of 151 cells per day. The add-one estimator predicted that the data are consistent with a most likely kin equal to 96 cells per day. When we extended the period of PPC presence in the plasma until the last study day when it was last observed for each patient, we found that the data constrained kin to be up to 112 cells per day for pt. 135, with the add-one estimator constraining kin to be up to 72. The kin estimates for the steady-state assumption are not significantly different from the kin estimates made above with kout dictating the reservoir decay rate, suggesting that the calculated maximum kin values consistent with all of the data are good approximations for the true maximum rate of inflow into the latent reservoir.
It is also helpful to view the maximum absolute flow rate into the reservoir in the context of the overall reservoir size in each patient. We approximate the percentage of the total reservoir that daily reservoir inflow represents by normalizing the daily reservoir inflow (the upper bound of the 95% confidence intervals for each kin calculated based on the pure data or the most likely kin calculated with an add-one estimate in Table 3) with the starting number of replication-competent cells in the latent reservoir (α(L1(0) + L2(0))) (Table 4). Our calculations indicate that for each patient, maximum daily flow of infected cells into the reservoir is very small compared to the total reservoir size. In fact, most maximum kin calculations were on the order of 0.01%–0.1% of the starting total reservoir size and even as low as < 0.001%. Because the total reservoir size decreases over time and we assume a constant kin, the values in Table 4 increase over time. On the time scale of when samples were taken from patients, however, none of the values in Table 4 would increase by more than 2-fold if assuming a 44-mo reservoir half-life, and most would not increase by more than 4-fold if assuming a 6-mo reservoir half-life.
We also consider whether our predicted maximum daily inflow of cells into the latent reservoir during HAART reflects a reduction from the predicted pre-HAART inflow. We have previously described how long each patient in this study had consistently suppressed viremia (79 mo for pt. 135, 35 mo for pt. 148, and 17 mo for pt. 154 from Table 1 of [45]) prior to enrollment in our study, and we have measured the size of each patient's latent reservoir at the beginning of our study (Table 2). From these values, and assuming a reservoir half-life of 44 mo (corresponding to kout = 0.000525 day−1), we are able to back calculate the most recent size of each patient's pre-therapy latent reservoir,
. The steady-state, pre-HAART reservoir inflow for each patient may be calculated by
. Based on this calculation and our lowest predicted upper bound on kin for each patient in Table 3, we find that HAART has reduced the daily inflow into the latent reservoir by at least (since we calculate the maximum reservoir inflow rate) a factor of 10.7 for pt. 135, 2.2 for pt. 148, and 18.8 for pt. 154, if we assume a 44-mo half-life for the resting CD4+ T cell reservoir. These fold reductions are several orders of magnitude larger if we assume a 6-mo half-life. These results indicate that HAART can drastically reduce the flow of cells into the resting CD4+ T cell reservoir.
The latent reservoir for HIV-1 in resting CD4+ T cells is the primary known barrier to eradication of HIV-1 infection. Therefore, eradication of HIV-1 infection depends on successful purging of the latent reservoir from an infected individual. Unfortunately, experimental evidence has shown that the latent reservoir is highly stable. Whereas the half-lives of other types of infected cells in the setting of HAART are on the order of days to weeks, the half-life of the latent reservoir has been reported to be on the order of months to years [7–9]. The longevity of the reservoir requires HIV-1-positive individuals to remain on HAART for their entire lives [53].
The basis for the stability of the latent reservoir is controversial. Some reports have shown an increase in the reservoir decay rate with an intensified HAART regimen [18], suggesting that low-level viral replication continuously replenishes the reservoir [11,18,19]. Because the latent reservoir resides within memory CD4+ T cells, which are inherently long-lived cells, some have hypothesized that the latent reservoir's longevity stems from its intrinsic stability [7–9,20,28]. Longitudinal studies of patients on standard HAART have shown no generation of new drug-resistant virus in plasma, suggesting a halt in viral replication [21]. These studies support the notion that the latent reservoir is intrinsically stable, consistent with the known properties of memory CD4+ T cells. It is difficult to find direct, experimental support for either argument, however, due to the lack of a readily accessible experimental model.
Opportunities do arise when patient-derived data may be used to gain unique insights into the latent reservoir. In this study, we offer a quantitative glimpse into the replenishment of the latent reservoir in the setting of HAART. Although several studies have suggested that reservoir replenishment might occur during HAART [11,18,19], there has been no quantitation of the replenishment rate. This has been due in part to the fact that there is no way of uniquely labeling latently infected resting CD4+ T cells and subsequently following that label. In this study we take advantage of a previously reported phenomenon where a unique, patient-specific viral sequence (PPC) dominated the residual plasma virus but could not be readily found in the patient's activated or resting CD4+ T cells [17,45]. For our study, we use and extend a previously reported, exhaustive data set of plasma and proviral sequences [45]. We hypothesized that replenishment of the latent reservoir by viral replication in the presence of the PPC would eventually lead to incorporation of the PPC into the latent reservoir. We used a simple mathematical model of latent reservoir dynamics to constrain the maximum rate of reservoir replenishment by viral replication in the setting of HAART.
Our model was constructed with as few assumptions as possible regarding the nature of reservoir dynamics. Our model does rely on the assumption that the PPC is replication competent, or at least capable of infecting and integrating into the genome of a CD4+ T cell. Our previous study strongly suggested that each patient's PPC was not different than other patient-specific plasma virus sequences in functionality (by direct examination of env, RT, and protease genes) or resistance to antiretroviral drugs or host-mediated immune responses [45]. Furthermore, detection of the PPC at a low frequency in resting CD4+ T cells [45] was highly suggestive of the idea that the PPC is replication competent. We therefore considered the PPC to be the same as other viral sequences with respect to replenishing the reservoir. However, even a non-infectious PPC would be indicative of minimal viral replication in the setting of HAART, because the PPC comprises such a large portion of the plasma virus (and it would be unlikely that all other plasma virus is produced from viral replication with no release from the reservoir). All other assumptions that were made would only artificially increase the calculated maximum reservoir replenishment rate. In particular, we assume that the PPC is first present in the plasma when we first detect it experimentally, whereas it may easily have been present in the patient much earlier. We have also assumed that any PPC sequence found in resting CD4+ T cells is due to infection of that cell after initiation of HAART. Clearly this is not the only possibility, as a PPC sequence may have entered a resting CD4+ T cell before HAART was initiated.
By applying this approach to data derived from three independent HIV-1-infected individuals on HAART who exhibited a PPC, we have been able to quite conservatively constrain the replenishment rate of the latent reservoir to be at most on the order of 100 cells carrying replication-competent virus per day. Given that the average size of the latent reservoir is approximately one million cells [1], we have therefore constrained the daily reservoir inflow to be approximately 0.01%–0.1% of the total reservoir size for the average HIV+ patient on HAART. Our results predict a substantial reduction in the reservoir inflow in the setting of HAART compared to pre-HAART levels. While we could not demonstrate a drastic reduction in reservoir inflow for pt. 148 due to the limited number of available sequences, we were able to show in other patients that HAART reduces the reservoir inflow by at least 10- to 20-fold from pre-HAART levels. Given pt. 154′s treatment history of frequent blips suggestive of low-level viral replication, that we are able to predict a ∼20-fold reduction in reservoir replenishment suggests that HAART would reduce the reservoir replenishment rate by even more in patients like pt. 135 and pt. 148, who exhibit no signs of potential viral replication. Subsequent analyses and procurement of additional sequence data may allow us to reduce the maximum replenishment rate of the reservoir even further. Our present results, however, do not establish whether or not there is any replenishment of the latent reservoir by low-level viral replication in the setting of HAART. Our analysis uses patient-derived data to conservatively constrain the maximum replenishment rate of the latent reservoir in the setting of HAART.
Of the three patients in our study, we detected the PPC in the resting CD4+ T cell compartment of one. We believe that the infrequent detection of PPC in resting CD4+ T cells of our study participants reflects the following: 1) few resting CD4+ T cells contain PPCs, and 2) replenishment of the latent reservoir in the setting of HAART must be slow, which is consistent with our results. Furthermore, while we can only infer a maximum daily inflow into the latent reservoir (since we cannot sequence the entire latent reservoir, but rather only a sample of the latent reservoir at each time point), the actual replenishment rate may easily be much lower than our calculated maximum replenishment rate and may even be zero. Finally, because the data have constrained the maximum reservoir inflow rate to be small compared to the total reservoir size, it may be that the flow of new cells into the reservoir does not significantly affect the decay rate of the latent reservoir in these patients (regime 1 versus regimes 2 and 3 described above).
The finding that the daily inflow into the reservoir is small compared to the overall reservoir size suggests that the decay of the reservoir in our patients (who have all been on HAART for several years) is more likely determined by kout (the intrinsic decay rate of latently infected cells) and not kin (new entry into the reservoir). Our model, however, predicts that the latent reservoir will eventually achieve a new steady-state level of kin/kout. If there is an inflow into the latent reservoir, despite HAART, further intensification of HAART may reduce the steady-state latent reservoir level or even lead to eradication of the reservoir. Because our results indicate that inflow into the reservoir must be very small and therefore probably does not affect the reservoir decay rate, an immediate benefit from HAART intensification may not be apparent. In fact, our results suggest that HAART intensification could at best cause the latent reservoir to decay with a half-life of on average 43.5 to 118 mo (Text S2).
The results of our analysis are important on a practical level. It has been suggested that intensification of HAART may stop residual viral replication in the setting of standard HAART and increase the decay rate of the latent reservoir [18]. HAART intensification poses a problem for physicians and patients because intensified HAART also leads to intensified drug toxicities such as lipodystrophy, hepatotoxicity, and gastrointestinal symptoms [54–56]. Drug toxicity not only leads to patient morbidity but also motivates non-adherence with subsequent development of drug resistance by the virus. Further analyses may enlighten the cost versus potential benefit of intensified HAART and hopefully maximize the clinical benefit or treatment for patients with minimal morbidity. |
10.1371/journal.pntd.0000199 | Improved Protective Efficacy of a Species-Specific DNA Vaccine Encoding Mycolyl-Transferase Ag85A from Mycobacterium ulcerans by Homologous Protein Boosting | Vaccination with plasmid DNA encoding Ag85A from M. bovis BCG can partially protect C57BL/6 mice against a subsequent footpad challenge with M. ulcerans. Unfortunately, this cross-reactive protection is insufficient to completely control the infection. Although genes encoding Ag85A from M. bovis BCG (identical to genes from M. tuberculosis) and from M. ulcerans are highly conserved, minor sequence differences exist, and use of the specific gene of M. ulcerans could possibly result in a more potent vaccine. Here we report on a comparison of immunogenicity and protective efficacy in C57BL/6 mice of Ag85A from M. tuberculosis and M. ulcerans, administered as a plasmid DNA vaccine, as a recombinant protein vaccine in adjuvant or as a combined DNA prime-protein boost vaccine. All three vaccination formulations induced cross-reactive humoral and cell-mediated immune responses, although species-specific Th1 type T cell epitopes could be identified in both the NH2-terminal region and the COOH-terminal region of the antigens. This partial species-specificity was reflected in a higher—albeit not sustained—protective efficacy of the M. ulcerans than of the M. tuberculosis vaccine, particularly when administered using the DNA prime-protein boost protocol.
| Buruli ulcer (BU) is an infectious disease characterized by deep, ulcerating skin lesions, particularly on arms and legs, that are provoked by a toxin. BU is caused by a microbe belonging to the same family that also causes tuberculosis and leprosy. The disease is emerging as a serious health problem, especially in West Africa. Vaccines are considered to be the most cost-effective strategy to control and eventually eradicate an infectious disease. For the moment, however, there is no good vaccine against BU, and it is still not fully understood which immune defence mechanisms are needed to control the infection. The identification of microbial components that are involved in the immune control is an essential step in the development of an effective vaccine. In this paper, we describe the identification of one of these microbial components, i.e., antigen 85A, a protein involved in the integrity of the cell wall of the microbe. Our findings obtained in a mouse model now need to be extended to other experimental animals and later to humans. Combination with a vaccine targeting the toxin may be a way to strengthen the effectiveness of the vaccine.
| Buruli ulcer (BU), also known as Bairnsdale ulcer, is an infectious, necrotizing skin disease caused by Mycobacterium ulcerans (M. ulcerans) occurring mostly in tropical and subtropical areas. Cases have been reported in several countries in West and Central Africa, in Central and South America, in Southeast Asia and in Australia. BU is emerging as a serious health problem, especially in West Africa, where it is the third leading cause of mycobacterial disease in immunocompetent people, after tuberculosis and leprosy. In some countries in Africa, thousands of cases occur annually and in these areas BU has supplanted leprosy to become the second most important human mycobacterial disease. The natural history of M. ulcerans infection and subsequent development of BU is not completely elucidated. M. ulcerans bacteria have been found in endemic areas in stagnant water or slowly moving water sources and in aquatic snails and carnivorous insects [1],[2]. So far, person to person transmission has not been reported. The infection causes initially a painless nodular swelling which can eventually develop into an extensive necrotizing lesion. M. ulcerans has the particularity to produce a family of toxin molecules, the so-called mycolactone (ML), polyketides that can suppress the immune system and destroy skin, underlying tissue and bone, causing severe deformities [3]–[5]. ML suppresses the in vitro TNF-α production by murine macrophages infected with M. ulcerans (4) and it strongly affects the maturation and the migratory properties of DC [5]. On the other hand, ML does not seem to affect the production of the inflammatory cytokine MIP-2, involved in the recruitment of neutrophils (4). M. ulcerans has an initial intracellular infection stage but virulent ML producing strains induce apoptosis of the infected cells and can subsequently be found extracellularly [3],[6]. Only few Mycobacterium species produce mycolactone toxins [7]. M. ulcerans isolates from different geographical areas produce different types of mycolactone, i.e. mycolactone A/B, C, D, E and F [8],[9].
The nature of immune protection against M. ulcerans infection remains unclear. In general, resistance to intracellular bacteria is primarily mediated by T cells with pivotal roles of Th1 type cytokines IFN-γ and TNF-α and this apparently is the case for M. ulcerans infection as well [10]. Progression of active Buruli ulcer is characterized by gradual down regulation of systemic and local Th1 type immune responses. Peripheral blood mononuclear cells from Buruli ulcer patients show reduced lymphoproliferation and IFN-γ production in response to specific stimulation with M. ulcerans [11]–[13]. Reduced IFN-γ response does not seem to be caused by decreased interleukin-12 production [14]. Also, semi-quantitative RT-PCR analysis demonstrated high IFN-γ and low IL-10 levels in early, nodular lesions whereas low IFN-γ and high IL-10 mRNA levels are observed in late ulcerative lesions [13]. Using a similar RT-PCR comparison of granulomatous versus non-granulomatous lesions, Phillips et al demonstrated higher expression of IL-12p35, IL-12p40, Il-15, IL-1β and TNF-α in patients from the former group and higher expression levels of IL-8 (human homologue of MIP-2) in the latter group [15]. Finally, Kiszewski et al have also confirmed that in ulcerative lesions without granuloma, there is increased expression of IL-10 and higher bacillary counts. [16].
It is not yet clear whether antibodies play a protective role against BU but the humoral immune response during M. ulcerans infection may be useful for serodiagnosis of BU. In contrast to tuberculosis and leprosy, immunoglobulin IgG antibody production against M ulcerans can be found even in early stages of infection [17]. IgG antibodies cannot be used to readily discern between patients and family controls, but primary IgM antibody responses against M. ulcerans culture filtrate proteins can be detected in sera from 85% of confirmed BD patients and only in a small proportion in sera from healthy family controls [18]. Antibody responses against the M. ulcerans homologue of the M. leprae 18-kDa small heat shock protein -that has no homologues in M. bovis and M. tuberculosis- can be used as serological marker for exposure to M. ulcerans [19].
BU results in considerable morbidity. Because of the late detection of the disease, treatment is principally by excision of the lesion, sometimes necessitating skin grafting [20]. WHO is currently recommending combined rifampicin and streptomycin treatment of nodules for eight weeks in the hope of reducing the need for surgery [21],[22]. Unfortunately, there is no specific vaccine against BU for the moment [23]. M. bovis BCG (Bacille Calmette et Guérin) vaccine, used for the prevention of tuberculosis, has been reported to offer a short-lived protection against the development of skin ulcers [24]–[26] and to confer significant protection against disseminated cases of BU, e.g. osteomyelitis, both in children and in adults [27],[28]. The precise M. ulcerans antigens that induce a protective immune response are poorly defined. The complete genome sequence of M. ulcerans has recently been published and will hopefully help to advance research and identification of relevant genes [29]. The 65 kD heat shock protein is expressed in considerable amounts by M. ulcerans bacilli in vitro and in vivo, and is immunogenic for both B and T cells in mice. Nevertheless, vaccination of mice with plasmid DNA encoding Hsp65 from M. leprae, having 96% sequence identity with Hsp65 from M. ulcerans, limited only weakly the progression of experimental M. ulcerans infection in tail [30]. We have previously reported that vaccination with BCG or with plasmid DNA encoding Ag85A from M. bovis BCG can partially protect B6 mice against footpad challenge with M. ulcerans [31]. Antigen 85 is a major secreted component in the culture filtrate of many mycobacteria such as M. bovis BCG, M. tuberculosis and M. avium subsp. paratuberculosis [32]. The antigen 85 complex (Ag85) of M. tuberculosis is a family of three proteins, Ag85A, Ag85B and Ag85C, which are encoded by three distinct but highly paralogous genes and that display an enzymatic mycolyl-transferase activity, involved in cell wall synthesis [33],[34]. Members of the Ag85 family rank among the most promising tuberculosis vaccine candidates, and are actually being tested in clinical trials, formulated as Hybrid-1 fusion protein of Ag85B with ESAT-6 or as recombinant Modified Vaccina Ankara virus encoding Ag85A booster vaccine of BCG [35],[36]. We have previously sequenced the gene encoding Ag85A from M. ulcerans and reported that it shares 84.1% amino acid sequence identity and 91% conserved residues with the gene encoding Ag85A from M. tuberculosis (which is identical to the Ag85A gene of M. bovis BCG) [31]. The genes encoding Ag85B and Ag85C of M. ulcerans have recently been sequenced as well and – as for M. tuberculosis- were localized on different loci in the genome [29].
Here, we report on a comparison of the immunogenicity and protective efficacy of vaccines encoding Ag85A from M. tuberculosis and from M. ulcerans. Vaccines were administered as plasmid DNA, purified protein in adjuvant or in a DNA prime-protein boost protocol. We and others have previously reported that DNA priming followed by protein boosting is an effective means to increase the potential of DNA vaccines [37]–[40].
C57BL/6 mice were bred in the Animal Facilities of the IPH-Pasteur Institute Brussels, from breeding couples originally obtained from Bantin & Kingman (UK). Mice were 8–10 weeks old at the start of the experiments. Female mice were used for immune analysis and male mice for the protection studies. This study has been reviewed and approved by the local Animal Ethics Committee (file number 030212/05).
Virulent M. ulcerans type 1 strain 04-855 from a Benin patient was isolated at the Institute for Tropical Medicine in Antwerp, Belgium. Bacteria grown on Löwenstein-Jensen medium were maintained and amplified in vivo in footpad of the mice. M. bovis BCG strain GL2 was grown for 2 weeks as a surface pellicle at 37°C on synthetic Sauton medium and homogenized by ball mill as described before [41].
Plasmid DNA encoding the mature 32 kD Ag85A from M. tuberculosis in V1J.ns-tPA vector was prepared as described before [31],[42]. The gene encoding Ag85A from M. ulcerans was amplified by PCR without its mycobacterial signal sequence using BglII restriction site containing primers and ligated into the same V1J.ns-tPA vector. The primers used were 5′-GGAAGATCTTGAGCGCTTGGTACTAGGC-3′ (forward) and 5′-GGAAGATCTTTTCGCGGCCGGGCCTGCCGGTGGA-3′ (reverse). In these plasmids the Ag 85A gene is expressed under the control of the promoter of IE1 antigen from cytomegalovirus, including intron A and it is preceded by the signal sequence of human tissue plasminogen activator.
Hexa-histidine tagged Ag85A protein from M. tuberculosis was purified from recombinant E. coli as described before [43]. The gene encoding the mature Ag85A protein from M. ulcerans was amplified by PCR from V1J.ns.tPA-85A vector. The primers used were 5′-CGCGGATCCGCGTTTTCGCGGCCGGGCCTGCCGTGGAA-3′ (forward) and 5′-CCCAAGCTTGGGCTAGGCGCCCTGGGTGTCACCG-3′ (reverse) with respectively BamHI and Hind III restriction sites. Ag85A gene was amplified without its mycobacterial signal sequence. Cloning in expression vector pQE-80L (QIAGEN), containing an NH2-terminal histidine-tag coding sequence, and purification were performed as described before [32]. Briefly, positives clones were screened on LB-ampicillin medium after ligation of the gene in the vector and transformation of E. coli DH5α cells. For expression, Top-10F' E. coli (Invitrogen) cells were transformed with plasmid encoding the 85A sequence. Recombinant protein was purified by immobilized metal affinity chromatography (IMAC) using gravity flow. The endotoxin level measured with the LAL kinetic chromogenic assay, was inferior to 10 EU/ml (endotoxin units per millilitre) or 0.03 EU/µg of purified protein (Cambrex Bioscience, New Jersey, America).
Peptides spanning the entire mature 295 amino-acid Ag85A sequence of M. tuberculosis were synthesized as 20-mers, with the exception of the 18-mer spanning aa 35–53 and the 21 mer-peptide spanning amino acids 275–295 [44]. Peptides spanning the entire 294-amino acid Ag85A sequence of M. ulcerans were synthesized as 20-mers. All peptides were purchased from Ansynth Service B.V., The Netherlands.
In experiment 1, B6 mice were anesthesized by intraperitoneal injection of ketamine-xylazine and injected three times intramuscularly (i.m) in both quadriceps muscles with 2×50 µg of control V1J.ns-tPA (empty vector), V1J.ns-tPA-Ag85A DNA from M. ulcerans or from M. tuberculosis (abbreviated as Ag85A-DNA Mu and Ag85A-DNA Mtb in the figures). For protein immunization, mice were injected three times subcutaneously (s.c) in the back with 10 µg of purified recombinant Ag85A (abbreviated as rec85A-Mu and rec85A-Mtb in the figures), emulsified in Gerbu adjuvant, i.e. water miscible, lipid cationic biodegradable nanoparticles, completed with immunomodulators and GMDP glycopeptide (GERBU Biochemicals). For the DNA prime-protein boost, mice were immunized twice i.m. with Ag85A DNA from M. ulcerans or from M. tuberculosis and boosted s.c. with 20 µg of recombinant Ag85A protein respectively from M. ulcerans or M. tuberculosis in Gerbu adjuvant (abbreviated as Ag85A-DNA/recMu and Ag85A-DNA/recMtb in the figures). All mice received the two first injections at 3 week intervals and the third injection was given two months later. For BCG vaccination, mice were injected intravenously, in a lateral tail vein, at the time of the first DNA injection with 0.2 mg (corresponding to 106 CFU) of freshly prepared live M. bovis BCG [41].
In experiment 2, B6 mice were injected intramuscularly (i.m) three times, at 3 weeks intervals, in both quadriceps with 2×50 µg of control V1Jns.tPA DNA or plasmid DNA encoding 85A from M. ulcerans or from M. tuberculosis. For protein immunization, mice were injected three times subcutaneously (s.c) in the back with 10 µg of purified recombinant Ag85A from M. ulcerans or from M. tuberculosis, emulsified in monophosphoryl lipid A (MPL-A) from Salmonella enterica serovar Minnesota (Ribi ImmunoChem Research, Hamilton, Mont)) solubilized in triethanolamine. For the DNA prime-protein boost, mice were immunized twice i.m. with Ag85A DNA from M. ulcerans or from M. tuberculosis and boosted s.c. with 20 µg of purified recombinant Ag85A protein respectively from M. ulcerans or from M. tuberculosis in MPL-A.
Naïve and vaccinated B6 mice were infected with M. ulcerans 3 months (Exp1) or 6 weeks (Exp2) after the last vaccination. 105 acid fast bacilli (AFB), obtained by in vivo passage in footpad, were injected in the right footpad of the vaccinated mice. The number of bacilli injected, suspended in Dubos Broth Base medium (Difco), was determined by counting under a microscope after Ziehl Neelsen staining. Viability of the M. ulcerans inoculum was checked by plating on 7H11 Middlebrook agar, supplemented with oleic-acid-albumin-dextrose-catalase enrichment medium. Yellow colonies were counted after 8 weeks of incubation at 32°C. The number of Colony Forming Units corresponded to the number of AFB.
Vaccinated mice were sacrificed 3 weeks after the third immunization (Experiment 1). Spleens were removed aseptically and homogenized in a loosely fitting Dounce homogenizer. Leucocytes (4×106 WBC/ml) from four mice per group were cultivated at 37°C in a humidified CO2 incubator in round-bottom micro well plates individually or pooled (as indicated) and analyzed for Th1 type cytokine response to purified recombinant his-tagged Ag85A (5 µg/ml), and synthetic peptides from M. ulcerans or M. tuberculosis (10 µg/ml). Supernatants from at least three wells were pooled and stored frozen at −20°C. Cytokines were harvested after 24 h (IL-2) and 72 h (IFN-γ), when peak values of the respective cytokines can be measured.
Interleukin-2 (IL-2) activity was determined in duplicate on 24 h culture supernatants using a bio-assay with IL-2 dependent CTLL-2 cells as described before [45]. IL-2 levels are expressed as mean counts per minute (cpm). Assay sensitivity is 10 pg/ml. A typical international standard curve of this assay has been published before [46].
Interferon-γ (IFN-γ) activity was quantified by sandwich ELISA using coating antibody R4-6A2 and biotinylated detection antibody XMG1.2 obtained from Pharmingen. The standard murine recombinant IFN-γ used was obtained from R&D. The sensitivity of the assay is 10 pg/ml.
Sera from immunized mice were collected by tail bleeding 3 weeks after the third vaccination. Levels of M. ulcerans specific total anti-Ag85A Igκ antibodies (Abs) were determined by direct enzyme-linked immunosorbant assay (ELISA) in sera from individual mice (four/group). The concentration of Ab was expressed by the optical density at a dilution of 1/100 of the sera. For isotype analysis, peroxidase-labeled rat anti-mouse immunoglobulin G1 (IgG1) and IgG2b (Experimental Immunology Unit, Université Catholique de Louvain, Brussels, Belgium) were used. Isotype titers were expressed as dilution endpoints (last serum dilution with an optical density (OD) value higher than a cut-off OD value calculated from the OD value plus three standard deviations (SD) of the secondary antibody only [42].
In experiment 1 (Gerbu adjuvant), protection was evaluated by enumeration of Acid Fast Bacilli (AFB) nine weeks after footpad infection. Briefly, the skin and bones were removed from infected foot pad. Tissues were homogenized in a Dounce homogenizer and suspended in 2 ml of Dubos broth based medium containing glass bead. The number of AFB in 20 fields (surface of 1 field: 0.037994 mm2×20 with the 22 mm ocular diameter used) was counted on microscope slides after Ziehl-Neelsen staining. In experiment 2 (MPL-A adjuvant), protection was evaluated by monitoring foot pad swelling after M. ulcerans infection. The swelling was measured with a calibrated Oditest apparatus with a resolution of 0.01 mm as described previously [47]. Animals were euthanized when footpad swelling exceeded 4mm according to the rules of the local ethical commission.
For cytokine production analysis, antibody production and AFB counting, statistical analysis was made according to one-way ANOVA test. Subsequent multiple comparison between the 7 different groups of animals and the antigens used was made by a Tukey's correction test. Statistical results are represented in the figure by *** (P<0.001), ** (P<0.01) and * (P<0.05). For the comparison of survival curves, logrank test was used.
Spleen cells from mice vaccinated with the three different vaccine formulations produced significant levels of IL-2 (Figure 1A) and IFN-γ (Figure 1B) after in vitro stimulation with purified recombinant Ag85A from M. ulcerans or from M. tuberculosis. As expected from the 91% sequence similarity between both antigens, highly cross-reactive immune responses were observed, mice vaccinated with M. ulcerans vaccines recognizing the M. tuberculosis antigen and vice versa. Nevertheless, a certain level of species specificity was observed, particularly in the IL-2 responses. Confirming previous results obtained with a M. tuberculosis DNA vaccine [37], boosting plasmid DNA vaccinated mice with purified M. ulcerans protein increased significantly Ag 85A specific IL-2 and IFN-γ responses.
Significant cross-reactive antibody responses were induced against Ag85A from M. ulcerans (and from M. tuberculosis, data not shown) in mice vaccinated with the M. ulcerans and M. tuberculosis vaccines (Figure 2).
Antibody responses in DNA vaccinated mice demonstrated considerable individual variation, and were markedly increased by the protein boost. Vaccination with purified protein in Gerbu adjuvant was also very effective in inducing high level antibody production. DNA vaccination induced very little IgG1 isotype antibodies but biased predominantly an IgG2b isotype response, confirming the well known Th1 inducing properties of intramuscular plasmid DNA. In contrast, vaccination with protein emulsified in Gerbu adjuvant induced antibodies of both IgG1 and of IgG2b isotype. Confirming previous findings, DNA prime- protein boost vaccination resulted in increased and less variable antibody titers of both isotypes [37]. Vaccination with recombinant 85A protein or with the DNA prime /protein boost protocol induced significantly higher levels of total IgG and IgG1 antibodies as compared to plasmid DNA vaccination alone (P<0.001).
Despite the high level of sequence similarity (91%) between Ag85A from M. tuberculosis and M. ulcerans but in view of the partial species-specific Th1 type immune responses observed in the previous experiment, we decided to characterize the H-2b restricted immunodominant T cell epitopes, using synthetic 20-mer peptides spanning the entire mature sequence of Ag85A from M. ulcerans and from M. tuberculosis. Figure 3 shows the IL-2 and IFN-γ production induced in response to M. ulcerans peptides in mice vaccinated with M. ulcerans DNA (white bars) or M. tuberculosis DNA (black bars). Spleen cells from B6 mice vaccinated with M.ulcerans-Ag85A DNA produced significant levels of IL-2 (Figure 3A) and IFN-γ (Figure 3B) when stimulated with M. ulcerans peptides both from the NH2-terminal and COOH-terminal part of the protein, whereas IL-2 and IFN-γ responses of B6 mice vaccinated with the M. tuberculosis plasmid were almost exclusively directed against M. ulcerans peptide spanning aa 241–260. M. ulcerans DNA vaccinated mice also recognized this peptide very effectively. Responses against the NH2-terminal peptides spanning aa 61–80 and 81–100 of M. ulcerans-Ag85A were only observed in M. ulcerans DNA vaccinated mice, indicating that this NH2-terminal region was responsible for the partial species-specificity. This confirmed a previous finding (Inserts in Figures 3A and 3B) on species-specific T cell responses induced following in vitro stimulation with a purified, partial M. ulcerans Ag85A protein, spanning aa 17–150 in mice vaccinated with DNA encoding Ag85A from M. ulcerans or M. tuberculosis. IL-2 and IFN-γ responses against M. ulcerans peptide spanning aa 261–280 were also species-specific and only detected in mice immunized with the M. ulcerans vaccine.
Responses against M. tuberculosis peptides showed a reciprocal pattern (Figure 4). Confirming previous findings [37] M. tuberculosis peptide spanning aa 261–280 was very well recognized in M. tuberculosis DNA vaccinated mice (Figure 4A and 4B). It was also recognized by M. ulcerans vaccinated mice. Both DNA vaccinated groups also reacted against M. tuberculosis peptide spanning aa 241–260, previously found to contain the immunodominant H-2b restricted epitope recognized in BCG vaccinated and M. tuberculosis infected B6 mice (10). Responses against this M. tuberculosis peptide were even higher in M. ulcerans than in M. tuberculosis DNA vaccinated mice. IFN-γ responses against M. tuberculosis peptides spanning aa 121–140 and 141–160 were only observed in mice vaccinated with the M. tuberculosis DNA, whereas a cross-reactive immune responses was found against M. tuberculosis peptide spanning aa 81–100. A sequence comparison of identified immunodominant Th1 peptides of Ag85A from M. ulcerans and from M. tuberculosis, showing conserved and non-conserved amino acid changes is presented in Table 1.
Mice were challenged three months after the third vaccination with 105 AFB of M. ulcerans in the footpad. Nine weeks later, when a significant swelling of the footpad appeared in the control mice vaccinated with empty vector, all animals were sacrificed and the number of AFB in the infected footpad was determined by Ziehl-Neelsen staining. As shown in Figure 5, a significant and strong reduction in the number of M. ulcerans AFB was observed in mice previously immunized with all three types of vaccine. Vaccination with specific M. ulcerans antigen using the DNA prime-protein boost protocol with Gerbu adjuvant conferred the highest protection with an almost one-hundred fold reduction in number of AFB as compared to the control group. This protection was comparable in magnitude to the protection conferred by the BCG vaccine. Difference between the vaccinated groups was not significant (ANOVA test; p>0.05).
In a second experiment, protective efficacy of the vaccines was determined by weekly monitoring appearance and size of footpad swelling and survival as previously reported [47]. Mice were euthanized when footpad swelling was >4 mm. In this experiment mice were challenged with 105 AFB of M. ulcerans 04-855 at 6 weeks after the last immunization. The evolution of footpad swelling is shown in Figure 6A and 6C whereas the survival curves are represented in Figure 6B and 6D. In mice vaccinated with empty control vector, footpad size started to increase 5 weeks after M. ulcerans infection whereas in BCG vaccinated mice, footpad swelling was delayed for 7–8 weeks (Figures 6A and 6C). Vaccination with DNA encoding Ag85A from M. tuberculosis or from M. ulcerans delayed onset of foot pad swelling by only 2 to 3 weeks (Figure 6A). DNA prime/protein boost protocol using the M. tuberculosis Ag85A did not increase vaccine efficacy (Figure 6C) whereas vaccination with DNA encoding Ag85A from M. ulcerans boosted with the recombinant Ag85A-MPL-A protein from M. ulcerans delayed onset of foot pad swelling to the same extent as the BCG vaccine by 7 to 8 weeks (Figure 6C). Survival curves reflected the footpad swelling pattern. Median survival time of mice vaccinated with empty vector was 10.5 weeks, whereas BCG vaccination delayed significantly the moment when mice had to be euthanized, resulting in a median survival time of 17.5 weeks (Figures 6B and 6D) (p<0.001 compared to empty vector vaccinated mice; p<0.01 compared to 85A-DNA Mu vaccinated mice). M. ulcerans DNA vaccinated mice demonstrated a median survival time of 13 weeks (p<0.01 compared to empty vector vaccinated mice according to the log rank test). Similar results were observed in mice vaccinated with DNA encoding Ag85A from M. tuberculosis (median survival time 12.5 weeks) (Figure 6B). Boosting DNA vaccinated mice with protein from M. tuberculosis did not increase protective efficacy of the DNA vaccine but priming with DNA encoding Ag85A from M. ulcerans and boosting with recombinant Ag85A from M. ulcerans was very effective in increasing the protection (Figure 6D) (p<0.001 compared to M. ulcerans DNA alone, p<0.01 compared to DNA encoding Ag85A from M. tuberculosis boosted with the protein of M. tuberculosis). Median survival time in the M. ulcerans DNA primed- M. ulcerans protein boosted mice was 17 weeks. This protection was comparable to that conferred by BCG (p>0.05).
Buruli ulcer belongs to the family of neglected tropical diseases [48]. In 1998 the first International Conference on Buruli Ulcer was organized in Côte d'Ivoire, expressing the poor knowledge about this disease and calling on the international scientific community to support control and research efforts. Currently, no specific vaccine exists against this disease. In 1957, Fenner demonstrated that a high degree of protection was conferred, in an experimental mouse model, against challenge infection with small doses of M. ulcerans by prior inoculation with M. ulcerans, M. balnei and M. bovis BCG (BCG). Footpad and intravenous BCG administration gave considerable protection against a small dose and a slight protection against a large dose of M. ulcerans given in the other footpad [49]. More recently we have shown in a similar experimental mouse model that BCG vaccine protects to some extent against infection with M. ulcerans but that a booster vaccination with the same BCG vaccine does not increase the protective effect [31],[47]. In 1969, a clinical study performed in Uganda reported on a protection rate of 47% of the BCG vaccine. However, protection turned out to be short-lived and was only detected in the first 6 months following BCG vaccination [24]. In 1976, Smith et al reported another BCG vaccination trial against Buruli ulcer in Uganda giving similar short lived (one year) protection rates of about 50% [25]. Although not very effective at preventing the classical skin lesions of Buruli ulcer, the BCG vaccine seems to exert a significant protective effect against its severe, disseminated osteomyelitis form both in children and in adults [26],[27].
A more effective M. ulcerans vaccine would certainly help to control this debilitating disease that affects particularly children. Unfortunately, the nature of the protective immune response and the precise antigens involved are not fully defined at the moment. Based on biopsy specimens, M. ulcerans was originally thought to reside exclusively as free extracellular bacilli, implying that humoral responses might be protective. However, Coutanceau et al recently demonstrated that the initial phase of M. ulcerans infection proceeds by internalization of bacilli by phagocytic cells and that the extracellular stage results from mycolactone inducing host cell death [6],[50]. Therefore, recognition of the early intracellular stage by an effective Th1 type immune response may contribute to the control of the infection, that is in so far as it can help to reduce the mycolactone production. Hence, magnitude of mycobacteria-specific Th1 type immune response is a plausible correlate of protection that can be used to analyze the potential of new, experimental vaccines.
In this study, we focused on a plasmid DNA vaccine encoding Ag85A from M. ulcerans. Protective efficacy was evaluated using two approaches, in one experiment by enumerating the number of AFB in the footpad at nine weeks after M. ulcerans challenge and in the other experiment by monitoring footpad swelling and long term survival of the mice. We have previously reported that footpad swelling is correlated with bacterial replication and can be used as an alternative read-out for protection against infection [47]. DNA prime–protein boost strategy using specific M. ulcerans antigen 85A was clearly the most effective, reducing about one hundred fold the bacterial number and offering a protection of comparable magnitude as the one induced by the BCG vaccine. Nevertheless, and as for the BCG vaccine, immune protection was not sterilizing and eventually all mice developed footpad swelling. We hypothesize that the vaccines reduced or delayed temporarily mycolactone production by the virulent type 1 strain 04-855 but that immunity was not strong enough to completely block the ML synthesis. Targeting ML production by specific antibodies or by interfering with its synthesis might help to overcome this problem. A study made by Fenner, in 1956 showed that the apparition of footpad swelling depends of the number of viable AFB in the inoculum, small doses of bacilli showing delayed appearance of footpad lesion [51]. As we used a high inoculum size of 105 AFB in our studies, it is possible that more sustained protections could have been observed if we had administered a lower number of bacteria.
The Gerbu adjuvant is less well known as immunomodulator than other adjuvants such as alumn or monosphoshoryl-lipd-A (MPL-A) [52]. Here we have shown that this adjuvant has a strong Th1 inducing capacity, as indicated by the elevated levels of antigen-specific IL-2 and IFN-γ that could be detected in spleen cell cultures from mice vaccinated with protein in this adjuvant. Antibodies of both IgG1 but also of IgG2b isotype were induced, which was another indication of its Th1 favouring properties. Vaccination with recombinant M. ulcerans Ag85A protein in Gerbu adjuvant induced comparable Th1 cytokine and antibody levels as the prime-boost DNA vaccination. This protein vaccine also induced considerable protection (albeit somewhat lower that the DNA based vaccine) as indicated by significantly reduced number of AFB in the footpad at nine weeks after M. ulcerans challenge. We have previously shown that DNA vaccination induces a broader T cell repertoire (more protein epitopes recognized) than infection with tuberculosis [53],[54], vaccination with BCG [44] or with protein [46] and this may explain the better protection conferred by the DNA prime-protein boost vaccination. It is also possible that immune memory induced with this combined immunization protocol was stronger and longer lasting than immune memory induced with protein only vaccination.
Analysis of the H-2b restricted Th1 T cell epitopes of antigen 85A from M. ulcerans and from M. tuberculosis revealed some extent of species specificity, both in the NH2-terminal and in the COOH-terminal half of the protein. In contrast to the response induced with DNA encoding M. tuberculosis Ag85A, which was preferentially directed against Ag85A peptide spanning aa 261–280, T cell response induced with DNA encoding the M. ulcerans protein was directed preferentially against peptide spanning aa 240–259. Remarkably, mice vaccinated with the M. tuberculosis DNA reacted more strongly to this peptide region of M. ulcerans (25,000 cpm of IL-2/5,000 pg of IFN-γ) than to the same region in M. tuberculosis (10,000 cpm of IL-2/1,000 pg of IFN-γ). We have previously reported that B6 mice vaccinated with DNA encoding Ag85B from M. tuberculosis also react more strongly to 85B peptide spanning aa 244–260 than to peptide spanning aa 262–279 [46]. Sequence analysis of the 241–260 region of Ag85 revealed that the Ag85A sequence from M. ulcerans is more similar to the Ag85B sequence of M. tuberculosis (only 1 aa (A–D) change in position 242) than to the Ag85A sequence of M. tuberculosis (4 aa changes). Interestingly, it was demonstrated by Yanagisawa et al that vaccination of B6 mice with killed M. tuberculosis triggered preferentially a vβ11+ CD4+ T cell response against the peptide spanning amino acids 240 to 254 of Ag85B [55]. All these data taken together seem to indicate that the M. ulcerans Ag85A241–260 region is more immunogenic than the corresponding M. tuberculosis Ag85A region and this may explain the better protective efficacy that we have observed with the species specific M. ulcerans vaccine.
In conclusion, our results show that specific Ag85A-DNA priming followed by protein boosting is an effective way to induce robust Th1 type immune responses and strong protection against experimental footpad infection with M. ulcerans in mice. This is a promising vaccination approach that warrants further analysis. Combination with vaccines targeting mycolactone or with vaccines targeting enzymes involved in mycolactone synthesis may be a way to strengthen its protective efficacy. |
10.1371/journal.pcbi.1003172 | Full Design Automation of Multi-State RNA Devices to Program Gene Expression Using Energy-Based Optimization | Small RNAs (sRNAs) can operate as regulatory agents to control protein expression by interaction with the 5′ untranslated region of the mRNA. We have developed a physicochemical framework, relying on base pair interaction energies, to design multi-state sRNA devices by solving an optimization problem with an objective function accounting for the stability of the transition and final intermolecular states. Contrary to the analysis of the reaction kinetics of an ensemble of sRNAs, we solve the inverse problem of finding sequences satisfying targeted reactions. We show here that our objective function correlates well with measured riboregulatory activity of a set of mutants. This has enabled the application of the methodology for an extended design of RNA devices with specified behavior, assuming different molecular interaction models based on Watson-Crick interaction. We designed several YES, NOT, AND, and OR logic gates, including the design of combinatorial riboregulators. In sum, our de novo approach provides a new paradigm in synthetic biology to design molecular interaction mechanisms facilitating future high-throughput functional sRNA design.
| Is our current knowledge of in vivo RNA-RNA interactions and thermodynamics enough to perform the unsupervised computational design of fully synthetic sequences encoding functional RNAs in living cells? Recent work gave a positive answer for the challenging problem of designing activating riboregulators. This was done by integrating theory and computation to develop a physicochemical framework for the design of regulatory RNA systems, using Watson-Crick interactions and optimization algorithms. Still, the objective function was not directly validated, preventing using with confidence the methodology for other systems. We here validate experimentally an objective function relying on free energies of RNA complex activation and formation, which allows extending the framework to produce logic devices that can be implemented to program gene expression. We demonstrate that it is possible to design increasingly sophisticated and modular functions, pointing our results out that energy-based optimization methods can perform the large combinatorial search required for RNA design.
| Small non-coding RNA (sRNA) has raised a big interest because of the predictability and modularity of its binding with a large variety of molecules and macromolecules [1]. Given this functional potential, the use of sRNAs to control protein expression has triggered a new way to engineer integrated regulatory networks [2]. Although rational techniques have been successfully applied to redesign natural systems [3], [4], engineer synthetic ones [2], [5]–[7] and assemble modular structures [8]–[10], de novo sequence design still remains difficult because of the size and complexity of multi-state systems. To overcome this, we propose an evolutionary computation design strategy [11], where all design specifications are automatically assembled to yield an optimal solution.
In this work, we demonstrate a full design automation of RNA sequences that implement diverse riboregulatory mechanisms, able to produce several sRNA-based logic gates that are functional in living cells. We generalize our previous work [11] on the design of riboregulators for activating protein expression, which could be considered as YES gates, to derive objective functions to design riboregulators implementing several logic gates. Furthermore, we experimentally validate our objective function by considering mutants of natural and synthetic riboregulators [11], [4], and this allows assessing the generality of the methodology.
By generalizing the positive riboregulation paradigm, where an sRNA interacts through Watson-Crick pairing with a target mRNA to trigger a conformational change enabling ribosome docking, we can extend the methodology to design arbitrary logic gates, accounting for new regulatory mechanisms, such as anti-termination, and implementing constrained design strategies (Fig. 1). For that, we exploit antisense and allosteric RNA [12], [13], two conserved mechanisms based on precise secondary structures, and whose major role has been reported over the last years in bacteria [14], but also in humans [15] and plants [16]. Our method starts from random sequences to proceed with successive rounds of a mutation operator, followed by selection using an objective function that accounts for the free energies of all possible reactions and the secondary structures of all species. Previous work on full design automation of nucleic acids was focused on in vitro annealing of small DNAs [17]–[20], hammerhead ribozymes [21], or ribosome binding sites (RBSs) [22].
In the following, we will start by formulating the RNA design problem as an inverse problem to program gene expression. This is based on an optimization method that minimizes an ab initio objective function, which contrasts with other approaches [4]. We will evaluate such an objective function by engineering and characterizing our own mutant library of synthetic riboregulators activating gene expression. Afterwards, we will show and exemplify how to design sRNA-based logic gates, including complex gates involving synergistic interactions of different sRNAs as inputs. Finally, we will discuss the results stressing the limitations of our methodology.
Riboregulation is based on conformational changes, after interaction, in the structures of RNA molecules, which allow controlling protein expression. To design such regulatory RNAs, we optimize the potential energy curve defined in the transition state theory [23], minimizing the free energies of the transition and hybridization states. We assume that the individual folding state is formed before intermolecular RNA-RNA interaction, because its time scale is of milliseconds whereas hybridization takes seconds or even minutes [24], [25]. The interaction mechanism is guided by means of the seed region (nucleation site; the first nucleotides that get paired) to form an intermediate complex at the transition state [3], [11]. Then, both RNAs are destabilized to form a complex with a new structure and minimal energy.
Here, we consider the structures of all individual species as design specifications. To address the computational design, we firstly have to find sequences folding into predefined structures and, second, find sequences able to interact specifically among them to form complexes displaying the correct behavior. The structural constraints are exploited to considerably reduce the combinatorial space and accelerate the design of nucleic acid sequences. Our computational procedure optimizes at the same time all RNA sequences of the circuit. During the optimization, we do not impose constraints in nucleotide sequence, such as stems with high GC-content or loops with YUNR motifs, which have been found in natural systems [12]. Importantly, our designs are just based on basic physicochemical principles and not on additional fitting, allowing the solution of the full design problem.
But, is the proposed objective function predictive enough to allow the designability of multi-state RNA devices? To illustrate this question, we constructed here a library of mutants of one of our previously designed circuits (the device RAJ11 [11], implementing a YES logic gate as shown in Fig. 1B). Then, we represented the experimental values of the measured activation fold against the objective function calculated for those mutants (Fig. 2A). To give further support to our objective function, we evaluated it for a set of mutational variants of the IS10 antisense RNA system [4], implementing a NOT logic gate (Fig. 1A), and then we represented those values against the experimental repression folds reported (Fig. 2B). This natural system constitutes an independent validation. The objective function here (Eq. 13) accounted for the free energy of formation and the length of the seed in the sRNA-mRNA interaction. Fig. 2 shows a good correlation (without any fitting) for our objective function and experimental data, which supports the designability of those devices.
We first applied our design methodology to obtain sRNA-based repression and activation. Many known riboregulators impart a repressive action on their targets by promoting accelerated degradation through endoribonucleases, which initiate turnover of both RNAs [26]. Instead, we here account for sRNAs that bind specifically to a segment of its target mRNA in order to inhibit translation (NOT logic function) [4]. The most intuitive mechanism consists in blocking the Shine-Dalgarno sequence, which is generally located about eight base pairs upstream of the start codon (AUG), for preventing ribosome docking (Fig. 1A). For instance, in E. coli plasmid F, sRNA FinP directly binds to the 5′ untranslated region (UTR) of protein TraJ [12]. We constructed the following objective functions (definitions of ΔGkin and ΔGstr in section Methods) to solve the optimization problem(1)These functions are associated to each entry of the truth Table, and then the solution of this problem will yield NOT logic gates. In Fig. 3, we show several computational designs of this logic device. We applied our methodology with different natural occurring structures involving one, two or three hairpins for the trans-repressing sRNAs. In our designs, we used the Shine-Dalgarno sequence AGGAGA.
Although the majority of sRNA-mediated regulation in E. coli consists in repression, an sRNA can also operate as an activator (YES logic function) [2]. In this case, the sRNA trans-activates a cis-repressed gene by its 5′ UTR. After interaction, the conformational change in the 5′ UTR releases the Shine-Dalgarno sequence and allows translation (Fig. 1B). For instance, in E. coli, sRNA DsrA is responsible of activating the expression of sigma factor RpoS, which modulates the stress response [13]. Hence, we constructed the following objective functions(2)The solution of this problem will produce the intended function specified in the truth Table. This problem is much complex that the previous one because here the two RNA species have structure. In Fig. 4, we show several computational designs of YES logic gates based on conformational changes in the 5′ UTRs of the target genes. We applied our methodology with different structures for the trans-activating sRNAs, while maintaining a common structure for the 5′ UTR. We also attempted the computational design of a synthetic RNA able to interact with the RpoS 5′ UTR, and then enhance the translation rate. Fig. S2 shows the sequences and structures obtained.
In addition, we exploited our methodology to design NOT logic gates based on structured 5′ UTRs. Here, the trans-activating sRNA interacts with the 5′ UTR to induce a conformational change that blocks the Shine-Dalgarno sequence (Fig. 1C). The objective functions to solve the corresponding problem read(3)where the difference with Eqs. (1) relies on the imposition that the RBS must be paired at the intramolecular level. Fig. 5A shows a computational design implementing this regulatory mechanism. We also designed riboregulators with activation activity based on a mechanism of anti-termination [27]. This design relies on a trans-regulating sRNA able to destabilize the structure of a terminator, which is here the cis-regulating element, resulting in a complex that allows the progression of the RNA polymerase (Fig. 1D). This mechanism can also entail kinetic effects [3], where the interaction has to occur before RNA polymerase reads through the terminator. This may impose a narrow time window for operation, which we speculate surmountable provided a given free energy threshold and a high ratio sRNA/mRNA. In this case, the objective functions were(4)where the 5′ UTR encodes for a terminator that is formed in absence of the sRNA. The solution of this problem will also satisfy the truth Table for YES. Fig. 5B shows a computational design of a YES logic gate based on this mechanism. In the final structure of the complex, the terminator hairpin is destabilized and the poly(U) tail does not have any effect.
We then applied our methodology for the design of higher-order riboregulatory devices. Taking the NOT logic gate shown in Fig. 5A as a reference, we performed the design of a new 5′ UTR for cis-repression and that was able to respond to the same riboregulator, in this case working as an activator. The optimization problem read(5)where the difference with Eqs. (2) relies on the imposition that the sRNA sequence is constant. Likewise, the same sRNA will have the ability to both repress and activate protein expression (coupled YES/NOT logic gate). Exploiting further this modularity, we carried out the design of an OR logic gate using the 5′ UTR sequence just designed. We now enforced the design of a new sRNA that had also the ability of releasing the RBS, maintaining constant the 5′ UTR sequence. The optimization problem had then only one instance, given by(6)Thus, the resulting system will integrate two sRNAs capable of activating the release of the RBS contained in a single 5′ UTR. Subsequently, we verified there was no interference between the two sRNAs, although this could have also been incorporated into the design process. Fig. 6 shows the integrative circuit (multi-input, multi-output) that we finally obtained with this strategy based on serial design of constrained YES gates.
Motivated by the previous results, we carried out the design of cooperative riboregulations. The regulatory function of multiple-sRNA complexes has not been reported in prokaryotes (all natural systems for riboregulation involve two RNA species, at most interacting with proteins such as RNA chaperones or endoribonucleases [28]), which further encourages the exploration by means of computational methods. To illustrate the power of our approach, we focused on the design of synergistic activation (AND logic function), where two trans-regulating sRNAs first interact among them to form a complex that will then activate translation (Fig. 1E). To solve the optimization problem, we constructed the following objective functions(7)As in the previous cases, these functions are associated to each entry of the truth Table, and hence the solution of this problem will yield AND logic gates. In Fig. 7, we show two different designs of this logic, combinatorial device. By themselves, the trans-regulating sRNAs cannot release the RBS. However, the dimer they form has a distinct structure that allows interplaying with the 5′ UTR.
In conclusion, we have followed a bottom-up approach to design RNA devices with YES, NOT, AND, and OR logic functions, based on first physical principles. These logic gates implement multi-state sRNA devices for which there was no design method before, and that can be interconnected to create more complex logic programs. Although we could solve intermolecular inverse folding problems [29], it was not possible the systematic design of multiple RNA species implementing arbitrary logic gates. For their design, each entry of the truth Table imposes a structural specification. Here, we accounted for the free energies of all possible reactions (thermodynamic potential) to solve this multi-objective inverse problem by optimization. Because our methodology does not require natural sequences (with the exception of key motifs such as the Shine-Dalgarno sequence), we have solved the full design problem of regulatory RNA for implementing logic programs in living cells.
Our approach has, however, some limitations, which prospect further research in the field. One of them is the use of the secondary structure to model riboregulation. This type of regulation could involve pseudoknot interactions and even non-canonical base pairing, for which three-dimensional models could better capture the interaction features [30]. In addition, our model does not account for RNA chaperons (e.g., Hfq) [31], nor co-factors such as Mg2+ or Zn2+, nor kinetic binding effects, which might have an impact on the designs. Another restraint of the current method is the enforcement of a given structure for all single species in the circuit (although not for the complex ones), because this constrains the sequence space of possible solutions [11]. By leaving unconstrained those structures, we could perform additions and/or deletions (not only replacements) of nucleotides during the optimization, and we would need to include into the function ΔGstr a new term for the stability (e.g., based on free energy). Finally, the convergence of the algorithm is highly reduced when evolving systems with multiple species, making necessary to reduce the sequence space by reusing functional modules to obtain more sophisticated systems.
Despite these limitations, we have demonstrated the power of computational design (through heuristic optimization) to overcome the complexity in obtaining fully synthetic riboregulation, exploring the vast combinatorial space of sequences. The proposed objective function was shown predictive enough to allow the designability of multi-state RNA devices, as ΔGkin explained differences in experimental repression fold for a set of mutational variants of the IS10 antisense RNA system (Fig. 2) [4]. Moreover, we recently validated experimentally some designs of YES logic gates in bacteria, encouraging further work [11]. Even though, the design problem does not require a perfect prediction, and similar or even lower correlations can be sufficient to tackle this problem, such as in the case of automated RBS design [22]. Of course, more sophisticated objective functions will be developed in the coming years to improve the design of functional RNAs.
The combination of ΔGkin and ΔGstr, for every possible conformational state (intra- or intermolecular) of a given genotype, results in an effective free energy that defines a fitness landscape. In case of riboregulation, the total search space can be about 1040 sequences [11], and typical optimizations that lead to sufficiently good solutions consist of 106–107 iterations. Indeed, the generalized problem of finding the nucleotide sequences of multi-species ensembles that will fold into specified conformations has an exponentially large number of solutions. It remains however a question how to distinguish several optimized sequences (assuming equal energetic features). For instance, differences in intracellular stability of the species will affect the ratio sRNA/mRNA, and then be key for the regulatory activity. Additionally, the kinetics of RNA folding, binding, and turnover will have significant impact on the performance of designed RNA circuits [3], [10]. All these criteria, either from first principles or from experimental feedback, will be exploited to enhance the design methodology.
Our present methodology is general and could be applied to obtain designs based on further mechanisms. In addition, instead of attempting full designs, it permits reusing complete known sequences (natural or synthetic) to constrain the design of new logic systems. This capacity enables the creation of a large variety of combinatorial sRNA systems, increasing sophistication at a reduced computational cost. Moreover, our approach can be used to analyze potential RNA sequences for a given functional circuit as a reverse engineering tool. The designed sRNA-based logic gates can be combined with transcription regulation to generate more complex functions [32], and also be integrated into libraries of models for the computational design of more complex networks involving transcription and post-transcription regulation [33]. Yet, our full design automation approach together with high-throughput screening techniques will propel the construction of modular and orthogonal devices for synthetic biology [34].
We considered riboregulation (RNA-RNA interaction) in terms of thermodynamics [29], [35], [36], assuming that the system reaches an equilibrium state. We first applied an inverse folding strategy over the structures of all individual species. Then, neutral mutations in structure were evaluated with an objective function intended to optimize the intermolecular folding states. To obtain an intermolecular folding satisfying the release or blockage of the RBS, in principle, we needed to maximize the partition function (Z) of the whole system. Using the reaction coordinate of the system (r), defined as the number of intermolecular Watson-Crick interactions (i.e., r = 0 represents individual folding) [11], Z can be written as(8)where G(r) is the effective free energy of the state with reaction coordinate r (where G(0) represents the free energy of the no-interaction state, with G = 0 for the unfolded state), R the gas constant, and T the temperature. Here, we are interested in G(r) at the reaction coordinates for the transition, G(rtrans), and final intermolecular (hybridization) states, G(rhyb), to define our functions ΔG, the free energy of formation, and ΔG‡, the free energy of activation, by(9)
To compute the free energy and secondary structure of all species (single and complexes) of a system, we used the ViennaRNA [37] and MultiRNAFold [38] (when having more than two RNA species) software. We only considered the minimum free energy state discarding the suboptimal ones. Here, we did not consider pseudoknots. Afterwards, the designed sequences were analyzed with the Nupack software [29], which is able to compute ensemble properties such as Z. In this work, we used the Mfold 3.0 RNA energy parameters [39], and always considered T = 37°C (which gives RT = 0.61 Kcal/mol).
In an RNA-RNA interaction between species A and B, an intermediate complex at the transition state ([A:B]‡) is formed mediated by the seed. Then, a fast reaction inducing a conformational change occurs. Denoting kon and koff the forward and reverse constants, respectively, to form [A:B]‡, and khyb the hybridization constant to form the final complex (A:B), the mass action kinetic model reads(10)where δ1 and δ2 are the degradation constants. Assuming that koff + khyb is much greater than δ1 (sRNA degradation takes several minutes [13]), we can obtain in steady state [A:B]‡ = AB/KM, where KM = (koff + khyb)/kon is the Michaelis constant. Hence, A:B (and also the translation rate) will be in steady state proportional to khyb/KM, assuming there is no saturation.
The constant kon can be obtained by fitting in vitro DNA hybridization data, where only the length of the seed (α), irrespective to the sequence, determines the kinetic constant following a Boltzmann factor [25]. Moreover, we can say that the constant khyb is determined by ΔG (the free energy of formation between A + B and A:B) also with a Boltzmann factor. This allows us to write(11)Therefore, the resulting model reads(12)where Gp is a fitted parameter to account for the average energetic contribution of one nucleotide. Gp = −1.28 Kcal/mol [25]. Finally, we proposed ΔG + αGp as the objective function to optimize RNA-RNA interactions. This formulation is in part equivalent to maximize Z, because from the Arrhenius equation [23] ΔG‡ and α should have a linear relationship.
Our evolutionary algorithm consists in a Monte Carlo Simulated Annealing [40], which can be parallelized to evolve a population of sequences. Our approach consists in optimizing an objective function accounting for the interaction and structure of the RNAs that lead to the target behavior.
The design specifications comprise the secondary structures of all single RNAs, critical subsequences of nucleotides (e.g., RBS), the reaction free energies, and the structure of the output complex. The algorithm starts from pure random sequences satisfying the structural and subsequence constraints, although it can also be specified an initial sequence. If the subsequence constraints do not allow satisfying the structures, the algorithm stops. Eventually, we can introduce a relaxation in the structural constraints (through an harmonic constraint) allowing having species with dissimilar structures to their targets. Subsequently, an iterative process of mutation and selection is implemented (see scheme of the algorithm in Fig. S3). The mutation operator consists in either random or directed nucleotide replacements. We do not consider additions or deletions, so the length of the RNAs is maintained constant. To speed up the convergence, we generated a mutation operator that only created useful mutations, e.g., mutations that are always guaranteed to contribute for an interaction among RNA species. We do this by taking a word (i.e., set of consecutive nucleotides) from one sequence, making its reverse complementary, and randomly inserting it into another sequence. Initially, the length of this word is three, and it is reduced to one (i.e., single point mutation) during the optimization process. Those mutations speed up the in silico evolution. If a nucleotide that has to be mutated belongs to a stem, its pair in the stem is also mutated with the corresponding nucleotide with the aim of preventing the disruption of the secondary structure and improving the convergence. We avoid sequences having consecutive repeats of four or more identical nucleotides.
The objective function is a weighted sum of two terms to be minimized. The first term (ΔGkin) accounts for the reaction kinetics of the system. For that, we compute the ΔG and α of all possible reactions, having between species A and B(13)Notice that ΔGkin is a negative-valued variable. We will minimize or maximize ΔGkin if the reaction must occur or not (in order to obtain the specified behavior). Maximizing ΔGkin is equivalent to minimize −ΔGkin. During the optimization we exclude sequences forming homodimers. In addition, we considered ΔGsat = −15 Kcal/mol and αsat = 6 as arbitrary saturation levels (i.e., levels from which there is no need for further minimization). These values can be enlarged to get designs with lower ΔGkin, although at a cost of altering the convergence. The second term (ΔGstr) accounts for the structural change of the output RNA. For that, we use a Hamming distance (d) between the current and target structures, being(14)This indicates that species A (which can be single or complex) is evolved to display the target structure, or substructure, Str (e.g., RBS paired, then repressing protein translation). Gp is used to rescale the distance in terms of free energy. We note that ΔGstr is a positive-valued variable, which we will minimize.
100 ng of plasmid pRAJ11 coding for the riboregulatory device RAJ11 were subjected to 30 cycles of PCR amplification with divergent primers I (5′-CCGCGAAGACCGGCACGGNNNGGTTGATTGTGTGAGTCTGTC-3′, N is A, C, G or T; BpiI recognition and cleavage sites underlined) and II (5′-GGCGGAAGACGCGTGCTCAGTATCTCTATCACTG-3′, BpiI recognition and cleavage sites underlined) in a volume of 20 µL with 0.4 U of the high fidelity Phusion DNA polymerase (Thermo Fisher Scientific) in the presence of HF buffer (Thermo Fisher Scientific), 3% dimethyl sulfoxide, 0.2 mM each dNTP and 0.5 µM each primer. Reactions consisted of an initial denaturation of 30 s at 98°C followed by 30 cycles of 10 s at 98°C, 30 s at 55°C and 1∶15 min at 72°C, with a final incubation of 10 min at 72°C. After PCR, 10 U of DpnI (Thermo Fisher Scientific) were added to each sample to digest the template plasmid and incubated for 1 h at 37°C. Reaction products were electrophoresed in a 1% agarose gel in TAE buffer (40 mM Tris, 20 mM sodium acetate, 1 mM EDTA, pH 7.2) and the gel stained with ethidium bromide. The 4460-bp long DNA product corresponding to the full-length plasmid was eluted from the gel, digested with BpiI for 1 h at 37°C (Thermo Fisher Scientific) and finally subjected to self-circularization with 5 U of T4 DNA ligase (Thermo Fisher Scientific) for 1 h at 22°C. Reaction products were purified by chromatography with silica gel spin columns (DNA Clean and Concentrator, Zymo Research) and electroporated in E. coli DH5α. Recombinant bacteria were selected in plates with 50 µg/mL ampicillin. Plasmids were purified from liquid cultures of selected clones (Wizard Plus SV Miniprep DNA Purification System, Promega) and analyzed by electrophoresis in 1% agarose gels in TAE buffer, followed by ethidium bromide staining. Forty-five plasmids whose electrophoretic mobility matched that of parental pRAJ11 were subjected to sequence analysis with primer III (5′-GAATTCGCGGCCGCTTCTAGAGC-3′) to find out the particular sequence in the randomized trinucleotide position introduced by primer I. Eleven mutant clones (see Table S3) were selected for further analysis, as well as the wild-type sRNA RAJ11 and the null system RAJ11m (Fig. S5).
Cultures (2 mL) inoculated from single colonies (three biological replicates) were grown overnight in LB medium at 37°C and 220 rpm. Cultures were then diluted 1∶100 (in 2 mL of LB), and were grown for 3 h in the same conditions (to reach an OD600 about 0.5). Ampicillin was used as antibiotic at 50 µg/mL. Then, 500 µL of each culture were centrifuged for 2 min at 13,000 rpm, and resuspended in the same volume of water. Subsequently, we loaded the multiwell plate with 200 µL for each sample, which was assayed in a Victor X5 (Perkin Elmer) to measure absorbance (600 nm absorbance filter) and fluorescence (485/14 nm excitation filter, 535/25 nm emission filter, for GFP). Background values of absorbance and fluorescence, which corresponded to water, were subtracted to correct the signals, and the normalized fluorescence was calculated as the ratio of fluorescence and absorbance (Fig. S4). Hence, we calculated the fold changes of activation (relative changes in GFP protein expression in absence or presence of sRNA).
|
10.1371/journal.pgen.1003131 | Comparative Genomic Analysis of the Endosymbionts of Herbivorous Insects Reveals Eco-Environmental Adaptations: Biotechnology Applications | Metagenome analysis of the gut symbionts of three different insects was conducted as a means of comparing taxonomic and metabolic diversity of gut microbiomes to diet and life history of the insect hosts. A second goal was the discovery of novel biocatalysts for biorefinery applications. Grasshopper and cutworm gut symbionts were sequenced and compared with the previously identified metagenome of termite gut microbiota. These insect hosts represent three different insect orders and specialize on different food types. The comparative analysis revealed dramatic differences among the three insect species in the abundance and taxonomic composition of the symbiont populations present in the gut. The composition and abundance of symbionts was correlated with their previously identified capacity to degrade and utilize the different types of food consumed by their hosts. The metabolic reconstruction revealed that the gut metabolome of cutworms and grasshoppers was more enriched for genes involved in carbohydrate metabolism and transport than wood-feeding termite, whereas the termite gut metabolome was enriched for glycosyl hydrolase (GH) enzymes relevant to lignocellulosic biomass degradation. Moreover, termite gut metabolome was more enriched with nitrogen fixation genes than those of grasshopper and cutworm gut, presumably due to the termite's adaptation to the high fiber and less nutritious food types. In order to evaluate and exploit the insect symbionts for biotechnology applications, we cloned and further characterized four biomass-degrading enzymes including one endoglucanase and one xylanase from both the grasshopper and cutworm gut symbionts. The results indicated that the grasshopper symbiont enzymes were generally more efficient in biomass degradation than the homologous enzymes from cutworm symbionts. Together, these results demonstrated a correlation between the composition and putative metabolic functionality of the gut microbiome and host diet, and suggested that this relationship could be exploited for the discovery of symbionts and biocatalysts useful for biorefinery applications.
| The symbiotic gut microbiome of herbivorous insects is vital for their ability to utilize and specialize on plants with very different nutrient qualities. Moreover, the gut microbiome is a significant resource for the discovery of biocatalysts and microbes with applications to various biotechnologies. We compared the gut symbionts from three different insect species to examine whether there was a relationship between the diversity and metabolic capability of the symbionts and the diet of their hosts, with the goal of using such a relationship for the discovery of biocatalysts for biofuel applications. The study revealed that the metabolic capabilities of the insect gut symbionts correlated with insect adaptation to different food types and life histories at the levels of species, metabolic pathway, and individual gene. Moreover, we showed that the grasshopper cellulase and xylanase enzymes generally exhibited higher activities than those of cutworm, demonstrating differences in capabilities even at the protein level. Together, our findings confirmed our previous research and suggested that the grasshopper might be a good target for biocatalyst discovery due to their high gut cellulytic enzyme activities.
| Insects represent one of the most diverse groups of organisms on the planet that can adapt to the extremely diverse eco-environments. In particular, herbivorous insects can exploit a wide range of the plant species as food sources [1]. Insect gut symbionts play an essential role in the insect adaptation to various food types and they have been shown to be important for lignocellulosic biomass degradation, nutrient production, compound detoxification, and environmental adaptation [2]–[7]. Disrupting insect gut symbionts can significantly reduce the fitness of insects and can even cause serious diseases such as CCD (Colony Collapse Disease) [8]. Moreover, insect gut symbionts also were shown to be maternally inheritable from generation to generation, which suggests the symbiotic microbiota is a dynamic component of the competitive evolution between plants and herbivorous insects as well as a driving force for insect speciation [9], [10]. For these reasons, insect gut symbionts have been the subject of extensive studies in recent years [10]. Previous studies highlighted several important features of some insect gut symbionts including their reduced genome size, convergent evolution, co-speciation, and complementary function with the host genome [11]–[15]. Recent studies also expanded our understanding of the roles of insect gut symbionts in non-conventional functions like nitrogen recycling, reproductive manipulation, pigment production and many other aspects related to insect fitness [16], [17].
Despite the progress toward understanding insect-symbiont relationships, there is still much to be learned especially with regard to facultative symbionts. Moreover, limited research has focused on comparing the gut symboints from insect species that specialize on different food sources. For this reason, we systemically compared the gut enzyme activities and microbial diversity in several insect species relevant to biotechnology applications [2], [3], [18]. Previous studies comparing gut symbionts from woodbore (Cerambycidae sp., (Coleoptera)), silkworm (Bombyx mori (Lepidoptera: Bombycidae)), and grasshopper (Acrida cinerea (Orthoptera: Acrididae)) suggested that the insect gut cellulytic enzyme activities were generally correlated with the lignocellulosic biomass composition in the food consumed [2]. Furthermore, the comparison of the microbial community structure of gut symbionts from woodbore, silkworm, grasshopper, and cutworm (Agrotis sp. (Lepidoptera:Noctuidae)) using DGGE (Denaturing Gradient Gel Electrophoresis) revealed significant differences in symbiotic community correlating with food adaptation [3]. Despite the progress, an in-depth understanding of the eco-evolutionary adaptation to food types requires metabolic and phylogenic analysis that cannot be offered by traditional approaches like DGGE [18]. Most of the previous comparative studies of symbionts from different insect species were either carried out with DGGE or focused on one or few symbiotic species [19], [20]. Compared to those conventional techniques, new platforms like metagenomics could help define the function of symbionts in the food adaptation of insects and promote discovery of biocatalysts for biotechnology applications [18].
From the deep sea to the human intestine system, metagenome analysis has emerged as a major approach to study the composition, function, and evolution of various microbiota [21]. Metagenome analysis and metabolic reconstruction of the termite gut symbiotic microbiota revealed potential functionality in these microbiomes that might be required for biomass degradation, nutrient synthesis and other functions essential to the insect [22]–[24]. Moreover, those studies also highlighted the potential for biotechnology application of insect gut symbionts, since many potential glycosyl hydrolases (GH) family enzymes have been identified from the termite gut [24]. Further studies revealed the potential complementary function between the host and symobionts enzymes for highly efficient biomass degradation [23]. Despite the progress, previous research mainly focused on the metagenome sequencing of symbionts in single insect species or the same symbioint in different insect species [17], [25]–[27]. Few studies have systematically compared the metagenomes of symbiotic microbiota from insect species with distinctly different diets, environmental adaptations, or life histories. This type of comparative metagenomics approach has the potential to substantially improve our understanding of the adaptive significance of insect gut symbionts for insect diet specialization as well as facilitates the discovery of novel biocatalysts for biorefinery applications.
In this study, we selected three insect species that are from different insect orders and have different diets and life histories characteristics: grasshopper (Acrida cinerea (Orthoptera), cutworm (Agrotis ipsilon) (Lepidoptera) and termite, Nasutitermes sp. (Isoptera: Termitidae). The grasshopper is a polyphagous insect specializing on different plant leaves, mainly from the monocot grass species. Previous studies revealed that the grasshopper diet contains about 37.2% of forbs, 58% of grasses and sedges and 4.8% of others [28]. The cutworm is also a polyphagous, generalist that can adapt to a broad range of food sources including cabbage, asparagus, bean, and other crucifers [29]. In contrast, the termite is monophagous insect that specializes on lignocellulosic biomass as a food source. The three insects also differ in life cycle. The cutworm is a holometabolous insect that undergoes complete metamorphosis with a pupal stage [30], whereas the grasshopper and termite are hemimetabolous, having incomplete metamorphosis and juveniles with morphologies similar to adults [31].
Metagenome data from the gut symbiotic microbiota of grasshopper and cutworm were generated using Illumina Genome Analyzer, and these metagenome data were compared with the updated sequencing data from gut symbionts of the wood-feeding higher termite [24]. As one of the first comprehensive comparisons of insect gut symbiotic metagenome, the goal was to examine the relationships between the taxonomic and potential metabolic diversity of the insect gut microbiomes and the diets and life histories of their insect hosts at the community, metabolic pathway, and molecular levels. The analysis indicated that the composition of gut symbionts was correlated with their function in biomass degradation and nutrient biosynthesis. The metabolic reconstruction revealed the presence of specific pathways relevant to the utilization and transport of diverse carbohydrate sources in cutworm and grasshopper. The diversity, phylogenetic, metabolic, and functional analyses all supported the hypothesis that insects and their gut symbionts co-evolved with the food preferences of the insect toward optimal capacities in biomass degradation, macromolecule intake and utilization, complementary nutrient synthesis, and other aspects related to insect life style. In addition, we cloned 24 biomass degrading enzymes based on the predicted gene models and characterized four of them. Enzyme assays revealed that grasshopper cellulytic enzymes were generally more active than the cutworm cellulytic enzymes, which confirmed the presence of functional diversity at the protein. The enzyme characterization indicated that insect guts were useful resources for discovering novel biocatalysts for biorefinery applications.
The metagenome sequencing results were summarized in Table 1. The sequence assembly rendered more than 20,000 of predicted gene models for the gut symbionts from grasshopper and cutworm, respectively. In order to analyze the composition-function relationship, we compared the grasshopper and cutworm gut microbiota with the updated termite gut microbiota sequences (JGI IMG Database GOLD ID: GM00013 and Sample ID: GS0000048), with respect to the phylogenetic diversity, microbial abundance, putative gene function, and metabolic capacity. As described above, the three host species are from distinct insect orders and have different diet specializations and life histories.
Relative abundance of symbiotic microbial species in each insect gut was estimated based on the species distribution of the gene-coding sequences as annotated by the BLAST search. The cluster analysis of bacterial species distribution for the gut symbionts was shown in Figure 1. It should be pointed out that Figure 1 only represented a rough estimation of the microbial species distribution because of the genome size variations in different symbionts, which complicated the data interpretation. Nevertheless, the comparison of the relative abundance of the bacteria phyla in the microbiota from the three different insect species revealed that the microbiota composition was rather different from each other and these differences might be relevant to the functions they provided for their insect hosts. The dominant groups differed among the three insect species. For the cutworm, the phylum Bacilli was the dominant group (24.14%), followed by Clostridia (4%), Erysipelotrichi (3.64%) and γ-proteobacteria (1.43%) (Figure 1). For the grasshopper, the most common bacterial genes were from γ-proteobacteria (25.16%), followed by Erysipelotrichi (3.51%), Clostridia (1.27%), and Bacilli, (0.84%), respectively (Figure 1). For both species, the most abundant groups comprised about 25% of the diversity, whereas the second most abundant groups comprised less than 5%.
Even though the insects differed in microbial composition, there were some similarities that likely were related to function. Both Clostridia and Bacilli species have been shown to be the major groups of microbes responsible for biogas production and biomass conversion in microbial communities [32]. Many Clostridia species such as C. thermocellum and C. ljungdahlii are anaerobic Firmicutes known to have a robust capacity to use cellulose, hemicellulose, and other carbohydrate [33]–[35]. The presence of a large proportion of Clostridia was likely to be important for lignocellulosic biomass degradation [34], [36]. However, the predominance of the γ-proteobacteria in grasshopper was unexpected, because γ-proteobacteria has not been shown previously to be involved in biomass utilization. However, recent work revealed that γ-proteobacteria might be important nutrient providers for host insects. For example, γ-proteobacteria as facultative or obligate endosymbionts were shown to play essential roles for insects like tsetse fly in the utilization of low nutrient food sources [37]. Similarly, the predominance of γ-proteobacteria in grasshoppers might be important for the utilization of the grasses, which characteristically have high fiber content.
Compared to the grasshopper and cutworm microbiomes, the microbial composition of the termite microbiome reflected its unique adaptation to utilization of woody species, where both the Clostridia and the Spirochaetes species were predominant (Figure 1) [24]. Additionally, the termite microbiome was composed of several major groups with more than 5% abundance. Morphologically diverse spirochaetes were consistently present in the hindgut of all termites [38], and was found as ectosymbionts attached to the surface of cellulose-digesting protists [39]. Overall, the microbial populations of the cutworm, grasshopper and wood-feeding termite gut systems appeared to consist of taxa with known capacities for degrading and utilizing the different types of foods on which their insect hosts specialize.
In addition to gene-coding sequence-based analyses, we also implemented two types of phylogenetic analyses. First, two partial 16S rRNA clone libraries were established from the PCR amplified 16S rRNA sequences using 515F/1492R primers. Sanger sequencing was used to sequence individual 16S rRNA clones as summarized in Table S1. The phylogenetic analysis was presented in Figure 2. The second phylogenetic analysis was based on the annotation of the contigs derived from the metagenome sequence assembly. The assembled contigs were first aligned to the 16S rRNA genes from the recent release of RDP database using blastn. The analysis resulted in 188 and 102 contigs assigned to be 16S rRNA for cutworm and grasshopper, respectively (Table S1). The most similar partial or complete 16S rRNA sequences from the database were used for the multiple sequence alignment and phylogenetic analysis using Maximum likelihood method (RAxML). The analysis results were presented in Figure S1. The results from the two types of analysis generally were consistent; although the phlygenetic analysis based on the annotated contigs (Figure S1) provided a deeper coverage of microbial species and a better representation of uncultured species.
The phylogenetic analyses (Table S1, Figure 2, Figure S1) revealed three features. First, proteobacteria represented the most diverse group of the microbes in the microbiomes of both grasshopper and cutworm. Among the proteobacteria, γ-proteobacteria was the predominant taxa and the 16S rRNA sequences from cutworm and grasshopper formed two distinct clades, indicating the relatively independent evolution of the gut microbiome in the two species. The 16S rRNA-based phylogenetic analysis correlated well with the microbial abundance analysis using gene models (Figure 1). The studies confirmed the differences in abundance, phylogeny, and evolution of gut symbionts between cutworm and grasshopper. A second feature of the analyses was that the cutworm had more species of gut symbionts than grasshopper (188 vs. 102, Figure S1). We speculated that the greater diversity of symbionts in the cutworm gut as compared to that of the grasshopper might be relevant to its being both more of a dietary generalist. A third feature was the discovery of large number of uncultured species or unknown species. Uncultured species referred to the species that cannot be cultured in standard medium, whereas unknown species referred to those lacking taxonic information. Due to the deeper coverage of metagenomic sequencing compared to the PCR cloning library, Figure S1 showed almost 60% sequences were from uncultured or unknown species. The results highlighted our limited knowledge of the diversity of insect gut symbionts. It was proposed that the existence of many unculturable species might be related to the significant reduced genome and limited metabolic capacity of some symbiotic microbes [40]–[43]. The phenomena indicated that the metabolic capacity of insect gut microbiota should be considered as a whole instead of based on individual species.
Another observation was that 14 and 10 16S rRNA sequences were assigned to Acetobacter pasteurianus (AP011163) for cutworm and grasshopper, respectively (Figure S1). Acetobacter strains belong to acetic acid bacteria (AAB), which are often found in various categories of fruits, flowers, and fermented foods [44] and some insect guts [45]. Acetobacter might have originally been acquired from the food sources of cutworm and grasshopper and subsequently become a more permanent symbiont for the two species or might occur as a transient resident. Acebacter can produce alcohol dehydrogenase (ADH), which could potentially contribute to lignin oxidation for lignin degradation/modification in termite guts [46], [47]. Overall, the phylogenetic analysis indicated correlations between microbial composition and function and insect diet preference.
Metagenome sequencing provided more detailed functional comparisons of different gut symbionts using pathway analysis based on COGs (Clusters of Orthologous Groups) and KEGG (Kyoto Encyclopedia of Genes and Genomes) [48], [49]. KEGG maps the genes within the biological pathways to derive potential functions [50], whereas COG analysis uses evolutionary relationships to group functionally relevant genes [51]. The annotation of the cutworm and grasshopper gut microbiomes yielded 11,317 and 8954 hits for the COG database as well as 900 and 1105 hits for the KEGG pathways, respectively.
D-ranks analysis was used to evaluate the relative enrichment of COG and KEGG gene categories in the cutworm and grasshopper gut symbiotic metagenomes compared to the termite metagenome. The enrichment or under-representation of COG categories were as shown in Figure 3. Both cutworm and grasshopper gut symbionts were enriched in several metabolic pathways compared to termite gut symbionts. Cutworm gut symbionts were enriched with genes for carbohydrate transport and metabolism, and defense mechanisms (P<0.05) relative to grasshopper symbionts. The diversity in carbohydrate metabolism genes correlated well with the taxonomic diversity of the gut microbiomes (Figure S1) and were consistent with the hypothesis that the greater diversity in species composition and carbohydrate metabolism observed in the cutworm may be related to the broader diet preference and more complicated life histories of the cutworm compared to those of the grasshopper.
The ontology analysis based on KEGG revealed similar patterns as shown in Table S2, where flagella assembly in cell motility and type III secretion system (P<0.05) are more enriched in termite gut symbionts than those of cutworm and grasshopper, although it is unclear why this would be so. Overall, the metagenomic composition of genes in all categories reflected their potential function in adaptation to insect diet and life history. A more detailed functional relevance can be derived from examination of specific pathways.
Metabolic reconstruction provided comparison of potential biocatalyst functionality in four general COG categories and thus a means of relating the metabolic diversity and capability of the microbiome to the insect diet and life style.
The ultimate goal of this research was to discover novel biocatalysts for biorefinery applications. We therefore cloned and characterized several enzymes for functional validation. A total of 24 ORFs of predicated plant polysaccharides degradation enzymes were PCR amplified using primers based on the assembled sequences (Figure S2). A total of 22 out of 24 ORFs amplified and the sequences of all of the amplicons were consistent with the assembled sequences (Figure S2). The results highlighted the reliability of the Illumina metagenomic sequencing and assembly to identify degredation enzymes. Our research represents one of the few metagenome sequencing efforts to rely mainly on the Illumina Genome Analyzer [69].
We further characterized an endoglucanase (CW-EG1 and GH-EG1) and a xylanase (CW-Xyn1 and GH-Xyn1) from both the grasshopper and cutworm guts, respectively. The selected enzymes were expressed and purified by a His-trap nickel column, as indicated by SDS-PAGE (Figure S3). The enzyme performance under different temperature and pH conditions was as shown in Figure S4. All four of the enzymes exhibited activity, and the activities were significantly influenced by temperature and pH. Most enzymes had temperature optima at 60∼70°C and pH optima at 7.0–9.0 (Figure S4). This pH range correlates with the fact that many insect gut systems have a slightly basic environment [70] Considering that many traditional filamentous fungi enzymes had optimal activity in the weakly acidic pH range, the insect gut enzymes provided complementary capacity for biomass degradation.
We further compared the specific activity of the same category of enzymes from cutworm and grasshopper gut microbiome. Interestingly, for both cellulase and xylanase, the grasshopper gut enzymes were significantly higher than those of cutworm (P<0.05, Figure 5). The result correlated with our previous analyses of gut content activities, even though the differences could also result from the choice of enzymes and other factors [2]. The adaptation to relatively higher temperature made the enzymes good candidates for some biomass conversion applications.
Together with many recent studies, our research indicated that insect gut symbionts are substantial resources for enzyme discovery for biorefinery applications. The relationship between the diversity and potential functional capabilities of the gut microbiomes and insect food preference is particularly relevant improvements in biomass degradation, and thus should be explored for biotechnology applications [71]–[75]. Due to the technical limitations, we particularly focused on the bacterial symbionts in this study. Nevertheless, the fungal and protozoal symbionts in insect guts were also widely studied for their biomass degradation capacity. These eukaryote symbionts should be investigated for their roles in biomass deconstruction, food and life history adaptation in the follow-up studies.
Metagenome analysis requires comprehensive coverage of most multiple species in the sample [76]. To obtain sufficient high-quality DNA for sequencing with Illumina Genome Analyzer, approximately 2000 third to fifth instar grasshoppers and 50 fourth to fifth instar cutworms were dissected to extract genomic DNA from gut symbionts. A recently developed indirect DNA extraction method was modified for the insect gut metgenomic DNA extraction [77]. The extracted metagenomic DNA were quantified by a Nano Drop ND-1000 spectrophotometer and characterized by electrophoresis. Moreover, the quality of the DNA was verified by PCR amplification of conserved 16S rRNA for bacteria and conserved 18S rRNA for insect host contamination [29]. The results confirmed that the metagenomic DNA is free from host DNA contaminations, because the 18S rRNA did not amplified.
Metagenome sequencing of cutworm and grasshopper gut symbiotic microbioata was carried out using Illumina Genome Analyzer II (Illumina, Inc. CA, USA) with paired-end 76 base sequencing. Library construction was carried out following the manufacture's recommendation using Illumina Paired-End Sequencing Kit (Cat. No. PE-102-1001). Briefly, 2 to 5 µg metagenomic DNA was sheared by nebulization to generate DNA fragments and the ends were repaired with Klenow, followed by several steps to add the adapters. Adapter-ligated DNA fragments of length 300–350 bp were isolated from a 2% agarose gel using QIAquick Gel Extraction Kit. The fragments were then amplified by 11 cycles of PCR reaction to generate the DNA library at a concentration of 20–35 ng/µl. The median size of the library was evaluated using 2% agarose gel. The PHIX Control V2 Library was prepared by Illumina (Cat. No CT-901-2001) and used for sequencing. Approximately 5 pmol DNA libraries were subjected to cluster generation and sequenced by DNA core of Institute of Plant Genomics and Biotechnology. The images were processed using version 0.3 of the GAPipeline software supplied by Illumina.
After base-calling with GAPipeline software, the remaining 44,155,246 (cutworm) and 58,033,340 (grasshopper) reads (each is about 76 bases) were trimmed and assembled using Velvet version 0.7.55 (http://www.ebi.ac.uk/~zerbino/velvet/, European Bioinformatics Institute, EMBL-EBI). The resulted assembly consisted of 64,065 and 78,991 contigs for cutworm and grasshopper, respectively.
The draft assembled contigs (≥100 bp) were loaded into IMG/M (http://www.jgi.doe.gov/m) [78]. Before further analysis, the IMG/M system first carried out a gene model validation process, including editing overlapping CDSs, correcting start codons, and identifying missed genes and pseudogenes [78]. The predicted coding sequences (CDSs) and some functional RNAs were recorded with start/end coordinates in the contigs. The predicted genes were assigned to COGs (clusters of orthologous groups) based on RPS-BLAST (reverse position specific BLAST) and NCBI's Conserved Domain Database (CDD), using an e-value threshold of 10−2 without low-complexity masking [79]. Genes were also probed against Pfam database using HMMER search (http://hmmer.janelia.org/) [80], [81]. Protein-coding sequences were further annotated for molecular function and pathways using KEGG pathways. In addition, the metagenome sequences and gene models were binned to rank domain, phylum, and class using PhyloPythia [82].
The phylogenetic analysis of 16S rRNA was carried out with two types of analyses. First, two clone libraries were prepared using PCR products amplified from cutworm and grasshopper gut metagenome DNA with one pair of primers broadly targeting the V3–V9 region of 16S rRNA. The primer sequences were 515F (5′-GTGCCAGCAGCCGCGGTAATACCTTGTTACGACTT-3′) and 1492R (5′-GGTTACCTTGTTACGACTT-3′) [83]. 87 and 97 near complete 16S rRNA V3–V9 region sequences were obtained for cutworm and grasshopper gut microbiome, respectively. The 16S rRNAs was then used for phylogenetic analysis.
In addition to sequencing of the V3–V9 region, we also sought to reach a deep coverage of symbiotic species by analyzing the assembled metagenome sequences. 16S rRNA sequences were identified using BLASTN (E<1×10−5 and a sequence length hit >50 nt) search against the SSU rRNA genes from release 16.3.3 of the RDP database (http://rdp.cme.msu.edu/) [84], and the European Ribosomal RNA database (http://www.psb.ugent.be/rRNA/index.html). Due to the high similarity, it is usually difficult to isolate the 16S rRNA genes from de novo assembly of metagenome data. A total of 96 and 53 partial and near complete 16S sequences were extracted from 188 and 102 assembled contigs for cutworm and grasshopper gut microbiomes, respectively. The sequences were then aligned with the NAST aligner [85], and imported into an ARB database (http://greengenes.lbl.gov) [86]. The nearest aligned full length sequences were used for classification and phylogenetic tree construction using RAxML [87].
Phylogenetic analysis was carried out using the Minimum Evolution method with the sum of branch length = 5.0 [88]. The evolutionary distances were computed using the Maximum Composite Likelihood method with 1000 replicates of bootstrap tests [89].
In order to compare the metabolic pathways for different microbiota, the coding sequences were analyzed with KEGG and COG (Clusters of orthologous groups). Both grasshopper and cutworm symbiotic metagenome and updated termite metageome data (JGI IMG Database GOLD ID: GM00013 and Sample ID: GS0000048) [24] were compared. For KEGG analysis, all coding sequences were converted into KEGG orthologous (KO) groups, and the KEGG pathway annotation was extracted based on the latest release of KEGG version (Release 55.1, September 1, 2010). The COG assignment was based on RPS-BLAST and NCBI's Conserved Domain Database (CDD). Only 4.95%, 3.48%, and 6.41% of predicted genes were assigned to KEGG pathway for grasshopper, cutworm, and termite gut microbiome, respectively. 39.4%, 44.41%, and 53.56% of coding sequences were assigned to COG terms for grasshopper, cutworm and termite gut microbiome, respectively.
In order to further define the enrichment or under-representation of a KEGG pathway or a COG term in a certain microbiome, two metrics were used in this study. For the comparison of a protein family between a query metagenome and a reference metagenome, the D-scores were calculated using a binomial distribution. We calculated the D-score using (f1–f2)/sqrt(p*q * (1/n1+1/n2)), where f1 = x1/n1 = frequency of functional occurrence in query group, f2 = x2/n2 = frequency of functional occurrence in reference group, p = (x1+x2)/(n1+n2) = probability of occurrence, q = 1−p = probability of non-occurrence. Specifically, x1 was the number of a given function in query group, x2 was the number of a given function in reference group, n1 was total counts of all function occurrences in query group, and n2 was total counts of all function occurrences in reference group. Further analysis involved D-rank, a normalization ranking for each pair wise comparison. D-rank was calculated by adding the D-scores of all protein families assigned to a certain functional category and then normalized by the square root of the number of total categories [90], [91].
In order to verify the quality of sequence assembly and discover novel biocatalysts, 24 predicted coding genes for carbohydrate degrading enzymes were amplified, among which 22 showed positive results. Among the 22, four were expressed and analyzed. The same batch of sequenced metagenomic DNAs were used as template for PCR amplification. The PCR mixture (50 µl) contained 5 µl of 10× PCR buffer, 4 µl of MgCl2 (25 mM), 1 µl of dNTP, 1 µl of each primer (10 mM), 37 µl of sterile Milli-Q water, 0.5 µl of Taqpolymerase (AmpliTaq Gold DNA Polymerase, Applied Biosystems, CA, USA), and 0.5 µl of DNA templates. PCR were carried out under the following conditions: an initial denaturation at 94°C for 5 min; 35 cycles of denaturation at 94°C 30 s, annealing at 55°C 1 min, and extension at 72°C for 1.5 min. The final step of the PCR was an extension step at72°C for 7 min, followed by cooling at 4°C. The PCR products were analyzed by gel electrophoresis. Two predicted endoglucanase genes and two xylanase genes were cloned and expressed as described by Shi et al (2011) [29]. Briefly, the endoglucanase and xylanase genes were cloned into pET161 vector (Cat No. K160-01, Invitrogen, USA) with a 6×His-tags. The enzyme expressions were induced in BL21 (DE3) cells with 0.5 mM IPTG at 25°C for 5 hours. The expressed enzymes were purified through a 5-ml nickel affinity column in AKTA FPLC system (GE healthcare, USA). Cellulase and xylanase activities were measured by the amount of reducing sugars released using dinitrosalicylic acid [92]. One unit was calculated as 1 µmol reducing sugar released per minute using glucose as standard.
This Whole Genome Shotgun project was deposited at DDBJ/EMBL/GenBank under the accession AKYZ00000000 and AKZA00000000 for grasshopper and cutworm, respectively. The version described in this paper is the first version, AKYZ01000000 and AKZA01000000. The Genbank ID for the four enzymes was as follows; cutworm EG1 is JX434086; grasshopper EG1 is JX434088; cutworm XYN1 is JX434089; and grasshopper XYN1 is KC155983.
|
10.1371/journal.pgen.1008272 | Effector gene reshuffling involves dispensable mini-chromosomes in the wheat blast fungus | Newly emerged wheat blast disease is a serious threat to global wheat production. Wheat blast is caused by a distinct, exceptionally diverse lineage of the fungus causing rice blast disease. Through sequencing a recent field isolate, we report a reference genome that includes seven core chromosomes and mini-chromosome sequences that harbor effector genes normally found on ends of core chromosomes in other strains. No mini-chromosomes were observed in an early field strain, and at least two from another isolate each contain different effector genes and core chromosome end sequences. The mini-chromosome is enriched in transposons occurring most frequently at core chromosome ends. Additionally, transposons in mini-chromosomes lack the characteristic signature for inactivation by repeat-induced point (RIP) mutation genome defenses. Our results, collectively, indicate that dispensable mini-chromosomes and core chromosomes undergo divergent evolutionary trajectories, and mini-chromosomes and core chromosome ends are coupled as a mobile, fast-evolving effector compartment in the wheat pathogen genome.
| The emerging blast disease on wheat is proving even harder to control than the ancient, still-problematic rice blast disease. Potential wheat resistance identified using strains isolated soon after disease emergence are no longer effective in controlling recent aggressive field isolates from wheat in South America and South Asia. We construct a high-quality assembly of an aggressive, recently-isolated wheat blast fungal strain and the first assembled mini-chromosome genome sequence of wheat and rice blast pathogens. We report that recent wheat pathogens can contain one or two highly-variable dispensable mini-chromosomes, each with an amalgamation of fungal effector genes and other sequences that are duplicated or absent from indispensable core chromosome ends. Well-studied effectors found on different core chromosomes in rice pathogens appear side-by-side in wheat pathogen mini-chromosomes. The rice pathogen often overcomes deployed resistance genes by deleting triggering effector genes. We propose that the fast-evolving effector-rich compartment of the wheat blast fungus is a combination of core chromosome ends and mobile mini-chromosomes that are easily lost from individual strains. Localization of effectors on mini-chromosomes would therefore accelerate pathogen adaptation in the field.
| Wheat blast is an explosive emerging disease capable of 100% yield losses. Little resistance is available in cultivated wheat varieties, and fungicides are not effective under disease favorable conditions [1,2]. The disease emerged in Brazil in 1985 and spread within South America, limiting wheat production (Fig 1). Wheat blast jumped continents in 2016, causing major yield losses in Bangladesh with this first report [3,4]. Wheat blast has now established in South Asia, enhancing fears about further disease spread, disruption of global grain trade by this seed-borne pathogen, and endangerment of global food security [5]. Wheat blast is caused by a wheat-adapted lineage of Magnaporthe oryzae (synonymous with Pyricularia oryzae) [6], known as the Triticum pathotype (MoT). MoT strains are distinct from rice pathogens in the M. oryzae Oryza pathotype (MoO) and millet pathogens in the Eleusine (MoE) and Setaria (MoS) pathotypes (S1 Fig). A serious turf grass disease emerged in the United States in the late 1980s, caused by the Lolium pathotype (MoL) with ryegrass as its major host. Although some MoL strains can infect wheat [7], MoT strains are distinguished as highly aggressive wheat pathogens that are so far restricted to certain countries in South America and South Asia (Fig 1A).
Although little is known about wheat blast, studies on rice blast disease have identified numerous effector genes, generally encoding small proteins that are specifically expressed in planta and play roles in host invasion [8–10]. Some effectors, termed avirulence (AVR) effectors, determine either rice cultivar or host species specificity through blocking infection upon recognition by corresponding cultivar- or species-specific resistance (R) genes and triggering hypersensitive resistance. For example, strains of several M. oryzae pathotypes are not able to infect weeping lovegrass, Eragrostis curvula, because they carry a host species-specific AVR effector PWL2 [11,12]. Planting of wheat varieties lacking the R gene Rwt3 in Brazil likely enabled MoL strains with the corresponding host species-specific AVR effector PWT3 to adapt to wheat, and subsequent loss of PWT3 function played a role in the wider emergence of the MoT subgroup [13]. So far, characterization of 11 MoO AVR effectors together with their corresponding R gene products has identified direct or indirect protein interactions that control rice cultivar specificity [9,14]. In contrast, understanding how individual effectors function in host invasion has been difficult due to apparent functional redundancy. That is, deletion of individual effector genes rarely dramatically impacts the pathogen's ability to cause disease.
Effector genes in diverse filamentous eukaryotic pathogens generally reside in rapidly evolving, transposon-rich chromosomal regions, which, together with slowly evolving core chromosome regions containing housekeeping genes, results in a 'two-speed' genome [15,16]. M. oryzae effectors from the Oryza pathotype are known to reside in transposon-rich regions, often near chromosome ends [9,17]. Two AVR effector genes [18,19] have been localized to dispensable mini-chromosomes (also known as supernumerary, accessory or B chromosomes [16,20,21]) that show non-mendelian inheritance and are present in some, but not all individuals in a population [22–24]. Effectors are associated with frequent presence/absence polymorphisms between and/or within the different M. oryzae lineages [18,25]. Deletion of the corresponding AVR effector gene could be a response to deploying R genes in a crop. In one well-studied case, AVR-Pita1, which corresponds to the periodically-deployed Pita rice R gene, has been mobile in the M. oryzae genome [18]. Specifically, AVR-Pita1 is found on different chromosomes in different strains, often near telomeres, and sometimes on mini-chromosomes. Understanding AVR effector gene dynamics is key to combating the ability of the blast fungus to rapidly overcome deployed R genes and to developing sustainable disease control.
Wheat blast disease is proving even harder to control than the ancient, still-problematic rice blast disease. Potential wheat resistance identified using strains isolated soon after disease emergence in 1985 are no longer effective in controlling recent aggressive field isolates from wheat in South America and South Asia. The global threat now posed by wheat blast disease makes it critical to generate genomic resources to further understand the wheat blast fungus. Here, a reference genome of an aggressive MoT strain was generated and compared to genomes of early and recent wheat pathogens and other host-adapted strains. We report that the genome structures of the 7 wheat blast core chromosomes have not diverged significantly from the rice blast core chromosomes. However, mini-chromosomes present in zero, one or two copies in different strains serve as a highly variable compartment for effector genes.
We sequenced and generated a near-complete genome assembly of the highly aggressive Bolivian field isolate B71 [4,26], which exhibits high sequence similarity with MoT isolates from Bangladesh (Fig 1A, S1 Table and S1 Fig). An assemblage containing 31 contigs (S2 Table) was produced from >12.4 Gb of whole genome shotgun (WGS) PacBio long reads (S2 Fig). Genome polishing utilizing ~10 Gb Illumina sequencing data corrected 37,982 small insertions and deletions as well as 350 base-pair substitutions in the PacBio draft assembly (S1 Data). Corrected assembled contigs were in the range of 44.2% to 52.5% GC content with the exception of a contig of 28.4%, which was predicted to be from mitochondria of B71 owing to its high similarity (99% identity) to the mitochondrial sequence of M. oryzae rice pathogen 70–15 [27]. A circularized B71 mitochondrial sequence was obtained after removing redundant sequences at the contig ends.
We developed a novel scaffolding technology, LIEP (Long Insert End-Pair sequencing) to improve the continuity of the assembly (Fig 2A). Briefly, LIEP involved construction of millions of vectors, each of which contains a unique DNA barcode pair of 22 nt and 21 nt random barcodes. Barcodes for each vector were sequenced to establish a sequence database of barcode pairs. The vectors were then used to construct clones with 20–30 kb long inserts of B71 genomic DNA flanked by the two vector barcodes. Both ends of the insert were sequenced, generating clone-end sequences with paired barcode sequences. Barcode sequences were used to recover clone-end pairs. All steps were performed with pooled clones rather than individual clones. After scaffolding, a small contig (~12 kb) with the poor support from Illumina reads was discarded. Scaffolding and filtering condensed the assembly to 12 contigs, which were then reoriented and renamed based on the MG8 genome assembly of rice pathogen 70–15 [27]. Consequently, the final B71 genome assembly (B71Ref1) is comprised of ~44.46 Mb in seven chromosomes and five unanchored scaffolds (Fig 2B).
Telomere repeat sequences (TTAGGG)n or M. oryzae telomeric retrotransposons (MoTeRs) that integrate in telomere repeats [28] were identified on both ends of chromosomes 2, 4, 5, 6, 7 and on one end of chromosome 1, indicating that B71Ref1 is a near end-to-end assembly. The B71Ref1 and MG8 assemblies show high end-to-end co-linearity for chromosomes 2, 4, 5, and 7 (Fig 2C, S2 Data). A two-megabase rearrangement was identified between chromosomes 1 and 6, of which part of chromosome 1 of MG8 was located on chromosome 6 of B71. The rearrangement was supported by eight pairs of LIEP sequences (S3 Fig) and by 50 single PacBio long reads. This rearrangement is not MoT specific because it was also observed in a MoO field isolate, evidenced by a long PacBio assembled sequence spanning both chromosome 1 and chromosome 6 of MG8 [29]. A large sequence in B71Ref1, from 1.3 to 2.9 Mb on chromosome 3, was absent in MG8. The unanchored 70–15 MG8 contig, supercont8.8, was mapped at the beginning of B71 chromosome 7, implying supercont8.8 is the missing end of chromosome 7 in the MG8 reference genome. None of five unanchored scaffolds of B71Ref1 can be mapped to MG8, with the requirement of, at minimum, a 10-kb match and 95% identity. Annotation of B71 identified 12,141 genes, with 1,726 harboring signal peptide domains (Fig 3, S3 and S4 Data). Of the 248 highly conserved core set of eukaryotic genes, 243 (98.0%) orthologs from the B71 annotation were identified by CEGMA, compared to 97.6% orthologs in MG8. Therefore, completeness and annotation of the B71 genome are at least comparable to that of MG8, which was produced using Sanger sequencing and multiple technologies.
Comparison of RNA-Seq data of MoT-infected wheat from the field in Bangladesh [3] and culture-grown MoT identified 335 and 153 genes that were only expressed in planta and in culture, respectively (SI Materials and Methods) (S5 Data). Secretion signal domains occurred in 173 in planta-specific genes, and in 18 culture-specific genes. The in planta-specific genes included homologs of five MoO effector genes, including PWL2 and PWL4 (an inappropriately expressed homolog from a weeping lovegrass pathogen that fails to block infection of Eragrotis spp.) [11,12], AVR-Pib and AVRPiz-t that determine rice cultivar specificity [30,31], and the cytoplasmic effector BAS1 [32] (Fig 3G and S5 Data). The remaining 168 in planta-specific genes were considered putative effectors (S6 Data). Both known and putative effector genes tended to be located towards the ends of core chromosomes (Fig 3G). We also generated RNA-Seq data from both B71 in planta leaf samples enriched with fungus at 40 hours post inoculation (HPI) and from B71 grown in liquid medium, which was referred to as the second RNA-Seq experiment. Differential expression analysis identified 2,891 up-regulated genes and 2,429 down-regulated genes of in planta B71 samples as compared to in vitro cultured samples. Considering genes with high fold changes in expression (at least 16x fold-change) between the two groups, we found many more highly up-regulated genes than highly down-regulated genes in planta (863 vs. 44). Of 174 known or putative effector genes, 110 were highly up-regulated at 40 HPI.
We sequenced eight additional field isolates, including less-aggressive early strain T25 isolated in Brazil in 1988 [26], five other MoT strains, a MoL strain, and a MoE strain (S1 Fig and S1 Table) [6]. A read depth approach was employed to detect genomic copy number variation (CNV) between B71 and each isolate, focusing on the identification of genomic regions with conserved copy number (CNequal), higher copy number (CNplus), or lower copy number (CNminus) in non-B71 isolates (S4 Fig). Among ~41.7 Mb of low repetitive regions, 36.4 Mb (87.3%) exhibited CNequal among all nine isolates. In total, 4.9 Mb (11.8%) displayed CNV between B71 and at least one other isolate, with 2.7 Mb (6.5%) being CNplus and 3.4 Mb (8.2%) CNminus (Fig 3D, 3E and 3F). Ten effector homologs [9] (PWL4, AVR-Pik-chr3, AVR-Pi54, BAS1-chr1, BAS2, BAS3, BAS4, AVR1-CO39, AVR-Pi9, and AVRPiz-t) resided in CNequal regions (chromosome identifier added to distinguish paralogs). Four (AVR-Pii-chr3, AVR-Pib, PWL2, and BAS1) were in CNminus regions and four (PWT3, AVR-Pii-scaf1, AVR-Pib, and AVR-Pik) in CNplus (S3 Table). CNV analysis of effector genes was supported by Illumina draft assemblies of the eight strains (S4 Table). Sequences from Illumina draft assemblies also showed sequence variation of some effector genes among these strains, such as DNA insertions in PWT3 and AVR-CO39, two AVR genes governing host specificity [13,33,34]. Thus, some AVR homologs are equal in copy number and highly conserved across all strains, while many are subject to sequence changes, including copy number changes. Of 1.2 Mb genomic sequences exhibiting CNplus in some isolates but CNminus in others, ~819 kb (68.5%) were from the five scaffolds (scaf1-5), which constitute only 4.3% of the genome. CNV variation of sequences in the B71 scaffolds indicated they are absent in the less aggressive MoT strain T25 (Fig 4A). The P3 and B71 comparison, however, suggested that most scaffold sequences are duplicated in P3, an aggressive isolate from Paraguay in 2012 (Fig 4B). In summary, extensive copy number variation was observed among M. oryzae field isolates, especially in five scaffolds that were not anchored to the seven chromosomes.
Variability in the five scaffolds led us to hypothesize that some or all scaffolds might correspond to mini-chromosome sequences in B71. Electrophoretic karyotypes of B71 using contour-clamped homogeneous electric field (CHEF) electrophoresis confirmed that B71, indeed, contained a mini-chromosome or multiple mini-chromosomes of ~2.0 Mb in size (Fig 4C). Mini- and core chromosomal DNAs were separately excised from the gel for Illumina sequencing. The five scaffolds were highly over-represented among reads obtained from the mini-chromosome DNA and highly under-represented among the core chromosome reads, confirming that all five scaffolds are from the mini-chromosome (Fig 4D). Roughly equal mean depths of B71 WGS reads mapped on all seven core chromosomes or the mini-chromosome supported that B71 contains a mini-chromosome. The mini-chromosome contains 192 protein-coding genes. Of those, 58.9% (113/192) of the genes were expressed (S5 Data). Approximately half expressed genes (N = 56) were highly regulated in expression with at least 16 fold changes comparing 40 HPI in planta samples with in vitro cultured samples, and, significantly, they were all up-regulated in planta, which indicated that genes in the mini-chromosome are likely to be associated with pathogenicity. Of 113 expressed genes, 23 were functionally annotated. Notably, the mini-chromosome contains four of all six genes in the genome that encode plasma membrane fusion proteins, and all four were highly up-regulated in planta at 40 HPI. Three functionally annotated genes exhibited in planta specific expression in the field samples or the B71 in planta leaf sheath samples, namely BSY92_12116, BSY92_11977, and BSY92_12070, encoding endochitinase B1, a gentisate 1,2-dioxygenase, and a heat-labile enterotoxin (a putative effector gene), respectively. A transcriptional regulatory gene, an Sge1 homologous gene (BSY92_12088), governing expression of secondary metabolite biosynthetic genes [35] was highly up-regulated in planta. Most other functionally annotated expressed genes are associated with putative enzymatic activities. A gene BSY92_11993 encoding ubiquitin-like-specific protease 2 was expressed in both in planta and in vitro cultured samples, but it was highly up-regulated in planta. Gene ontology (GO) enrichment analysis identified that cysteine-type peptidase activity (GO:0008234, p-value = 0.0001) was over-represented in genes on the mini-chromosome (S7 Data). Eight out of all 11 genes associated with cysteine-type peptidase activity are located on the mini-chromosome, and 7 out of these 8 were expressed in either in planta or in vitro cultured samples.
Known effector genes PWL2 and BAS1 (S5 and S6 Figs), which are located on different core chromosomes in MG8, were located immediately adjacent to one another and surrounded by various transposon sequences on the B71 mini-chromosome (Fig 5A). This configuration was supported by 211 PacBio long reads and by Sanger sequencing of a PCR product obtained with a PWL2 and BAS1 primer pair (S7 Fig). No PWL2 or BAS1 homologs, with at least 70% identity, were identified on core chromosomes, supported by an under-represented sequencing coverage on the PWL2 or BAS1 regions from CHEF sequencing of B71 core chromosomes (Fig 5D). Both genes exhibited in planta-specific expression on the mini-chromosome (Fig 5B and S8 Fig). Therefore, mini-chromosomes harbor effector genes that show similar in planta-specific expression patterns to effector genes residing on core chromosomes.
Further CHEF analyses showed no evidence of mini-chromosomes in T25 and supported at least two mini-chromosomes in P3, consistent with predictions from the CNV results. The P3 mini-chromosomes are ~1.5 Mb and ~3 Mb in length (Fig 4C). Sequences of both P3 mini-chromosomes exhibited similarities to the B71 mini-chromosome but also marked differences (Fig 4D and Fig 5E). The large P3 mini-chromosome contained both PWL2 and BAS1 genes (Fig 5C and 5D), plus it harbored ~33 kb (assembly location 6,007 to 6,039 kb) of duplicated DNA from a region near the end of chromosome 6. This duplicated DNA segment included a homolog of the MoO effector AVR-Pib [30]. In contrast, the small P3 mini-chromosome lacked the PWL2 and BAS1 genes, but it contained a duplication of approximately 0.39 Mb of the chromosome 7 end (assembly location ~3.65 to 4.04 Mb) (Fig 5D). Retention of this segment in the core chromosome explains the large CNplus segment at this region of P3 chromosome 7 (Fig 4B). The CNV result indicated that both sequences of ends of chromosome 6 and chromosome 7 found in separated mini-chromosomes have only one extra copy, supporting that P3 mostly likely has no more than two mini-chromosomes. Notably, this segment contained five putative effector genes and a homolog of the known MoO effector gene AVR-Pik [36]. Another notable region from the end of chromosome 3 was present in both P3 mini-chromosomes, but not present in the B71 mini-chromosome (Fig 4D). Sequencing P3 core chromosomes identified sequences homologous to the B71 mini-chromosome that were not present in B71 core chromosomes (Fig 4D). Taken together, these three MoT mini-chromosomes contain different sets of known or predicted effector genes and other core-chromosome end sequences, which are either missing or duplicated on the core chromosomes of the same or other strains. The highly variable structure of MoT mini-chromosomes indicates frequent acquisition of sequences from core chromosomal ends.
Repeat annotation showed approximately 12.9% of the B71 genome consisted of transposons and other repetitive elements, and transposons accounted for 9.7% and 52.8% of the core and mini-chromosomes, respectively (Fig 6A and S5 Table). Many of the transposons that were over-represented in the mini-chromosome occurred frequently on chromosome arms, particularly at chromosome ends (S9 Fig). Four transposon subclasses made up a greater proportion of the total transposon sequences on the mini-chromosome versus core chromosomes, including three LINEs (Tad1, Jockey and I) and the DNA transposon TcMar-Fot1 (Fig 6B). These four are among the top five elements enriched in the core chromosomal 20% ends relative to the 20% middle core chromosome regions (Fig 6C). Besides similarities in transposon composition between chromosome ends and the mini-chromosome, alignment of the B71 mini-chromosome sequence to core chromosomes identified duplications of >10 kb fragments with at least 95% identity. Duplications were located at ends of chromosomes 3, 4, and 7 (S9 Fig), and they were highly enriched for telomere-associated MoTeRs (LINE/CRE element). Therefore, a subset of MoT transposons is implicated in dynamic interactions between MoT mini-chromosomes and core chromosome ends.
Nucleotide composition analysis indicated that, overall, repetitive sequences along core chromosomes were highly negatively correlated with GC content (Fig 3B and 3C, S10 Fig). However, the highly negative correlation did not hold in the mini-chromosome, which is highly repetitive while maintaining relatively high GC content (S10 Fig). Repetitive sequences in many fungi, including M. oryzae MoT strains, are subject to repeat-induced point (RIP) mutation resulting in C-to-T or G-to-A transitions and, thereby, leading to reduced GC content [37–40]. Given higher GC content of repetitive sequences in the mini-chromosome versus core chromosomes, we explored the possibility of different levels of RIP in these genomic regions by assessing their RIP-type mutation rates. Of six high-abundance transposons examined, all exhibited reduced levels of RIP-type mutations in the mini-chromosome relative to core chromosomes (S11 Fig). We examined transposons MGR583 (LINE/Tad1 element) and Pot2 (DNA/TcMar-Fot1 element) that are present with multiple copies in both core and mini-chromosomes. RIP analysis indicated that no sequences of MGR583 (N = 7) or Pot2 (N = 22) from the mini-chromosome were subjected to extensive RIP-type mutations, while 14/20 MGR583 and 3/19 Pot2 from core chromosomes contained abundant RIP-type mutations (Fig 6D and S12 Fig). Therefore, unlike transposons in core chromosomes, transposons in MoT mini-chromosomes do not appear to be inactivated by the RIP genome defense mechanism.
The B71 reference genome for the wheat blast fungus has shown a high degree of macrosynteny for the core chromosomes relative to the rice pathogen reference genome 70–15 (MG8), which supports the recent report maintaining M. oryzae as a single species [6]. In contrast, mini-chromosomes present in B71 and another recent MoT field isolate P3 (P3-large and P3-small mini-chromosomes) are highly variable, with each one containing shared and different MoO effector homologs, putative effector genes, and other sequences from core chromosome ends. The B71 and P3-large mini-chromosomes contain the only copies of known MoO effectors PWL2 and BAS1 in these strains and neither gene was present in the early MoT strain T25, which we show lacks mini-chromosomes. PWL2 and BAS1 are located on different core chromosomes in 70–15, but they are found side-by-side on the B71 mini-chromosome. Both effectors show similar in planta specific expression on the MoT mini-chromosomes and on the MoO core chromosomes. Only the P3-large mini-chromosome contains a homolog of the MoO AVR-Pib gene, and only the P3-small mini-chromosome contains a homolog of AVR-Pik. Each mini-chromosome contains many other sequences that are either duplicated from core chromosome ends or missing from core chromosomes altogether. In one case, a P3 core chromosome sequence was homologous to the B71 mini-chromosome but not present in B71 core chromosomes. Taken together, our findings provide new insight on the M. oryzae two-speed genome [15] previously known to involve effector localization in transposon-rich regions near chromosome ends. We expand understanding of this effector compartment to include two apparently interchangeable regions, non-dispensable core chromosome ends coupled to dispensable mini-chromosomes.
We show that the M. oryzae accessory mini-chromosomes have a unique set of properties relative to accessory chromosomes in other fungi, including the well-studied accessory chromosomes in Fusarium species [41,42] and in Zymoseptoria tritici (syn. Mycosphaerella graminicola) [43,44]. M. oryzae mini-chromosomes, like lineage-specific chromosomes in Fusarium spp. [41,42] and the mini-chromosome in Leptosphaeria maculans [45] contain multiple genes associated with virulence and host-specificity. However, extensive recombination with core chromosomes has been so far only observed in M. oryzae mini-chromosomes. The rich set of accessory chromosomes in Z. tritici lack genes with an obvious role in pathogenicity, although some contribute quantitative pathogenicity effects in some strains [46]. The Z. tritici accessory chromosomes appear to be relatively ancient based on apparent survival through at least one speciation event [43,46]. M. oryzae mini-chromosomes resemble the accessory chromosomes of F. poae in lacking signs of the fungal specific genome defense mechanism known as RIP [47], therefore differing from the mini-chromosome and AT-isochore regions of L. maculans for which RIP appears to be a major mechanism for effector gene mutation during response to R gene deployment [48,49]. The gene and transposable element crosstalk between the core and supernumerary genomes reported in F. poae does not preferentially involve effectors and core chromosomes ends such as we report for M. oryzae [47]. Although supernumerary chromosomes in many systems appear heterochromatic, with low levels of gene expression [21], effector genes in M. oryzae mini-chromosomes show in planta specific expression characteristic of these genes on core chromosomes. Therefore, mini-chromosomes in the wheat blast pathogen differ in degree of variability of effector gene content and extent of recombination with core chromosome ends compared to dispensable chromosomes characterized so far in other fungi.
The mechanism for sequence exchange between core- and mini-chromosomes is unknown. However, the enrichment in mini-chromosomes of multiple subclasses of LINE retro-transposons and a DNA transposon that are also enriched at core chromosome ends, points to a transposon-mediated recombination mechanism involving non-allelic homology. Such a mechanism has been shown to facilitate genome rearrangements in another phytopathogenic fungus [50]. In contrast to seemingly RIPed core chromosome copies, the multiple copies of both MGR583 (LINE element) and Pot2 (DNA element) in the mini-chromosomes are nearly devoid of RIP-type mutations. This suggests that transposons on the mini-chromosomes remain active, facilitating multiplication and recombination. Telomere-associated MoTeR elements, found in MoL strains but not in MoO strains, are present on MoT mini-chromosomes. MoTeR elements have been reported to account for the extreme sequence variability of MoL telomeres compared to MoO telomeres [28], suggesting these elements might enhance mini-chromosome dynamics in MoT and MoL strains through destabilization of telomere regions. Transposon-rich genomic regions have been linked to increased sequence and structural variation in fungal plant pathogens [15,25,46]. Therefore, transposon-rich mini-chromosomes that also carry a number of genes, including many putative effectors, likely serve as genomic hotspots promoting genomic variation. Exceptional genomic variation produced in mini-chromosomes, and capable of flowing into core chromosomes, could accelerate the evolutionary potential of the pathogen.
Dynamic interchange between mini-chromosomes and core chromosome ends would contribute to AVR-Pita1 effector gene mobility, which is especially characteristic of rice pathogens [18]. M. oryzae rice pathogens are notorious for their ability to rapidly overcome deployed R genes. AVR-Pita1 and AVR-Pita2, which each confer avirulence to rice carrying the corresponding Pita resistance gene, belong to a subtelomeric gene family (S4 Table) and show a high rate of spontaneous mutations, including frequent deletions [51,52]. AVR-Pita1 and AVR-Pita2 occur in zero, one or more copies in different M. oryzae isolates and show highly variable genomic locations, usually near ends on core chromosomes 1, 3, 5, 6, 7; in 3 separate locations on chromosome 4; and on supernumerary chromosomes [18]. In contrast, avr-pita3, which lacks AVR activity, is stably located on chromosome 7 across the host-adapted lineages of M. oryzae. Therefore, extremely high genomic mobility, particularly of AVR-Pita1, appears to be a response to the periodic deployment of the Pita gene in rice. Mini-chromosomes would provide a population-wide repository for AVR genes that are deleted from individual strains and a means for rapid loss of AVR gene function from individual strains, because mini-chromosomes are frequently lost during meiosis and mitosis [22–24,53]. Individual strains lacking AVR-Pita1 could regain it through acquiring AVR-Pita1 containing mini-chromosomes from other individuals through the parasexual cycle and lateral gene transfer [18]. This would explain how the gene became integrated into new locations on the core chromosomes, typically at chromosome ends. The dynamic coupling we report between mini-chromosomes and core chromosome ends supports the multiple translocation hypothesis for AVR genes responding to periodic negative selection pressure of R gene deployment. Collectively, we propose that the mini-chromosome plays a role for gene movements like a shuttle, in which mutation, duplication, loss, and rearrangements of DNA occur at a faster pace than normal genomic changes, hence, accelerating genomic evolution for adaptation.
Growing evidence suggests that avirulence-conferring PWL family members (S4 Table) may be undergoing multiple translocation similarly to AVR-Pita family members [18]. PWL2 from a rice isolate and PWL1 from an Eleusine isolate each confer avirulence toward Eragrostis spp. [11,12]. The well-studied PWL2 gene, like AVR-Pita1, occurs in zero to four copies in different strains and is subject to frequent spontaneous deletion [12]. Genetic analyses showed that PWL2 and PWL1 map to different chromosomal locations, with PWL1 linked to a telomere. Homology between the PWL2 and PWL1 genes begins 70 bp upstream of the PWL1 initiation codon and ends immediately after the stop codon, and sequences beyond this conserved region are completely unrelated. In contrast, the apparently allelic, non-AVR conferring PWL3 and PWL4 genes mapped to a third genomic region and share conserved flanking sequences [11]. Two copies of PWL2 are present on chromosomes 3 and 6 in the reference rice genome MG8, the intact PWL2 sequence was found in three assembled contigs of a highly aggressive rice isolate 98–06 [54], and we report that PWL2 resides on a mini-chromosome in some wheat pathogens. Further research is needed to track chromosomal dynamics of PWL2, as well as PWL1, in host-adapted forms of M. oryzae. AVR-Pita1 effector gene mobility is reported to be in response to periodic deployment of the corresponding Pita gene in rice, raising the question of comparable selection pressure that might be acting in the Eragrostis system. Introduction of weeping lovegrass, native to South Africa, and other Eragrostis spp., around the world for forage and erosion control in the past decades could have provided conditions promoting loss and recovery of PWL family members.
Our results will inspire further exploration of function and evolutionary roles of mini-chromosomes in the fungal phytopathosystem, and facilitate answering important questions for blast on wheat and other cereal crops. Our early MoT strain T25, isolated in 1988, lacks mini-chromosomes, as was previously reported for 7 other MoT strains isolated in Brazil between 1986 and 1988 [23]. This raises the question of whether mini-chromosomes have contributed in any way to the enhanced aggressiveness characteristic of recent field isolates such as B71 and P3. It is critical to monitor further evolution, including potential recombination with other M. oryzae pathotypes, of the complex MoT population in South America and the initially clonal MoT population in South Asia [1,4]. Localization of the PWL2 host species-specificity gene on mini-chromosomes in wheat pathogens raises the question of a role for mini-chromosomes in host jumps. Effector gene dynamics, so far only associated with a small number of MoO AVR effector genes corresponding to periodically deployed R genes, raises the question of what roles known MoO AVR effector homologs and BAS1 (lacking known AVR activity in MoO strains) play in wheat infection by MoT strains. Finally, it is critical to identify and deploy effective wheat blast resistance.
Detailed description of materials and methods is included in S1 Text.
All M. oryzae strains examined were field strains from South America (S1 Table). MoT isolates B71, T25, and P3 were isolated in Bolivia (2012), Brazil (1988), and Paraguay (2012), respectively. All work with living wheat blast fungus in the U.S. was performed with proper USDA-APHIS permits and monitoring in BSL-3 laboratories in the Biosecurity Research Institute at Kansas State University.
Single spore isolates of each pathogen strain were cultured in complete medium for mycelium propagation. Mycelium was harvested and frozen using liquid nitrogen. To avoid excessive mitochondrial DNA, mycelial nuclei were collected by gradient centrifugation as described [55]. The CTAB (cetyltrimethylammonium bromide) DNA extraction method was applied to isolate genomic DNA from the nuclear samples [56].
The 3–20 kb WGS libraries were constructed using B71 nuclear genomic DNAs. The library was sequenced with P6-C4 chemistry on ten SMRTcells of PacBio RS II. Nuclear genomic DNAs were also subjected to 2x250 bp paired-end Illumina sequencing. To increase the assembly continuity, LIEP was devised and used to generate 20–30 kb long-distance paired sequences for scaffolding. PacBio long reads were assembled using the Canu pipeline [57]. Self-correction using PacBio reads did not correct all PacBio sequencing errors. Illumina reads and the Illumina assembly sequences assembled using DISCOVAR de novo [58] were both utilized for further error correction. The resulting assembled contigs were scaffolded using LIEP long-distance paired sequences with the software SSPACE [59].
Two RNA-Seq experiments were performed. In the first RNA-Seq experiment, an in vitro cultured mycelium sample was collected for the total RNA extraction using RNeasy Plant Mini Kit. Total RNA was used for RNA sequencing on a MiSeq to generate 2x150bp paired-end data. Clean data after adaptor and quality trimming were de novo assembled using Trinity [60], which were then aided in genome annotation.
In the second RNA-Seq experiment, we attempted to compare B71 gene expression in planta and in vitro culture with three biological replicates in each group. RNAs of in planta samples were isolated from B71-infected epidermal cells of leaf sheaths from 3–4 weeks old wheat plants at 40 HPI. The B71 in vitro culture RNAs were extracted from mycelium grown in liquid swirling cultures with minor modifications to the method of Mosquera et al. [32]. The total RNAs were subjected to library preparation for mRNA sequencing to produce single-end 75bp reads. Clean data after adaptor and quality trimming were aligned to the B71Ref1 reference genome with STAR [61]. Read counts per genes were used for differential expression analysis with DESeq2 [62] with 1% false discovery rate (FDR) as the threshold to declare significantly differentially expressed genes between in planta and in vitro culture groups [63].
A Maker pipeline was used for the B71 genome annotation [64]. Both evidence-driven prediction and ab initio gene prediction were employed [65]. Transcriptional evidence was provided using assembled sequences from RNA sequencing data of the B71 strain that was cultured in media and field wheat leaf samples infected by Bangladesh wheat blast strains, which were genetically almost identical to B71. CEGMA was used to assess the completeness of the genome assembly or annotation [66].
Publically available RNA-Seq data of MoT infected wheat were used as in planta expression data to compare with in vitro culture RNA-Seq data from the first RNA-Seq experiment. Field RNA-Seq data includes samples 5 and 7 from Bangladesh wheat fields [3]. These MoT isolates have been demonstrated to be almost identical to B71. All data from samples 5 and 7 were merged to represent field in planta transcriptomes. Genes with read abundance higher than 0.1 FPKM (fragment per kilobase of coding sequence per million reads) in either in planta or in culture samples were considered to be expressed genes. Genes with read abundance higher than 1 FPKM from the in planta data set but no reads from the cultured sample were considered to be in planta specific expression. In planta specific genes containing classical signal peptide domains [67] were considered putative effectors.
Read depth approach was employed to identify CNV between each of some M. oryzae strains and B71 for each of sequence bins (e.g, 300 bp). Segmentation with the R package of DNACopy was performed to identify genomic CNV segments merged from multiple bins [68].
MoT protoplasts were prepared and mixed with 1.5% low melting-temperature agarose [23]. Suspensions were loaded into disposable plug molds. Protoplasts in plugs were lysed with proteinase K and washed. A Biorad CHEF electrophoresis system was used for separating chromosomes embedded in the plugs. After the CHEF gel electrophoresis, DNAs from individual mini-chromosomes, and from core chromosomes as one unit, were excised and purified from the agarose gels. Purified DNAs were subjected to Illumina 2x151 bp paired-end sequencing.
Repetitive sequences were identified using MGEScan [69], LTR_Finder [70], LTRharvest [71,72], and RepeatModeler (github.com/rmhubley/RepeatModeler). Merging discovered repetitive sequences and previously characterized M. oryzae repeats [73] produced a non-redundant database, which served as a repeat library to identify repeats in the B71 genome using RepeatMasker (www.repeatmasker.org). Some transposable elements were subjected to analysis of RIP-type polymorphisms, nucleotide changes of C to T or G to A.
|
10.1371/journal.pbio.0060234 | Capturing Hammerhead Ribozyme Structures in Action by Modulating General Base Catalysis | We have obtained precatalytic (enzyme–substrate complex) and postcatalytic (enzyme–product complex) crystal structures of an active full-length hammerhead RNA that cleaves in the crystal. Using the natural satellite tobacco ringspot virus hammerhead RNA sequence, the self-cleavage reaction was modulated by substituting the general base of the ribozyme, G12, with A12, a purine variant with a much lower pKa that does not significantly perturb the ribozyme's atomic structure. The active, but slowly cleaving, ribozyme thus permitted isolation of enzyme–substrate and enzyme–product complexes without modifying the nucleophile or leaving group of the cleavage reaction, nor any other aspect of the substrate. The predissociation enzyme-product complex structure reveals RNA and metal ion interactions potentially relevant to transition-state stabilization that are absent in precatalytic structures.
| Enzymes use variations of a few standard approaches to catalyze reactions. One of these approaches, acid–base catalysis, is of such fundamental importance that it is common to both protein enzymes and RNA-based enzymes, or ribozymes. The hammerhead ribozyme is one such ribozyme that uses an invariant guanine residue as a general base in its catalytic reaction. By changing this to an adenine, we can slow the reaction rate 100,000-fold, permitting us to capture both active, precatalytic, and postcatalytic forms of the ribozyme. We have exploited this approach to obtain near-atomic–resolution three-dimensional structures of the hammerhead ribozyme both before and after catalytic self-cleavage. These structures provide complementary views of the chemical step of hammerhead ribozyme catalysis.
| The hammerhead ribozyme, since its discovery in satellite virus RNA genomes [1,2], has been a central focus of experiments designed to correlate RNA structure with RNA catalysis, as it is a comparatively small RNA whose biochemistry has been intensively investigated using a wide variety of approaches [3–5]. Recently, the discovery that natural hammerhead RNAs having tertiary contacts distant from the active site may enhance catalysis up to approximately 1,000-fold relative to “minimal” hammerheads [6–9] compelled renewed mechanistic and structural investigations.
Natural hammerhead ribozymes fall into two distinct classes [6] based upon the nature of the tertiary contacts between Stem I and Stem II (Figure 1). The most well-characterized member of the first class of natural hammerheads occurs within the satellite RNA of the tobacco ringspot virus (sTRSV), which is also the first hammerhead ribozyme discovered [10]. The best-characterized member of the second class of natural hammerheads occurs within the multimeric RNA transcript of the Schistosoma mansoni alpha repetitive sequence (Smα) repetitive DNA within the S. mansoni genome [11,12]. The structure [13] of a full-length Schistosome hammerhead [12] ribozyme-competitive inhibitor complex in which a substrate analog having a modified 2′-OMeC17 nucleophile was recently obtained, revealing how G12 becomes positioned to initiate cleavage as a general base, and how G8 may function as a general acid in hammerhead ribozyme catalysis. However, the substrate was inactivated by replacing the nucleophilic 2′-OH of the cleavage-site nucleotide (C17) with an inert ether linkage, thus potentially altering the active site environment.
We have now obtained two crystal structures from a full-length sTRSV hammerhead RNA with an unmodified cleavage site that has an active nucleophile. These include an active enzyme–substrate complex trapped just prior to catalytic cleavage from freshly grown crystals, and an active enzyme–product complex trapped prior to dissociation of the product, subsequent to cleavage (Figure 2) from crystals allowed to age for several weeks.
Instead of inactivating the nucleophile via methylation, as was done with the Schistosome hammerhead [12], the cleavage reaction in the case of the sTRSV hammerhead has been greatly decelerated with a G12A enzyme active site variant that lowers the pKa of the purine general base in the cleavage reaction from approximately 9.5 to approximately 3.5, which, assuming the observed log-linear rate dependence [6,14,15] on pH, potentially represents an approximately 106-fold decrease of the reaction rate. The G12A mutation in the context of a minimal hammerhead ribozyme has been reported previously to create a greater than 500-fold reduction in the cleavage rate [16]. More recently, a full-length peach latent mosaic viroid hammerhead ribozyme with a G12A substitution has been shown to have very limited cleavage activity [17]. We have measured an approximate 10−6-fold rate reduction for the G12A substitution in the full-length hammerhead, and have also shown the G12A modification retains the standard pH dependence of the hammerhead reaction rate (cf: Figure S4). The correlation between the pKas of various purine derivatives substituted at position 12 and the hammerhead ribozyme cleavage rate has been thoroughly examined using inosine, diaminopurine, and 2-aminopurine nucleotides substituted for G12 [18]. These results are all consistent with the purine at G12 functioning as a general base, as well as with the G12A mutant being a very poor, but not completely inactive, general base. By greatly slowing the reaction, the hammerhead RNA crystallizes prior to cleavage, but remains active in the crystal and slowly cleaves. We have exploited this property to obtain both reactant (precatalytic) and product (postcatalytic) structures of the active hammerhead ribozyme to 2.4 Å and 2.2 Å resolution, respectively.
In our study, two datasets were used; one, the reactant, diffracts to 2.4 Å resolution and the other, the cleavage product, diffracts to 2.2 Å resolution. In both datasets, two crystallographically independent 69-nucleotide hammerhead structures (Figure S1) occupy a P1 unit cell (a = 27.9 Å, b = 53.0 Å, c = 72.0 Å, α = 74.6°, β = 81.4°, γ = 75.6°) [19]. The only significant difference between molecule 1 and molecule 2 within the asymmetric unit is in the tertiary contact region, where the electron density for several of the nucleotides involved in the tertiary contact in molecule 2 is quite weak, indicating disorder and dynamic flexibility in a structure otherwise characterized by a well-resolved and easily interpretable electron density map. Two precatalytic (uncleaved) models were unambiguously constructed in the 2.4 Å electron density map and refined. Refinement of the reactant structure of the 2.4 Å data (Tables 1 and 2) clearly shows that both molecules in the asymmetric unit are in an uncleaved, precatalytic state, whereas both molecules in the asymmetric unit of the product 2.2 Å structure (Tables 1 and 2) are in a cleaved, postcatalytic state.
The precleavage or enzyme–substrate complex structure of the G12A sTRSV hammerhead RNA at 2.4 Å resolution reveals an active site (Figure 3A) very similar to that of the Smα hammerhead (Figure 3B), despite the presence of the 2′-OMe modification in the latter, and the G12A substitution in sTRSV hammerhead. Hence, it is reasonable to conclude that neither modification grossly perturbs the atomic structure of the hammerhead ribozyme active site. In this sense, the uncleaved sTRSV hammerhead and the Smα hammerhead structures are both useful internal experimental controls that put to rest any concerns that either the previous 2′-OMe modification or the current G12A substitution induces formation of a catalytically incompetent hammerhead ribozyme structure. A thorough analysis of two decades of experimental results obtained from biochemical and mechanistic investigations of the hammerhead ribozyme has been carried out [20,21] that confirms the assessment that the Smα hammerhead active site conformation, and therefore the similar sTRSV hammerhead active site conformation, indeed represent the catalytically competent structural state.
Some small differences between the sTRSV hammerhead enzyme–substrate complex structure and the corresponding Smα hammerhead enzyme–inhibitor complex do exist (Figure 3C). The unmodified 2′-OH of C17 in the latter appears to be slightly more in-line with the scissile phosphate (168.5° vs. 162°), and the position of A12 differs slightly, due to a different hydrogen-bonding interaction with A9 that replaces the G12/A9 sheared pairing (Figure 3A–3C). The primary difference is that the hydrogen bond between the exocyclic amine of G12 and N7 of A9 is, by necessity, absent in the G12A structure, so that only one hydrogen bond between A9 and A12 exists (Figure 3A) rather than three (Figure 3B). The net effect is that the positions of A9 and A12 in the G12A sTRSV structure change slightly compared with the G12 structure (2GOZ), as can be seen in the superposition of the active site residues (Figure 3C). The difference in absolute positions of the scissile phosphorus in the two superimposed structures is 1.7 Å. The geometry of the G12A sTRSV appears to be somewhat better suited to initiation of the cleavage reaction. Specifically, the angle between the N1 of A12, the 2′O nucleophile (C17), and the adjacent scissile phosphorus is 149°, and the distance between N1 and O2′ is 2.7 Å. The corresponding angle in the G12 structure with the modified substrate (2GOZ) is 139° and the N1 to O2′ distance is 3.5 Å. The in-line attack angle (between O2′, P, and O5′) is 168° in the G12A structure, versus 162° in the previous G12 structure. Hence, the slow cleavage rate appears to be primarily due to the result of the purine pKa shift from approximately 9.5 to approximately 3.5 upon G12A substitution, rather than due to a disadvantageous structural perturbation. Deprotonation of G12 must occur (Figure 2) to initiate the cleavage reaction, but G12 is almost certainly protonated in the 2GOZ crystal structure at pH 6.5, whereas A12 is normally deprotonated at neutral pH. In this sense, A12 may be a better (albeit much slower) representation of the activated ribozyme poised for general base catalysis, even though A12 is a much weaker base than G12 due to its much smaller pKa.
Refinement of a hypothetically uncleaved structure using the 2.2 Å resolution cleavage product dataset, obtained from the crystals allowed to age, revealed unique and significant (>3 σ) negative difference Fourier peaks (Figure 4A) positioned directly on the O5′ atoms of A1.1 of each molecule of the hypothetically uncleaved model (without noncrystallographic symmetry averaging applied), in addition to clear breaks in the sigma-A–weighted 2Fo-Fc maps [22–25] at the same locations (Figures 4B and S1B), thus demonstrating that the substrate RNA is predominantly in the cleaved state. The negative difference Fourier peak on molecule 1 (Figure 4A) is slightly more pronounced, and subsequent refinement of the structure in which a 2′,3′-cyclic phosphate was added to C17, and the phosphate linking it to A-1.1 was replaced with a terminal 5′-OH, provided a much better fit to the observed electron density (Figure 4B and 4C). Molecule 1 appears to be completely cleaved, whereas a small amount of molecule 2 may remain in the uncleaved form. Cleavage of molecule 2 is thus best interpreted as somewhat incomplete, and it is notable that possibly less-complete cleavage corresponds to the molecule in which the tertiary contact is less well defined, hinting that the tertiary contact may function as a molecular modulator in the life cycle of the satellite virus RNA that regulates cleavage and possibly religation activities. The internal equilibrium of the sTRSV hammerhead ribozyme greatly favors the cleaved over the uncleaved state, whereas the internal equilibrium of the Smα hammerhead is such that about 1/3 of the RNA is ligated [26].
The cleaved structure reveals several interactions potentially relevant to the catalytic mechanism (Figure 4C). In molecule 1 of the cleavage product structure, two Mg2+ ions appear to interact with the scissile phosphate, which is in the 2′,3′-cyclic form. In addition, the 2′-OH of G8, previously implicated as possibly the acid catalyst [13,18], makes a hydrogen bond to the more proximal nonbridging phosphate oxygen of the cyclic phosphate. N1 and N6 of A9 are also positioned about 4.5 Å from the same nonbridging cyclic phosphate oxygen atom, as is the Mg2+ ion bound to A9 phosphate. Although these latter distances, shown as orange and yellow dotted lines in Figure 4C, are too large to form bonding interactions in the product structure, it is plausible that they form stabilizing interactions within the trigonal bipyramidal oxyphosphorane transition-state structure to help disperse transiently accumulating excess negative charge, thus contributing to catalysis. (An analogous role for adenosine bases is observed in the hairpin ribozyme [27], and a requirement for either divalent metal ions or a high concentration of positive charge [28] in the hammerhead cleavage reaction is well known.) A second Mg2+ ion is observed in molecule 1 to coordinate directly with the other nonbridging cyclic phosphate oxygen, suggesting a possible role for the second Mg2+ ion in stabilizing the cleavage product or transition state. Although a single divalent metal ion has yet to be observed in a hammerhead crystal structure to bridge the scissile and A9 phosphates via a predicted inner-sphere coordination [29], the observed Mg2+ ion and A9 nucleotide base interactions nonetheless suggest how transition-state stabilization, especially at low ionic strength, may be facilitated. Since this postcatalytic structure represents the state of the molecule before product dissociation, due to trapping by the crystal lattice, we suggest that the structure reveals features relevant to the transition state and that are complementary to those in the uncleaved state.
The structure of the Schistosoma Smα hammerhead [13] revealed how the distal tertiary contacts stabilize a conformational change (relative to the minimal structure) within the active site of the hammerhead ribozyme. However, most of the naturally occurring viral hammerhead RNAs, including the sTRSV hammerhead, belong to the other class of hammerhead ribozymes in which a tetraloop on Stem II (typically the thermodynamically favored GNRA tetraloop) interacts with a closed loop on Stem I [6]. The Smα hammerhead and the sTRSV hammerhead tertiary contacts induce what are nearly identical conformational changes in the ribozyme's catalytic core, despite the fact that the sequences and structures of the two tertiary contact regions are radically different. In fact, only one tertiary base pair is common to both classes of hammerhead tertiary contacts (Figures 5, S2, and S3).
In both classes of hammerhead tertiary contacts, an apparently conserved [6] Hoogsteen base pair forms between an A in Stem-Loop II and a U in the nonhelical region of Stem I. The A in the Hoogsteen pair corresponds to position 46 in the sTRSV hammerhead and L6 in the Smα hammerhead, and the U corresponds to position 19 in the sTRSV hammerhead and B5 in the Smα hammerhead. Of the 13 natural hammerhead sequences considered in previous modeling studies [6], all possess this final A in the GNRA tetraloop capping Stem II, and ten possess this U adjacent to residue 1.6, suggesting the AU Hoogsteen pair is conserved due to its functional relevance, despite the fact that it evaded identification [6] before now. (The remaining three sequences have C instead of U, which can form an analogous Hoogsteen pair if protonated.) In the new sTRSV hammerhead crystal structure, the conserved AU Hoogsteen pair is found within a base triple in which another (apparently nonconserved) U from the Stem I loop forms an additional Watson-Crick base pair with the A from the Stem II loop (Figures 5, S2, and S3).
Until 2003, it was not recognized that a tertiary contact region possessing little recognizable sequence conservation is critical for optimal catalysis [6,7], and subsequently, it was discovered that the tertiary contacts, which impart an approximately 1,000-fold rate enhancement, induce a dramatic conformational change within the hammerhead ribozyme active site that activates it for catalysis [13]. We report here the first to our knowledge, full-length hammerhead ribozyme crystal structures in which the crystallized molecule is catalytically active, permitting capture of both the active precleavage enzyme–substrate and the postcleavage enzyme–product complexes. The former appears to be in an active conformation immediately preceding catalysis, and the latter is in a predissociated state that immediately follows catalysis. Each, therefore, provides complementary views of the unobservable transition state, the former immediately before, and the latter immediately after formation of the transition state, providing new mechanistic insights into ribozyme catalysis.
RNA sample preparation and crystallization have been previously reported [19]. Briefly, using in vitro transcription from a synthetic DNA template derived from the sequence of the sTRSV, a self-cleaving hammerhead ribozyme (69 nucleotides long) was synthesized. Since the wild-type transcription product cleaved to completion in the transcription reaction, a sequence having a mutation at position 51 (a G12A modification, using the canonical hammerhead numbering scheme), which resulted in a greatly reduced rate of cleavage, was transcribed and crystallized. The sample was purified on a fast protein liquid chromatograph (FPLC) using a diethyl aminoethyl (DEAE) ion exchange column, and the crystals were obtained by vapor diffusion as previously described [19]. For data collection, 30% (v/v) MPD was added gradually to the mother liquor, equilibrating the crystal-containing drops stepwise over a period of 3 d, before being flash frozen by liquid nitrogen stream. The cleavage-product 2.2 Å dataset, in which the RNA is predominantly cleaved, resulted from crystals that had aged substantially longer than the crystals used to collect the reactant dataset.
The native datasets from single crystals were collected at 100 K on an R-AXISIIC imaging plate detector coupled with a Rigaku Rotaflex X-ray generator and the MSC/Yale mirror optics. The datasets were processed with DENZO [30] and scaled with rotavata/agrovata implemented in the CCP4 program suite [24,31]. The final data statistics are shown in Table 1. The approximately 10% overall incompleteness was due primarily to the absence of crystal symmetry (the space group is P1).
The reactant crystal structure was determined to 2.4 Å resolution by piecewise molecular replacement using multiple copies of a seven base-paired poly-adenine standard A-form double helix (stem) as a model. The initial crystal content analysis indicated that there are two sTRSV hammerhead molecules in the asymmetric unit (VM = 2.23 Å3 Da−1, 56% solvent content, assuming RNA density is 1.7 g/cm3 [32,33]. The molecular replacement search for six stems of which three potentially constitute one hammerhead structure was carried out by Phaser [34] and a solution (Z-score 7.7 for a translation function that was 15% higher than the next possible solution) was obtained. Subsequent rigid body refinement using CNS [35] resulted in the Rfree and R values of 46.1% and 49.6%, respectively. The phases were improved by rounds of manual rebuilding and composite omit map calculation implemented in CNS, which enabled building of more than 80% of the structure. Finally, the proper connections and the correct sequence registrations were made with the aid of a newly determined full-length Smα hammerhead structure [13]. The subsequent refinement was carried out by Refmac [22] and Phenix [25,36]. The model was constructed, and Mg2+ and water sites were identified and validated using COOT [23].
The cleavage-product structure was then solved using the coordinates of the refined, uncleaved structure. The cleaved state of the substrate was identified using sigma-A–weighted (Fcalc-Fobs) difference Fourier maps calculated in both Refmac and Phenix, and then displayed in COOT. The detailed refinement statistics are shown in Table 2.
Coordinates and amplitudes for the cleaved (2QUW) and uncleaved (2QUS) structures are available in the Protein Data Bank (http://www.rcsb.org).
|
10.1371/journal.pbio.1002599 | Unique 5′-P recognition and basis for dG:dGTP misincorporation of ASFV DNA polymerase X | African swine fever virus (ASFV) can cause highly lethal disease in pigs and is becoming a global threat. ASFV DNA Polymerase X (AsfvPolX) is the most distinctive DNA polymerase identified to date; it lacks two DNA-binding domains (the thumb domain and 8-KD domain) conserved in the homologous proteins. AsfvPolX catalyzes the gap-filling reaction during the DNA repair process of the ASFV virus genome; it is highly error prone and plays an important role during the strategic mutagenesis of the viral genome. The structural basis underlying the natural substrate binding and the most frequent dG:dGTP misincorporation of AsfvPolX remain poorly understood. Here, we report eight AsfvPolX complex structures; our structures demonstrate that AsfvPolX has one unique 5′-phosphate (5′-P) binding pocket, which can favor the productive catalytic complex assembly and enhance the dGTP misincorporation efficiency. In combination with mutagenesis and in vitro catalytic assays, our study also reveals the functional roles of the platform His115-Arg127 and the hydrophobic residues Val120 and Leu123 in dG:dGTP misincorporation and can provide information for rational drug design to help combat ASFV in the future.
| African swine fever virus (ASFV) is highly contagious and can cause lethal disease in pigs. AsfvPolX catalyzes the gap-filling reaction during the DNA repair process of the virus genome; it is highly error prone and plays an important role in the strategic mutagenesis of the virus genome. Unlike the homologous proteins, AsfvPolX has several unique structural features, including a 5′-P binding pocket, a His115-Arg127 platform, and hydrophobic residues Val120 and Leu123, which can all affect the catalytic efficiency (especially during dG:dGTP misincorporation) of AsfvPolX. These properties, especially the 5′-P binding pocket, provide an ideal structural basis for designing of small molecules, which can specifically inhibit the activity of AsfvPolX and disrupt the DNA repair process of the ASFV genome.
| African swine fever virus (ASFV) is highly contagious and can cause lethal disease in both domestic pigs and wild boars [1]. ASFV is an endemic disease, and it remained restricted to Africa prior to 1957 [2]. Since then, ASFV has been found in many countries throughout Europe, including Sardinia in Italy, the Caribbean, the Caucasus region, and Russia, and has caused very serious economic problems in some local regions [3]. In 1971, more than 500,000 pigs were killed in Cuba to prevent a nationwide animal epidemic, which was labeled the “most alarming event” of 1971 by the United Nations Food and Agricultural Organization [4]. In recent years, ASFV has also been introduced to other continents such as Asia and is turning into a global threat [5,6]. Although ASFV has been extensively studied in the past, no vaccine or other useful treatment against this virus has been developed until now [7].
ASFV belongs to the genus Asfivirus, a unique member of the family Asfarviridae; it is a large, encapsulated, double-stranded DNA virus and is one of the most complex known viruses. The genome of ASFV is approximately 170–190 kb in size, encoding more than 150 proteins that function in various biological processes, such as gene transcription, DNA replication, and suppression of host immune response as well [8]. Swine macrophages and monocytes are the primary target cells of ASFV [9]. The DNA synthesis process of the virus is initialized in the host cell nucleus, whereas, the replication and virion assembly are completed in the cytoplasm, in which the virus genome is exposed to a damaging and mutagenic environment [10,11]. To overcome potential damage to the DNA such as apurinic/apyrimidinic (AP) sites and/or single strand breaks, the virus has evolved its own DNA repair system, composed of one AP endonuclease [12], one repair DNA polymerase (ASFV DNA Polymerase X [AsfvPolX]) [13], and one DNA ligase (AsfvDNAL) [14]. Unlike their homologous proteins, both AsfvPolX and AsfvDNAL can tolerate various base mismatches at the repair site; therefore, apart from their critical role in genome stability maintenance, these enzymes play an important role in the strategic mutagenesis of the ASFV genome.
Owing to their functional importance, the enzymes involved in the DNA repair system of ASFV have been extensively studied [15,16]. However, only limited structural information is available. To date, the structures of AP endonuclease and DNA ligase (DNAL) of ASFV have not been determined. AsfvPolX is composed of 174 amino acids, with several AsfvPolX nuclear magnetic resonance (NMR) structures being reported [17–19], which reveal the domain architecture of AsfvPolX and the formation of Hoogsteen pairing during the dG:dGTP misincorporation. AsfvPolX is the most distinctive DNA polymerase identified to date; compared to homologous proteins, such as rat DNA polymerase β (RatPolβ) [20], AsfvPolX lacks two important DNA-binding domains: the thumb domain and 8-KD domain. However, previous studies have indicated that AsfvPolX can efficiently catalyze the gap-filling reaction towards various substrates, including the stem-loop structured DNA utilized in the NMR structural study, recessed DNA, and regular gapped DNA (that is the natural substrate of AsfvPolX) [16]. The 5′-phosphate (5′-P) group of the downstream oligo of the gapped DNA can dramatically enhance the dG:dGTP misincorporation efficiency of AsfvPolX. However, the structural basis underlying both the natural substrate binding and the function of 5′-P of AsfvPolX remains elusive.
In the present study, we report on eight AsfvPolX crystal structures, including four AsfvPolX:DNA binary complexes and four AsfvPolX:DNA:dGTP ternary complexes. Our structures revealed a unique DNA binding mode of AsfvPolX that is different from the DNA binding modes observed in the homologous protein structures [20,21] and the AsfvPolX NMR structure [17–19]. AsfvPolX lacks the thumb domain and 8-KD domain conserved in the homologous proteins; however, our structures showed that AsfvPolX has one novel 5′-P binding pocket, which can facilitate the productive catalytic complex assembly. In combination with mutagenesis and in vitro catalytic assay, our studies also uncovered several unique structure features of AsfvPolX, which play an important role during the dG:dGTP misincorporation.
AsfvPolX belongs to the X-family DNA polymerases (PolXs), which can fill up the short gaps arising during DNA repair processes [22,23], particularly base excision repair (BER). The sequence similarities (S1 Fig) between AsfvPolX and other PolXs, including Bacillus subtilis PolX (BsPolX), Thermus thermophilus HB8 PolX (TtPolX), RatPolβ, and Homo sapiens Polβ (HsPolβ) are very low (about 30%); the identity between AsfvPolX and the homologous proteins is even lower (about 10%).
In this work, we solved eight AsfvPolX crystal structures (Table 1 and S2 Fig), including four AsfvPolX:DNA binary complex and four AsfvPolX:DNA:dGTP ternary complex structures; these structures represent two different reaction states: one prior to dNTP incorporation and one after the dNTP incorporation (S3 Fig). Four different types of DNA molecules, including blunt-ended DNA, recessed DNA, gapped DNA, and gapped DNA with 5′-P in the downstream oligo [gap(P) DNA], were captured in the structures; the detailed sequences and secondary structures of the DNA molecules were depicted in Fig 1. Besides the wild-type (WT) AsfvPolX, three mutant proteins, including H115F, H115F/R127A, and selenomethionine substituted L52/163M (Se-L52/163M, which was designed to facilitate the structure determination process using the single-wavelength anomalous diffraction [SAD] method), were also utilized in the structural studies. The crystals were grown under several different conditions (S1 Table), and, as revealed by the cell parameters and space group (Table 1), the packing of the AsfvPolX proteins was also different in most of the structures. However, unlike the NMR structures, which showed various conformations for AsfvPolXs, the AsfvPolXs in all our crystal structures adopted a conserved conformation (S2 Fig). The overall root mean square deviations (rmsd) of AsfvPolX in our structures are within the range of 0.36~0.59 Å, based on 174 pairs of Cα atoms. The rmsd (around 0.37~0.51 Å) between the N-terminal palm domains are similar to the overall structures. The most obvious conformational differences are observed in the 21EYNGQL27 region; this region is absent in the homologous proteins (S1 Fig), and it does not interact with DNA in any of our structures. The rmsd values between the finger domains are even lower, at approximately 0.23~0.36 Å.
Although the sequences and the secondary structures of the DNA molecules varied in the eight complex structures (Fig 1), superposition of the complex structures revealed one DNA-interacting site, which is common for all DNA molecules. This common DNA binding site is mainly composed of residues from two regions, 81CGERK85 from the palm domain and 135YKLNQY140 from the finger domain, and it forms extensive interactions with the DNA template strand. The AsfvPolX:DNA1 structure was utilized to demonstrate the detailed interactions (Fig 2), owing to the high resolution (1.7 Å). The binding site contains three positively charged residues (Arg84, Lys85, and Lys136), but only the NZ atom of Lys136 forms an electrostatic interaction with the OP1 atom of A4, positioning at the n-2 position, whereas, Arg84 and Lys85 mainly interact with the OP1 atom of A6 through their backbone N atoms. Three more direct H-bonds were also observed in the structure, one between the ND2 atom of Asn138 and the OP1 atom of A4, one between the OH group of Tyr140 and the OP1 atom of T5, and one between the backbone N atom of Cys81 and the OP1 atom of T7. The three nucleotides, T5, A6, and T7, are located at the n-3, n-4, and n-5 positions, respectively. No electrostatic interaction or direct H-bond forms between the backbone of G3 (locating at the n-1 position) and AsfvPolX; whereas, they interact with each other via water-mediated H-bond networks. Nucleotides, located at the positions from n-2 to n-4, also form water-mediated H-bonding with AsfvPolX, which further stabilizes the AsfvPolX:DNA complex.
In contrast to the extensive interactions between the template strand and the protein, the primer strand only forms one interaction with the protein in the AsfvPolX:DNA1 structure, the H-bond between the NE2 atom of Gln98 and the OP1 atom of C8 locating at the n-1' position; this interaction is not conserved in other AsfvPolX structures, suggesting that AsfvPolX mainly recognizes the substrate via the template strand. Besides the template strands, all the PolX homologous proteins, such as RatPolβ [20], also form extensive interactions with the primer strands (especially the nucleotides at the n-2′ and n-3′ positions) by means of the thumb domain that is missing in AsfvPolX. Together, these observations indicate that the substrate binding mode of AsfvPolX is unique among the PolX family proteins.
The natural substrates of AsfvPolX have a phosphate group (5′-P) on the 5′-end of the downstream oligo. Previous kinetic studies showed that the 5′-P can significantly increase the catalytic efficiency of AsfvPolX [16]; for instance, the reaction rate of correct dGTP incorporation against 1-nt gap(P) DNA is 15 times faster than that of corresponding DNA without 5′-P. In the homologous protein structures, the 5′-P groups were bound by the 8-KD domains [24,25], which is absent in AsfvPolX. To assess the importance of 5′-P, we carried out structural studies using three gapped DNA molecules: 1nt-gap DNA4, 2nt-gap(P) DNA5, and 2nt-gap(P) DNA6. The structure of 1nt-gap DNA4 is composed of one 15-nt template strand, one 7-nt primer strand, and one 7-nt downstream oligo without 5′-P. In the structure (Fig 3A), AsfvPolX (Se-L52/163M mutant) binds the 1nt-gap DNA4 at the blunt end instead of at the gap site. The template dG (G8) is located in the middle of the template strand and is more than 20 Å away from the active sites of AsfvPolX. Although dGTP was also present in the crystallization samples, it did not pair with G8.
The sequence of the 2nt-gap(P) DNA6 is identical to that of the 2nt-gap(P) DNA5 (Fig 3B), except that the template C9 is replaced with G9 in the 2nt-gap (P) DNA6. During the crystallization process, one ddATP (paired with T10 on the template strand) was incorporated into the 3′-ends of the primer strands of both 2nt-gap(P) DNA5 and 2nt-gap(P) DNA6; therefore, only a single-nucleotide gap was left on the two DNA molecules, which are referred to as 1nt-gap(P) DNA5 (Fig 3B) and 1nt-gap(P) DNA6 hereafter. Besides DNA, one dGTP was also captured in the two structures, which are referred to as AsfvPolX:1nt-gap(P) DNA5:dGTP and AsfvPolX:1nt-gap(P) DNA6:dGTP, respectively. As revealed by the AsfvPolX:1nt-gap(P) DNA5:dGTP structure, the dGTP pairs with the template C9 and is located at the active site of AsfvPolX (Fig 3C). Together with the Se-L52/163M:1nt-gap DNA4 structure, these structural studies suggest that the 5′-P of the downstream oligo can facilitate the complex formation of the productive AsfvPolX:DNA:dNTP.
In the AsfvPolX:1nt-gap(P) DNA5:dGTP structure (Fig 3C), the primary duplex (formed by the primer and the template strand) and the downstream duplex (formed by the downstream oligo and the template strand) all adopt B-form conformation. As depicted in S3 Fig, the conformations of the primary duplexes (especially the template strand regions) are similar in the AsfvPolX:DNA1 and the AsfvPolX:1nt-gap(P) DNA5:dGTP structures. The downstream duplex was tilted approximately 80° in respect to the primary duplex, and its duplex axis is almost perpendicular to the axis of αE in the AsfvPolX:1nt-gap(P) DNA5:dGTP structure (Fig 4A). The first base pair of the downstream duplex packs against the hydrophobic surface, which is composed of the CB2 atom of Ile124, the CB and CD atoms of Arg125, and the CB atom of Ala128, with Ile124, Arg125, and Ala128 all located in the middle region of the helix αE. As revealed by the rmsd value (1.8 Å), the overall conformations of our AsfvPolX:1nt-gap(P) DNA5:dGTP structure and the AsfvPolX:DNA:dGTP NMR structure are similar; however, in the latter, perhaps due to the interactions between the DNA loop and the side chains of Lys131 and Lys132, the downstream duplex was bent toward the helix αE (Fig 4B) [17].
To analyze the impact of the DNA structure on the substrate recognition by AsfvPolX, we also compared our AsfvPolX:1nt-gap(P) DNA5:dGTP structure with the crystal structures of HsPolβ (in complex with regular gap(P) DNA, protein data bank identification number [PDB ID]: 2FMS) [25] and TtPolX [in complex with stem-loop structured gap(P) DNA, PDB ID: 3AUO] [21]. The palm and finger domains of the three structures can be well superimposed; the rmsd values between the AsfvPolX structure and the two homolog structures are all around 1.8 Å. The overall structures of the primary duplexes are also similar in our AsfvPolX:1nt-gap(P) DNA5:dGTP structure, HsPolβ structure (Fig 4C), and TtPolX structure (Fig 4D), whereas, the orientations of the downstream duplexes in the three structures are very different from each other. In the HsPolβ and the TtPolX structures, the downstream oligos all interact with the 8-KD domains; although the orientations of the 8-KD domains are different, their interactions with the backbone and the 5′-P of downstream duplexes are conserved. Rather than our structures, the orientation of the 5′-P in the AsfvPolX NMR structure is similar to the one in the TtPolX structure (comparing Fig 4B and 4D).
The 8-KD domain is absent in AsfvPolX; however, all of our AsfvPolX:DNA:dGTP structures showed the 5′-P of the downstream oligo bound by a phosphate-binding pocket (referred to as the 5′-P binding pocket) located in the finger domain. The 5′-P binding pocket is highly positive in charge (Fig 5A). Two Arg residues (Arg125 and Arg168) and one Thr residue (Thr166) are involved in the pocket formation, and they form five H-bonds with the 5′-P of downstream oligo (Fig 5B and 5C). Arg125 forms one H-bond (2.9 Å), which is between its NH2 atom and the 5′-P OP3 atom. Arg168 forms two H-bonds: one (2.9 Å) between its NH1 atom and 5′-P OP3 atom and the other (2.7 Å) between its NH2 atom and 5′-P OP2 atom. The last two H-bonds (2.7 Å and 2.8 Å) are formed between the 5′-P OP1 atom and the backbone N atom and the side chain OG1 atom of Thr166, respectively.
Both Arg125 and Arg168 are variable in the PolX family (S1 Fig). Although some homologous proteins have Arg residues, for example, TtPolX has Arg268 (corresponding to Arg125 of AsfvPolX), and hsPolβ and RatPolβ have Arg328 (corresponding to Arg168 of AsfvPolX); none of them simultaneously have two Arg residues at the corresponding positions, indicating that the 5′-P binding pocket is unique to AsfvPolX. Supported by its strong electron density, the 5′-P binding pocket is well defined in the AsfvPolX:1nt-gap(P) DNA5:dGTP (Fig 5B) and the AsfvPolX:1nt-gap(P) DNA6:dGTP structures. However, superposition of all our structures showed that the 5′-P binding pocket can undergo large conformational changes when 5′-P is absent. For example, when compared with the AsfvPolX:1nt-gap(P) DNA5:dGTP structure, the guanidyl group of Arg125 is rotated approximately 180° along the CD—NE bond in the AsfvPolX:DNA1 structure (Fig 5C); although Arg125 and the C-terminal carbonyl group still interact with each other, they were both shifted away from the loop (where Thr166 and Arg168 reside). Together, these results indicate that the 5′-P binding pocket is not preformed, and its formation may follow an induced-fit mechanism.
To verify the biological relevance of the 5′-P binding mode observed in the structures, we constructed three AsfvPolX mutants (R125A, R168A, and R125/168A) and carried out isothermal titration calorimetry (ITC) and in vitro catalytic assay using a gap(P) DNA, DNA G31 (Fig 5D). The ITC results (Fig 5E, Table 2) showed that the DNA G31 binding affinity of the WT AsfvPolX are stronger than those of the R125A and R168A mutants; the dissociation values (Kd) are 1.64 μM, 6.21 μM, and 16.02 μM for the WT AsfvPolX, R125A, and R168A, respectively. No detectable DNA G31 binding affinity was observed for the R125/168A mutant. Consistent with the DNA binding affinities, the dG:dGTP misincorporation activities of the WT AsfvPolX are also much stronger than the three mutant proteins (Fig 5F and S4 Fig). After the 30-min reaction, there are 92% dG incorporation products generated for the WT AsfvPolX. Compared with the WT protein, the activities of the R125A and R168A mutants were lowered more than 2- and 10- fold, respectively; after the 30-min reaction, there are only 42% and 8% products formed for the R125A and R168A mutants. The activity of the double mutant (R125/168A) was even lower; it only generated about 3% product after 30 mins.
In addition to DNA G31, we also carried out the ITC (Table 2) and in vitro catalytic assays using two DNA molecules without 5′-P (S5A Fig): one 1nt-gap DNA (DNA G31a, which is identical to DNA G31 in sequence) and one 2nt-recessed DNA (DNA R2). As depicted in S5B and S5C Fig, both DNA G31a and DNA R2 have no detectable binding with the AsfvPolX proteins, including the WT AsfvPolX, R125A, R168A, and R125/168A mutants. Due to the weak binding, the dGTP misincorporation against both DNA G31a and DNA R2 is very slow, and mutation of Arg125 and Arg168 had no significant impact on the dGTP misincorporation activity of AsfvPolX. Compared to DNA G31a, the dGTP misincorporation rate against DNA R2 is slightly higher; in the presence of WT AsfvPolX, there were 26% and 40% products formed for DNA G31a (S6A and S6B Fig) and DNA R2 (S6C and S6D Fig), respectively, after 4 hr reaction. The AsfvPolX:1nt-gap DNA4 structure (Fig 3A) may provide one plausible explanation for this phenomena, i.e., besides the gap site, AsfvPolX can also bind to DNA G31a at the blunt end, which will inhibit the reaction.
Interestingly, in addition to the dG:dGTP misincorporation product band, one more newly formed DNA band was also simultaneously observed on the gel after the in vitro catalytic assay using DNA R2 (S6C Fig). According to the distances between the bands, the slower-moving band corresponds to the product having two dGTP incorporated; the second dGTP should be directed by the template dC at the 5′-end of DNA R2. The intensities of the two product bands are comparable to each other, suggesting that AsfvPolX itself can efficiently bypass dG:dG lesion. However, the detailed mechanism of this lesion bypass is unclear. Similar to dG:dGTP misincorporation, the dG:dG lesion bypass activity of AsfvPolX might play a role during the strategic mutagenesis of the ASFV genome. With longer reaction times (such as 3 and 4 hr), some very slow-moving bands are also observed on the gel, suggesting that AsfvPolX may have terminal transferase activity.
Both WT and mutant AsfvPolX proteins can efficiently catalyze the Watson—Crick paired dCTP incorporation against DNA G31a (S7A and S7B Fig) or DNA R2 (S7C and S7D Fig). Unlike the dGTP misincorporation against DNA G31 (Fig 5F and S4 Fig), the dCTP incorporations against DNA G31a and DNA R2 was not sensitive to the mutations on the 5′-P binding pocket; after 30 min reaction, there are more than 98% products formed for all the AsfvPolX proteins, including the WT AsfvPolX, R125A, R168A, and R125/168 mutants. All together, these observations suggested that the 5′-P and its recognition by AsfvPolX play a more critical role in the dG:dGTP misincorporation than the Watson—Crick paired incorporation.
Under our reaction conditions, the reaction rate of dG:dGTP misincorporation against DNA G31 is slower than that of dG:dCTP incorporation; however, previous studies demonstrated that the dG:dGTP misincorporation rate might be as fast as dG:dCTP incorporation under certain conditions [26]. One dGTP was captured at the active sites of both the AsfvPolX:1nt-gap(P) DNA5:dGTP and AsfvPolX:1nt-gap(P) DNA6:dGTP structures. In the former structure, the dGTP is in anti-conformation and forms Watson—Crick base pairing with the template dC (Fig 6A), whereas, in the latter structure, the dGTP adopts syn-conformation and forms Hoogsteen interactions with the template dG (Fig 6B), which is consistent with the AsfvPolX:DNA:dGTP NMR structure [17].
In previous studies, it was suggested that His115 played the most critical role in dG:dGTP misincorporation. In the AsfvPolX:1nt-gap(P) DNA6:dGTP structure (Fig 6B), His115 forms one interaction with the incoming dGTP, the hydrophobic interaction (3.4 Å) between its CE1 atom and the C8 atom of dGTP. Unexpectedly, His115 (via its NE1 atom) forms one H-bond (3.0 Å) with the NE2 atom of Arg127, and this interaction is conserved in all our WT AsfvPolX structures. To assess the impacts of His115 and Arg127 on the dG:dGTP incorporation, an in vitro catalytic assay using DNA G31 and five AsfvPolX mutants (H115D, H115E, H115F, R127A, and H115F/R127A) was carried out (Fig 6C and S8 Fig). Compared with the WT AsfvPolX, the dG:dGTP misincorporation activities of both H115D and H115E mutants were lowered more than 18- and 36- fold, respectively; after 30-min reaction, there are only 5% and 2.5% products formed for the H115D and H115E mutants, respectively. Asp115 and Glu115 may be able to form salt bridge with Arg127 and hold it in the conformation similar to the one in the WT AsfvPolX structures; however, the lower catalytic activities of H115D and H115E suggested that Asp115 and Glu115 could not mimic His115 in interacting with the dG:dGTP pairs, possibly because of their negative charges and higher hydrophilicity that are incompatible with the nucleobase of dGTP. The dG:dGTP misincorporation catalyzed by H115F was also very slow, with 8% product bands observed on the gel after the 30-min reaction. In contrast to H115F, R127A mutant can support the dG:dGTP misincorporation; although it is not as efficient as the WT AsfvPolX, R127A created 29% product after the 30-min reaction. Noticeably, the double mutation of His115 and Arg127 does not further reduce the dG:dGTP misincorporation rate; in contrast, there are 52% products formed in the presence of the H115F/R127A mutant after 30-min reaction, suggesting that the dG:dGTP misincorporation activity of H115F/R127A is higher than those of the H115F and R127A mutants.
To further investigate these observations, we solved the structures of H115F/R127A:1nt-gap(P) DNA6:dGTP (Fig 6D) and H115F:1nt-gap(P) DNA6:dGTP (Fig 6E). Similar to the AsfvPolX:1nt-gap(P) DNA6:dGTP structure (Fig 6B), the dGTPs in the two mutant structures adopt syn-conformations and form Hoogsteen interactions with the template dGs, with the overall conformations of the dGTPs in the three structures being very similar. The orientations of His115 in the AsfvPolX:1nt-gap(P) DNA6:dGTP structure and Phe115 in the H115F:1nt-gap(P) DNA6:dGTP structure are also similar (Fig 6F), whereas, compared to the WT AsfvPolX structure, the side chain of Arg127 in the H115F mutant structure rotates approximately 90° around the CG—CD bond and forms two H-bonds: one (3.1 Å) is between the NE atom of Arg127 and the backbone O atom of Leu137, and the other (2.7 Å) is between the NH2 atom of Arg127 and the O4′ atom of dT10. The dT10 pairs with ddA9′ at the post-insertion n-1′ site in all our PolX:1nt-gap(P) DNA6:dGTP structures. The relative orientations of the dT:ddA pairs are similar in the H115F and H115F/R127A structures; however, when compared with the WT AsfvPolX structures, both nucleobases of dT10 and ddA9′ in the H115F structure shifted up approximately 2 Å (Fig 6F).
Arg127 is highly conserved in the PolX family (S1 Fig), whereas His115 can be replaced by other aromatic residues in the homologous proteins, such as Tyr in TtPolX, RatPolβ, and HsPolβ, which are less efficient in catalyzing dG:dGTP misincorporation. A previous study showed that replacing His115 with Tyr115 could not maintain the dG:dGTP misincorporation activity of AsfvPolX; instead, it will completely prevent the complex formation between AsfvPolX and dG:dGTP mispair containing DNA molecules. Although it needs to be further verified, structural comparison (S9 Fig) suggested that two neighboring Phe residues (Phe102 and Phe116) may play a certain role during this process. In the homologous protein structures, the corresponding residues (which are Arg245 and Leu259 in TtPolX and Arg258 and Phe272 in HsPolβ) do not interact with each other, whereas Phe102 and Phe116 form stable stacking interaction and packs against the side chain of His115 in the AsfvPolX structures. Based on all these observations, we concluded that both His115 and Arg127 are important for dG:dGTP misincorporation. His115 and Arg127 form a platform, the His115–Arg127 platform, which can stabilize both the dG:dGTP base pair (at the insertion site) and, more importantly, the base pair at the postinsertion site from underneath. When the platform is broken in the mutant structures, the postinsertion site base pairs shift away. The interactions between Arg127 and Leu137 (and dT10) in the H115F mutant interfere with the dT:dA base pair rearrangement (to the catalytic conformation), which may cause the low dG:dGTP misincorporation rate.
AsfvPolX is a highly distributive DNA polymerase, and it follows an ordered Bi Bi mechanism [17–19]. The first substrate of AsfvPolX is dNTP, which can form a complex with AsfvPolX in the absence of DNA. Although we failed to determine any AsfvPolX:dNTP binary complex structure in the present study, our ternary structures can shed some light on the dNTP binding. In the structures, the triphosphate groups of the incoming dGTPs coordinate with the cations located at the catalytic site (S10A and S10B Fig). In addition, the triphosphate and 3′-OH groups of dGTP interact with Ser39, Arg42, and Asn48 of AsfvPolX (S10C Fig). These interactions are common for all four dNTPs (dGTP, dATP, dCTP, and dTTP).
AsfvPolX is most error prone to dG:dGTP misincorporation, and it also has very strong dGTP preference in the absence of DNA. His115 forms hydrophobic interaction with dGTP in the AsfvPolX:1nt-gap (P) DNA6:dGTP structure (Fig 6B). However, this interaction is not unique; it also forms between His115 and dC, dG, and dT in the AsfvPolX:DNA1 (S10D Fig), AsfvPolX:DNA2 (S10E Fig), and AsfvPolX:DNA3 (S10F Fig) structures, respectively. We further analyzed all our structures to study this strong dGTP preference and found some interactions that are unique for the dGTP (or dG) in syn-conformations. In the AsfvPolX:1nt-gap(P) DNA6:dGTP structure, the side chain of Val120 forms extensive hydrophobic interactions with the dG (Fig 7A). The CB2 atom of Val120 points to the center of the six-member ring of dG, and the distances between the CB2 atom and the six atoms (N1, C2, N3, C3, C4, and C6) of the ring system are all within the range of 3.4–3.6 Å, suggesting that these interactions are very stable. Similar interactions were also observed in the AsfvPolX:DNA2 structure. In the AsfvPolX:1nt-gap (P) DNA5:dGTP structure, the dGTP adopts an anti-conformation; instead of the six-member ring, the five-member ring of dG was placed next to Val120, but it only forms two hydrophobic interactions (around 3.4 Å) with the CB2 atom of Val120 (Fig 7B). In the AsfvPolX:1nt-gap(P) DNA6:dGTP and AsfvPolX:DNA2 structures, one hydrophobic interaction (3.5 Å) was also observed between the C8 atom of dGTP and the CD1 atom of Leu123, which forms one additional interaction (3.3 Å) with the backbone O atom of His115 (Fig 7C). Both Val120 and Leu123 are hydrophobic in nature, and they are not conserved in other PolX family proteins (S1 Fig).
To study their potential impacts on dNTP selection and dG:dGTP misincorporation, we constructed two mutants, V120A and L123A, and carried out ITC and in vitro catalytic assays. ITC analysis (Fig 7D) showed that V120A mutation can significantly reduce the dGTP binding affinity; the dissociation values (Kd) of V120A mutant and the WT AsfvPolX are 4.65 μM and 0.37 μM, respectively. L123A mutation also lowered the dGTP binding affinity, but the Kd value (1.70 μM) is lower than that of V120A, indicating that Val120 is more important for dGTP binding. In vitro catalytic assay results (Fig 7E and S11 Fig) suggested that these two residues are important for the dG:dGTP misincorporation activity of AsfvPolX. Compared with the WT AsfvPolX, the dG:dGTP misincorporation activities were lowered 1.3- and 3.4-fold for the L123A and V120A mutants, respectively; after 30-min reaction time, there are 69% and 27% products formed in the presence of the V120A and L123A mutants, respectively. These results also suggested that Val120 residue is more important for dG:dGTP misincorporation than the Leu123 residue. As summarized in Table 3, mutations of Val120 and Leu123 have little impact on the binding of dCTP and dTTP, but they can cause obvious reduction on dATP binding, which is similar to dGTP. However, compared to the dATP misincorporation, AsfvPolX is more effective in dGTP misincorporation; the aforementioned factors, such as the Hoogsteen pairing with template dG and the stabilization by the His115–Arg127 platform, should play important roles in this selection.
ASFV is contagious and can cause lethal diseases in domestic pigs and wild boars. AsfvPolX is the most unique DNA polymerase identified to date; it catalyzes the gap-filling reaction on the ASFV genomic DNA during the BER process. The sequence similarity between AsfvPolX and the homologous proteins is very low, and, as revealed by our crystal structures of AsfvPolX in complexes with various DNA molecules, AsfvPolX has a unique primary stem binding mode and several structure features, including a 5′-P binding pocket, a His115-Arg127 platform, and hydrophobic residues, which are unique to AsfvPolX. These unique structural features are involved in downstream oligo 5′-P recognition, dG:dGTP mispair stabilization, and dGTP stabilization, respectively. In combination with ITC analysis, mutagenesis, and in vitro catalytic assays, our studies further showed that these structural features are all important for the dG:dGTP misincorporation activity of AsfvPolX, the most frequent misincorporation catalyzed by AsfvPolX.
The ASFV genome is replicated and assembled in an oxidative environment, which can cause continuous damage to the virus genome. Although the fidelity is low, AsfvPolX is the sole DNA repair polymerase involved in the BER process; therefore, inhibiting the catalytic activity of AsfvPolX will disrupt the repair process of the virus genome. Compared with other gap-filling DNA polymerases, the most unique feature of AsfvPolX is the 5′-P binding pocket located in the finger domain. As observed in several of our structures, negatively charged ions (such as the SO42− ion present in the crystallization buffer) can bind at the 5′-P binding pocket. These observations can help facilitate future rational drug design targeting the 5′-P binding pocket.
Clearly, preventing dNTP binding by non-reactive dNTP analogs (especially dGTP and dATP analogs) is another way to block the BER process of ASFV, as has been proposed in previous NMR studies. Interestingly, the dNTP and 5′-P binding sites are only approximately 15 Å away from each other; therefore, they provide great opportunities for small molecules to prevent the simultaneous binding of dNTP and 5′-P, which should have better inhibitory effect and higher specificity.
The gene (S2 Table) containing the codon-optimized cDNA of full-length WT AsfvPolX was purchased from Shanghai Generay Biotech Co., Ltd, China. The gene was cleaved with BamHI and XhoI and resolved on agarose gel. The target fragment was recovered and recombined into the pET28-Sumo vector treated with BamHI and XhoI. The recombinant vector (coding the His-Sumo-AsfvPolX) was then transferred into the Escherichia coli BL21 DE3 competent cell. The plasmid DNA was extracted according to standard Miniprep protocols, and the sequence of the plasmid was confirmed by DNA sequencing.
The plasmid DNA of the L52/163M mutant was constructed using a site direct mutagenesis kit according to the manufacturer’s protocols, with the recombinant vector coding the WT His-Sumo-AsfvPolX used during this process. The His-Sumo-AsfvPolX plasmid DNA was also used as the template for the polymerase chain reactions (PCR) or overlap PCR during the preparation of all other AsfvPolX mutant constructs, including H115D, H115E, H115F, H115F/R127A, V120A, L123A, R125A, R125/168A, R127A, and R168A. The detailed sequences of the primers are listed in S2 Table. Other procedures, such as double digestion, DNA ligation, and transformation, are similar to those utilized during the WT AsfvPolX DNA construction. Sequences of all mutant plasmids were confirmed by DNA sequencing. All the recombinant strains were protected by 30% glycerol and stored in a −80°C freezer until use.
The frozen recombinant strains were revived in Lysogeny broth (LB) medium supplemented with 50 μg/mL kanamycin at 37°C overnight. Every 25-mL revived bacterium suspension was inoculated into 1 L LB medium supplemented with kanamycin (50 μg/mL) and cultured at 37°C with continuous shaking (225 rpm). The protein expression was induced at OD600≈0.6 by the addition of isopropyl β-D-1-thiogalacto-pyranoside (IPTG), with a final concentration of 0.2 mM. The induced cultures were then grown at 18°C for an additional 18 hr. The cells were harvested by centrifugation, and the pellets were resuspended in phosphate-buffered saline (PBS; 137 mM NaCl, 2.7 mM KCl, 10 mM Na2HPO4, and 2 mM KH2PO4). The suspension was centrifuged again and the pellets were stored in a −20°C freezer.
For the overproduction of the Se-Met substituted L52/163M AsfvPolX mutant, the revived recombinant strains from 50 mL overnight cultures were inoculated into 2 L LB medium supplemented with kanamycin (50 μg/mL) and grown at 37°C. When OD600 reached 0.4, the cells were harvested by centrifugation and resuspended in 100 mL M9 medium (47.7 mM Na2HPO4, 22 mM KH2PO4, 8.6 mM NaCl, and 28.2 mM NH4Cl). The resuspended cells were centrifuged again and transferred into 2 L of fresh M9 medium supplemented with 50 μg/mL kanamycin and 40 mg/L Se-Met (J & K). Following growth of the cultures at 37°C for 1 hr, the temperature was lowered to 18°C. Protein expression was induced by addition of IPTG with a final concentration of 0.1 mM. The induced cultures were then grown at 18°C for an additional 18 hr and the cells were harvested by centrifugation.
The cell pellets were resuspended in Ni binding buffer (Buffer A, 20 mM Tris pH 8.0, 500 mM NaCl, and 25 mM Imidazole pH 8.0) and lysed under high pressure via a JN-02C cell crusher. The homogenate was clarified by centrifugation (17,000 rpm) at 4°C for 1 hr, and the supernatant was loaded onto a Ni-NTA column (GE healthcare) equilibrated with Buffer A. The His-Sumo-AsfvPolX fusion protein was eluted from the column using elution buffer (Buffer B, 20 mM Tris pH 8.0, 500 mM NaCl, and 500 mM Imidazole pH 8.0) with a gradient. The fractions containing the desired fusion proteins were pooled and dialyzed against Buffer S (20 mM Tris pH 8.0, 500 mM NaCl, and 25 mM Imidazole pH 8.0) at 4°C for 3 hr; Ulp1 protease was also added to the sample during the dialysis process. The sample was again loaded onto a Ni-NTA column; the flow through containing the target AsfvPolX was collected and diluted with Tris buffer (20 mM, pH 8.0) to lower the NaCl concentration (the final concentration of NaCl was less than 150 mM). The diluted sample was loaded onto a HiTrap SP HP column (GE Healthcare), equilibrated with S binding buffer (20 mM Tris pH 8.0 and 100 mM NaCl), and eluted using S Elution Buffer (20 mM Tris pH 8.0 and 1 M NaCl) with a continuous gradient. The fractions containing the target protein were concentrated and loaded onto a Hi 16/60 Superdex G75 column (GE Healthcare) and equilibrated with Gel Filtration Buffer (20 mM Tris pH 8.0 and 500 mM NaCl). The purity of the proteins was analyzed by a SDS-PAGE gel. The protein was concentrated and snap-frozen using liquid nitrogen and stored at −80°C until use. To prevent the intermolecular S-S bond formation, 1mM DTT was present in all buffers. All the mutant proteins were purified using the same procedures.
All ITC experiments were performed on an ITC200 calorimeter (Microcal Inc.). The heat evolved following each titration point was obtained from the integral of the signal, and the data were analyzed using Microcal Origin software.
All DNA molecules utilized in this work were purchased from Shanghai Generay Biotech Co., Ltd. DNA G31, DNA G31a, and DNA R2 were utilized in the in vitro catalytic assay. DNA G31 and DNA G31a were assembled by mixing the template strand, primer strand, and downstream oligo in a molar ratio of 1:1:1 in Tris buffer (Buffer C, 20 mM, pH 8.0). DNA R2 was formed by a self-complementary DNA, which was also dissolved in Buffer C. The concentrations were 8 μM for all three DNA samples. Protein samples, including the WT AsfvPolX and all mutants, were diluted using Gel Filtration Buffer. A 10-μL reaction system (containing 3 μL Gel Filtration Buffer, 2 μL Buffer C, 1 μL 100 mM MgCl2, 1 μL 10 mM dCTP [or dGTP], 1 μL 8 μM DNA, and 2 μL protein) was established. The final protein concentrations are 0.2 μM and 1.6 μM for the Watson—Crick paired dCTP incorporation and the dG:dGTP misincorporation, respectively. The reactions were carried out at 37°C and quenched by the addition of 10 μL termination buffer (90% formamide, 20 mM EDTA, 0.05% bromophenol blue, and 0.05% xylene blue) at various time points indicated on the Figures. Each reaction was repeated for at least three times. Samples of 3 μL were loaded onto prewarmed 18% urea sequencing gels and run at 50–55 W and 48–50°C for 90 min. The gel was imaged using Typhoon FLA 9000, and the intensities of the substrate and product bands were quantified by ImageQuantTL and analyzed by GraphPad Prism programs.
All DNA molecules utilized in the structural studies were dissolved in ddH2O without annealing; the detailed sequences of the DNA molecules are listed in Fig 1. The crystallization samples were prepared by mixing proteins DNA, MnCl2, and dNTP (if present) at room temperature. The initial crystallization conditions for all crystals were identified at 18°C using the Gryphon crystallization robot system from the Art Robbins Instrument company and crystallization kits from the Hampton Research company. During the initial screening, the sitting-drop vapor diffusion method with the 3-drop Intelli-Plates was utilized, whereas, all the crystal optimization procedures were performed at 18°C using the hanging-drop vapor diffusion method. The compositions of the final crystallization conditions are listed in S1 Table.
All the crystals were cryoprotected using their mother liquor supplemented with 25% glycerol and snap-frozen in liquid nitrogen. The X-ray diffraction data were collected on beamline BL17U and BL19U at the Shanghai Synchrotron Radiation Facility (SSRF) at cryogenic temperatures and maintained with a cryogenic system. One single crystal was used for all structures; data processing was carried out using the iMosflm program [27,28] embedded in the CCP4i suite [29] or the HKL2000 or HKL3000 programs [30]. The data collection and processing statistics are summarized in Table 1.
The structure of Se-L52/163M:1nt-gap DNA4 was solved using the SAD method [31] with the AutoSol program [32] embedded in the Phenix suite [33]; the Figure of Merit (FOM) value was 0.36. The initial model (that covers approximately 75% of protein residues in the asymmetric unit) was built using the Autobuild program. The model was then refined against the diffraction data using the Refmac5 program [34] of ccp4i, which revealed the detailed orientations of the missing protein residues and 1nt-gap DNA4. During refinement, 5% of randomly selected data were set aside to use in free R-factor cross validation calculations. The 2Fo-Fc and Fo-Fc electron density maps were regularly calculated and used as guides for the building of the missing amino acids, DNA, and solvent molecules using COOT. All the other structures were solved using the MR method with the Phaser program of CCP4i suite. The Se-L52/163M:gap DNA4 structure (with the DNA and water molecules omitted) was used as the search mode. DNA molecules, Mn2+ ions, water, and other molecules were all built manually using COOT [35]. The structures of H115F/R127A:1nt-gap(P) DNA6:dGTP and H115F:1nt-gap(P) DNA6:dGTP were refined using the phenix.refine program [36] of Phenix; all other structures were refined using the Refmac5 program of CCP4i. The structural refinement statistics are also summarized in Table 1. Structural factors and coordinates have been deposited in the Protein Data Bank under accession codes 5HR9, 5HRB, 5HRD, 5HRE, 5HRH, 5HRI, 5HRK, and 5HRL.
|
10.1371/journal.ppat.1002092 | The Lipid Transfer Protein CERT Interacts with the Chlamydia Inclusion Protein IncD and Participates to ER-Chlamydia Inclusion Membrane Contact Sites | Bacterial pathogens that reside in membrane bound compartment manipulate the host cell machinery to establish and maintain their intracellular niche. The hijacking of inter-organelle vesicular trafficking through the targeting of small GTPases or SNARE proteins has been well established. Here, we show that intracellular pathogens also establish direct membrane contact sites with organelles and exploit non-vesicular transport machinery. We identified the ER-to-Golgi ceramide transfer protein CERT as a host cell factor specifically recruited to the inclusion, a membrane-bound compartment harboring the obligate intracellular pathogen Chlamydia trachomatis. We further showed that CERT recruitment to the inclusion correlated with the recruitment of VAPA/B-positive tubules in close proximity of the inclusion membrane, suggesting that ER-Inclusion membrane contact sites are formed upon C. trachomatis infection. Moreover, we identified the C. trachomatis effector protein IncD as a specific binding partner for CERT. Finally we showed that depletion of either CERT or the VAP proteins impaired bacterial development. We propose that the presence of IncD, CERT, VAPA/B, and potentially additional host and/or bacterial factors, at points of contact between the ER and the inclusion membrane provides a specialized metabolic and/or signaling microenvironment favorable to bacterial development.
| The obligate intracellular bacterial pathogen Chlamydia has developed strategies to invade, survive and replicate within the host genital, ocular and pulmonary epithelial surfaces. Chlamydia developmental cycle occurs in a membrane bound vacuole, the inclusion. The Chlamydia-dependent remodeling of the inclusion membrane leads to a unique and specialized compartment that allows for the specific interaction with cellular organelles and the acquisition of molecules and nutrients required for bacterial survival and replication. This study provides an example of how the insertion of C. trachomatis proteins into the inclusion membrane may allow the bacteria to manipulate the host cells to its own benefit. We showed that the lipid transfer protein CERT localized to C. trachomatis inclusion membrane and partly co-localized with the endoplasmic reticulum (ER) protein VAPB. Moreover, VAPB positive tubules made close contact with the inclusion membrane, suggesting the formation of ER-Inclusion membrane contact sites upon C. trachomatis infection. Finally, we have shown that CERT interacted with the C. trachomatis inclusion protein IncD. We propose that the presence of CERT, VAPB and IncD at ER-Inclusion membrane contact sites allow C. trachomatis to exploit the non-vesicular lipid transport machinery of the host cell and generate platforms specialized in metabolism and signaling events favorable to bacterial development.
| Chlamydia species are obligate intracellular Gram-negative bacterial pathogens that infect genital, ocular and pulmonary epithelial surfaces. Chlamydia are characterized by a biphasic developmental cycle that occurs exclusively in the host cell. The bacteria alternate between an infectious, metabolically inactive form called elementary body (EB) that is characterized by a condensed nucleoid, and an intracellular, metabolically active form named reticulate body (RB). Once internalized, Chlamydia resides in a membrane bound compartment, named the inclusion. Shortly after uptake, an uncharacterized switch occurs leading to the differentiation of EBs into RBs. The RBs then start to replicate until the inclusion occupies a large part of the cytosol of the host cell. Midway through, the developmental cycle becomes asynchronous and RBs start to differentiate back into EBs. At the end of the cycle, which last two to three days depending on the species, EBs are released from the host cell allowing infection of neighboring cells [1], [2].
To establish and maintain their intracellular niche, Chlamydia manipulate the host cellular machinery [3]. Once internalized [4]–[6], Chlamydia directs the trafficking of the nascent inclusion to a perinuclear localization via a mechanism involving microfilaments, microtubules and the motor protein dynein [7]. The inclusion is encased in a scaffold of host actin and intermediate filaments [8] and infection induces Golgi fragmentation and formation of Golgi ministacks that surround the inclusion [9]. ER tracks containing Chlamydia antigens have also been described in the vicinity of the inclusion membrane [10]. The inclusion does not interact with the endocytic pathway [7], [11], however it intercepts exocytic vesicles and lipids from the Golgi [12]. Resident protein and lipid constituents of multivesicular bodies are also delivered to Chlamydia inclusion [13], [14]. Some Rab GTPases are recruited to the inclusion membrane [15], and host lipid droplets are targeted to enhance intracellular survival and replication [16], [17].
The secretion of Chlamydia effectors proteins is thought to be important for a successful developmental cycle. To date, many Chlamydia effector proteins have been identified. Some, but not all, are type III secretion substrates and their function range from entry to establishing and maintaining the replicative vacuole [18], [19]. Some of these effectors are released into the host cell cytosol where they target cellular organelles or signaling pathways, while others act directly at the inclusion membrane. C. trachomatis encodes ∼50 putative inclusion membrane proteins, including the Inc proteins which are characterized by a large hydrophobic domain of 40 or more amino acids. So far the inclusion membrane localization of 22 of them has been confirmed [20] and only a few of them have established functions [18]. IncA is involved in the homotypic fusion of the C. trachomatis inclusions [21], [22]. IncB, CT101, CT222 and CT850 co-localized with activated Fyn and Src kinases in inclusion membrane microdomains. These microdomains also interact with the centrosomes and it has been proposed that these four inclusion proteins are involved in the interaction of the inclusion with the microtubule network [23]. CT229 and CT813 have been assigned putative functions in intercepting the host vesicular trafficking, based on their respective interaction with the host proteins Rab4 [24] and VAMP7-8 [25].
Our knowledge of the cellular processes that are targeted by Chlamydia has greatly increased over the past 10 years, but we have only begun to identify the host and bacterial factors required for bacterial development. To better understand the molecular mechanisms underlying C. trachomatis infection, we recently conducted an RNAi screen and identified CERT as a host factor involved in C. trachomatis infection (I.D. and H.A., unpublished). In non-infected cells, CERT is proposed to be involved in the non-vesicular transfer of ceramide at ER-Golgi membrane contact sites (MCSs). Our results indicate that ER-Inclusion MCSs are formed in C. trachomatis infected cells. We propose a model in which, through IncD-dependent recruitment of CERT to the inclusion, C. trachomatis exploit the non-vesicular lipid transport machinery of the host cell and generate platforms specialized in metabolism and/or signaling events favorable to its replication and development.
We recently conducted an RNAi screen and identified CERT as a host factor involved in C. trachomatis infection (I.D. and H.A., unpublished). CERT is proposed to be a functional component of ER-Golgi membrane contact sites (MCSs) (i.e. zone of close apposition (10–50 nm) between two organelles [26], [27]) involved in the non-vesicular transfer of ceramide from the ER to the Golgi [28]. In addition to the carboxy-terminal START domain [29] that binds ceramide, the ER-to-Golgi transfer process requires a central FFAT domain [30] which binds the ER resident proteins VAPA and VAPB (Vesicle-associated membrane protein-associated protein) [31] and an amino-terminal PH domain [32] which binds PI4P and Arf1 on the Golgi membrane [33], [34].
To further investigate the role of CERT in C. trachomatis infection, we first determined its cellular localization upon infection. In uninfected cells, the endogenous CERT protein was detected at the Golgi (Supplementary Figure S1). In C. trachomatis infected cells, the endogenous CERT protein was highly recruited to C. trachomatis inclusion as early as 8 h post infection when incoming bacteria had reached the perinuclear area of the host cells (Figure 1A). As the infection progressed, the inclusion remained CERT positive (Figure 1B) and CERT appeared to localize to the inclusion membrane as shown by co-immuno-staining with the inclusion membrane protein, IncA [35], [36] (Figure 1B). Co-immuno-staining of C. trachomatis infected cells with antibodies against CERT and the Golgi markers p115, GM130 or TGN46, did not reveal any co-localization of the two proteins (Figure 1A and 1C and Supplementary Figure S1), showing that the CERT signal did not correspond to the Golgi ministacks surrounding the inclusion in infected cells [9]. However, the CERT signal detected at the inclusion membrane partly overlapped with the ER resident proteins VAPA (data not shown) and VAPB (Figure 1D). These results showed that the lipid transfer protein, CERT, and the ER resident proteins VAPA/B localize to C. trachomatis inclusion, suggesting a the close apposition of the ER and the inclusion membrane.
Localization of the ER resident proteins VAPA and VAPB at the inclusion membrane and the previous report of ER tracks containing Chlamydia antigens in the vicinity of the inclusion membrane [10] led us to further investigate the potential interaction between the ER and C. trachomatis inclusion at the ultra structural level.
Electron microscopy analysis of C. trachomatis infected cells revealed the presence of ER tubules in close proximity of the inclusion membrane (Figure 2). On the inclusion section depicted in Figure 2, the ER tubules covered ∼25% of the inclusion membrane (Figure 2A and 2B). The longest tubule was ∼2 µm long (Figure 2B, panel 2) and the shortest tubules were ∼250 nm (Figure 2B, panel 3). Although, the number (1–6) and the size of the tubules (∼250 nm to 2 µm) varied among the inclusion sections analyzed, all inclusions were to some extent associated with ER tubules, suggesting a substantial coverage of the inclusion membrane. The distance between the ER tubules and the inclusion membrane was ∼10 nm (Figure 2C, panels 5 and 6).
We next investigated the cellular localization of CERT and VAPB with regard to the inclusion membrane and/or the ER tubules. Cryo-electron microscopy analysis confirmed the presence of ER tubules in close proximity (<20 nm) of the inclusion membrane (Figure 3A and 3B). Immunogold labeling showed that CERT localized to the inclusion membrane and was enriched at points of contact with the ER tubules (Figure 3A). VAPB was not observed on the inclusion membrane, but located to the ER tubules observed in close proximity of the inclusion (Figure 3B).
Altogether these results are consistent with the notion that, similar to the ER-Golgi MCSs observed in non-infected cells [26], [27], ER-Inclusion MCSs are established upon C. trachomatis infection and the non-vesicular ceramide transfer protein CERT and the ER resident proteins VAPB localize to these points of contact.
We next determined the CERT domain required for its recruitment to the inclusion membrane. As observed with the endogenous CERT (Figure 1), a CERT-GFP fusion protein containing the full length CERT was recruited to C. trachomatis inclusion membrane (Figure 4A, FL-CERT and Supplementary Figure S2, FL-CERT CTRLsiRNA). Similarly, a construct containing the PH domain only was also recruited to C. trachomatis inclusion (Figure 4B, CERT-PH). However, a CERT construct lacking the PH domain was mostly cytosolic and, except for very small patches, was no longer recruited to the inclusion membrane (Figure 4C, CERTΔPH and Supplementary Figure S2 CERTΔPH CTRLsiRNA). It is likely that, in these experiments, ER-Inclusion MCSs were formed. Thus, remaining patches observed in Figure 4C (and Supplementary Figure S2) with the CERTΔPH-GFP construct probably reflected binding of the fusion protein to the endogenous VAPs through its FFAT domain. Accordingly, the number of patches was further decreased in VAPA/B-depleted cells (Supplementary Figure S2 CERTΔPH VAPA&BsiRNA). Altogether, these results indicate that the PH domain of CERT is necessary and sufficient for CERT association with C. trachomatis inclusion membrane.
Our results suggested that, at ER-Inclusion MCSs, CERT interacted with the ER tubules by binding to the VAPs, and with the inclusion membrane via its PH domain. The interacting partner(s) on the inclusion membrane remained however to be identified. CERT belongs to a family of lipid transfer proteins, containing OSBP and FAPP [33], [37]. These proteins are characterized by a common amino-terminal PH domain, which governs their specific association with Golgi membranes. For OSBP and FAPP, this specificity is conferred by the recognition and binding of the PH domain to PI4P and Arf1. Both determinants are simultaneously required since reduced level of PI4P onto Golgi membrane or Arf1 inactivation reduces the Golgi association of OSBP and FAPP PH domains [38], [39]. Association of CERT PH domain with Golgi membranes also depends on PI4P [28], [40]. By analogy with OSBP and FAPP, it has been proposed that Arf1 might also control CERT association with Golgi membranes [33].
Because Arf1 is present onto the inclusion membrane [41] (Figure 5, top panels, Arf1-GFP), we tested the functional importance of Arf1 with respect to CERT recruitment to the inclusion. As expected, Brefeldin A treatment abolished Arf1 interaction with the inclusion [41] (Figure 5, bottom panels, Arf1-GFP). However, it did not affect endogenous CERT recruitment to C. trachomatis inclusion (Figure 5, bottom panels, CERT). We observed a similar result in Arf1-depleted cells (not shown). These results suggested that the PH domain of CERT mediates its recruitment to C. trachomatis inclusion through an Arf1-independent mechanism.
To identify the determinants underlying the recruitment of CERT to the C. trachomatis inclusion, we performed immuno-precipitation experiments with extracts obtained from cells expressing a FLAG-tagged version of CERT. Comparative mass-spectrometry analyses of uninfected and infected samples led to the identification of the C. trachomatis effector protein IncD as a potential binding partner for CERT (see Methods for details and Supplementary Figure S3 for C. trachomatis replication in HEK293 cells).
IncD is the product of the first gene of an operon containing three other inclusion proteins (IncE, IncF and IncG) [42] and displays a large central hydrophobic domain [42], [43] (Figure 6A). The IncD-G operon is expressed within the first two hours of C. trachomatis developmental cycle and all four proteins were shown to localize to the inclusion membrane [42]. Both spatial and temporal expression of IncD made it an attractive candidate for CERT interaction.
We tested the specificity of CERT/IncD interaction by co-immuno-precipitation experiments. While a CERT-GFP construct (but not GFP alone) co-immuno-precipitated with the FLAG-tagged version of IncD (Figure 6B, IP, lanes 1–2), the FLAG-tagged versions of IncE and IncF, two Inc proteins that had a cellular localization similar to IncD when over-expressed in eukaryotic cells (i.e. the ER) (Supplementary Figure S4), did not (Figure 6C). We next confirmed that CERT/IncD interaction is mediated through the PH domain of CERT. As expected, a FLAG-tagged version of IncD co-immunoprecipitated with a CERT-PH-GFP fusion protein, but did not co-immunoprecipitate with a CERTΔPH-GFP construct lacking the PH domain (Figure 6B, IP, Lanes 3–4). We confirmed the specificity of IncD interaction with the PH domain of CERT by showing that the FLAG-tagged version of IncD failed to co-immunoprecipitate the phosphoinositide binding domain of FAPP1 [38], p40phox (PX) [44] or PLC∂ [45] (Figure 6D). Finally, we confirmed the direct interaction between IncD and the PH domain of CERT by in vitro binding assay, using CERT PH domain and IncD purified as MBP and GST fusion proteins, respectively. As shown in Figure 6E, the PH domain of CERT interacted strongly with GST-IncD fusion protein, but not with GST alone. Altogether, these experiments demonstrated that the C. trachomatis inclusion membrane protein IncD binds specifically to the PH domain of CERT, suggesting that IncD is involved in CERT recruitment to the C. trachomatis inclusion.
We next investigated IncD localization onto C. trachomatis inclusion membrane. When C. trachomatis infected cells, were fixed and permeabilized using paraformaldehyde and saponin, respectively, IncD displayed inclusion membrane localization (Figure 7A, top panels). However, not all inclusions were IncD positive, although they were all positive for IncA (Supplementary Figure S5), suggesting that the IncD antigen was not efficiently revealed. When the cells were fixed and permeabilized using methanol, all inclusions were positive for IncA and IncD (Figure 7A, bottom panels).
We next investigated whether IncD co-localized with CERT onto the inclusion membrane. When C. trachomatis infected cells were fixed and permeabilized using paraformaldehyde and saponin, respectively, both markers localized to the inclusion membrane (Figure 7B, top panels). As previously observed (Figure 1), the CERT signal was not homogeneous and appeared more intense in some areas of the inclusion membrane, some of which also appeared enriched in IncD (Figure 7B, top panels, arrowhead). When the cells were fixed with methanol (Figure 7B, bottom panels), CERT localization was patchier than in paraformaldehyde-fixed cells and all patches were positive for both CERT and IncD. Altogether, these results suggested that IncD and CERT localize to the inclusion membrane and may be enriched/stabilized at specific macrodomains of the inclusion membrane.
To demonstrate the physiological relevance of CERT and VAPA/B recruitment to the C. trachomatis inclusion, we depleted CERT using two independent pools of CERT siRNA duplexes and VAPA and VAPB using a combination of pools of VAPA and VAPB siRNA duplexes. Transfection of the siRNA duplexes resulted in efficient depletion of all proteins (see Supplementary Figure S6 for CERT, VAPA and VAPB knock-down efficacy at the mRNA and protein level and Supplementary Table S1 for siRNA duplex sequences). In addition, CERT was no longer detected on the surface of C. trachomatis inclusion in CERT depleted cells (Supplementary Figure S7). Importantly, these experiments revealed that C. trachomatis inclusions were smaller in CERT-depleted cells as compared to control siRNA-treated cells (Figure 8A and 8B). The overall reduction in inclusion size upon CERT depletion also correlated with a reduction in infectious progeny (Figure 8C). Similar to the situation observed with CERT, depletion of VAPA and VAPB led to the formation of smaller inclusions (Figure 8A and 8B) and the number of infectious progeny produced was also reduced (Figure 8C). These experiments confirmed a role for CERT and VAPA/B in C. trachomatis development.
We showed that the lipid transfer protein, CERT, was recruited to C. trachomatis inclusion (Figure 1 and Figure 4A). The CERT signal was not evenly distributed onto the inclusion and appeared more concentrated in some areas. Although CERT localizes to Golgi membranes in uninfected cells, the CERT signal observed upon infection did not correspond to the Golgi ministacks surrounding the inclusion in infected cells [9], but rather corresponded to the inclusion membrane. These data indicate that CERT is enriched in some macrodomains of the inclusion membrane.
Our results also showed that the ER resident proteins, VAPA and VAPB, were recruited to C. trachomatis inclusion in areas enriched in CERT (Figure 1). VAPB however did not localized to the inclusion membrane but rather to ER tubules that were in close proximity of the inclusion membrane (Figure 3), suggesting a potential interaction between the ER and the inclusion.
Many cellular organelles have been shown in close apposition to or interacting with the Chlamydia inclusion, including the Golgi [9], mitochondria [46], multivesicular bodies [13] and lipid droplets [17]. Tracks of ER have also been reported in the proximity of C. trachomatis inclusion [10]. Our electron microscopy analysis (Figure 2 and Figure 3) is in agreement with this latest observation, since it revealed the close apposition of ER tubules to C. trachomatis inclusion and therefore suggests a direct interaction between the ER and the inclusion through the formation of ER-Inclusion MCSs.
MCSs are defined as a zone of close apposition (10–50 nm) between two organelles. In eukaryotes, the ER has been described as one of the partnering organelle and has been shown to make close contact with the yeast vacuole, the plasma membrane, the Golgi complex, endosomes and lysosomes, and mitochondria. Except for the yeast ER/Vacuole MCSs [47], bridging complexes, also referred to as structural components, that bring the two partnering membranes in close apposition and stabilize of MCSs, have not been identified. In contrast, an increasing number of proteins known to function on two contacting organelles have been identified and are referred as functional components [26], [27], including CERT, which is proposed to be a functional component of ER-Golgi MCSs.
The distance between the ER tubules and C. trachomatis inclusion membrane was 10–20 nm (Figure 2 and Figure 3), which is in agreement with the notion of ER-Inclusion MCSs. We cannot exclude that, as described for the yeast ER/Vacuole MCSs [47], specific structural components are involved in the formation and maintenance of the ER-Inclusion MCSs. These factors could be of mammalian and/or bacterial origin. Our results however indicate that, CERT, a previously proposed functional component of ER-Golgi MCSs, is probably also a functional component of ER-Inclusion MCSs.
CERT belongs to a family of lipid transfer proteins, containing OSBP and FAPP [33], [37]. These proteins are characterized by a common amino-terminal PH domain, which governs their specific association with Golgi membranes by recognizing and binding to PI4P (OSBP, FAPP and CERT) and Arf1 (OSBP and CERT). By analogy with OSBP and FAPP, it has been proposed that Arf1 might also control CERT association with Golgi membranes [33]. Because Arf1 and PI4P are also present at the inclusion membrane [41], it made them likely components involved in CERT recruitment to C. trachomatis inclusion membrane.
Moorhead et. al. study showed that OSBP-PH and CERT-PH (referred as GPBP in this study) associated with Chlamydia inclusion membrane. However, the association was not sensitive to Brefeldin A, suggesting that it was Arf1-independent. In agreement, we showed that Brefeldin A or Arf1 depletion did not affect endogenous CERT association with the inclusion membrane (Figure 5), confirming that Arf1 is not involved in this process.
Moorhead et. al. showed that the association of OSBP-PH with the inclusion membrane was however sensitive to a mutation in OSBP-PH domain that abolished PI4P binding (but not Arf1 binding). This result led the author to conclude that the localization of PI4P-binding PH domains to the inclusion reflected the presence of PI4P at the inclusion membrane. We have attempted to confirm the involvement of PI4P in the endogenous CERT association with the inclusion by depleting enzyme involved in PI4P synthesis. PI4KIIIß has been identified as the main source of PI4P on Golgi membranes and therefore a key regulator of CERT, OSBP and FAPP association with Golgi membrane [33], [37]. Moorhead et. al. have proposed that OCRL1 and PI4KIIα might contribute to the pool of Chlamydia inclusion membrane PI4P [41]. Treatment with a specific PI4KIIIß inhibitor (PIK93) or depletion of PI4KIIIß, OCRL1 or PI4KIIα proteins had no effect on CERT recruitment to C. trachomatis inclusion (not shown). Altogether, these results left open the role of PI4P in CERT association with the C. trachomatis inclusion membrane.
We have identified the C. trachomatis inclusion membrane protein IncD as an in vivo binding partner of CERT and we showed that IncD interacted with the PH domain of CERT (Figure 6). This interaction was highly specific because IncD did not interact with other phosphoinositide-binding domain and CERT did not interact with other Inc proteins (i.e. IncE or IncF). Our results therefore suggested a role for IncD in the association of CERT with the inclusion membrane.
In agreement with this notion, both IncD and CERT localized to C. trachomatis inclusion membrane (Figure 7). Various methods of fixation suggested that IncD and CERT displayed partial co-localization into patches. In addition, CERT appeared enriched in some areas that were in contact with VAPB positive tubules (Figure 1). These data suggest that the enrichment of CERT and IncD onto macrodomains of the inclusion membrane may correspond to the point of contacts with the ER where IncD/CERT/VAPB form a complex through the binding of CERT to IncD onto the inclusion membrane and to VAPB onto the ER tubules. Further characterization of IncD/CERT/VAPB interaction will be required to identify the determinants that drive the specific formation of this complex.
Altogether, our results suggest that C. trachomatis has evolved strategies to efficiently hijack CERT and relocate it to the inclusion membrane. Unfortunately, because genetic tools are not available to manipulate Chlamydia, we were not able to assay whether a C. trachomatis IncD mutant is impaired in CERT localization to the inclusion and whether its developmental cycle is perturbed. We were however able to show that C. caviae, a strain of Chlamydia that infect guinea pigs and that is lacking IncD, was not able to recruit CERT-GFP to its inclusion membrane (Supplementary Figure S8), confirming the specific role of IncD in CERT recruitment to the inclusion.
Although our in vitro binding study demonstrated that IncD interact with the PH domain of CERT in the absence of additional factors (Figure 6), we cannot exclude the involvement of additional host or C. trachomatis factors in vivo, including PI4P. If PI4P plays a role in CERT association with the inclusion membrane, it would be interesting to determine whether the analogy could be made between PI4P/Arf1 on Golgi membranes and PI4P/IncD on C. trachomatis inclusion membrane. On the other hand, if PI4P is not required for CERT association with C. trachomatis inclusion membrane, it would suggest that IncD may mimic the Arf1/PI4P determinants that usually drive CERT recruitment to Golgi membranes. We also note that IncD could act in concert with some yet to be discovered host and/or bacterial factors that localize at the inclusion membrane. Further structure/function analyses of IncD will be required to address these questions.
Our results showed that CERT or VAPA/B depletion was detrimental to C. trachomatis development (Figure 8), but the specific role of these proteins remains to be determined. Previous studies have shown that sphingomyelin-containing vesicles traffic from the Golgi ministacks that surround the inclusion, to the inclusion, where sphingomyelin is incorporated into the inclusion membrane as well as the cell wall of the bacteria. It was also shown that acquisition of sphingomyelin is essential for Chlamydia development [9], [12], [48]–[50]. Since CERT is involved in the non-vesicular trafficking of the sphingomyelin precursor ceramide from the ER to the Golgi, an overall defect in sphingomyelin synthesis could explain the defect in C. trachomatis development observed in CERT-depleted cells.
Although we did not measure sphingomyelin synthesis in CERT siRNA-treated cells, studies from the Hanada group using the LY-A cell line (in which CERT is inactive due to a mutation in the PH domain that prevents CERT association with Golgi membranes) and fluorescent ceramide or [3H]sphingosine, revealed that the LY-A cell line showed at least a 50% reduction in sphingomyelin biosynthesis [28], [51], supporting the hypothesis that CERT siRNA-treated cells might have reduced level of sphingomyelin. These studies however also suggested the existence of a CERT-independent pathway for sphingomyelin synthesis that could account for up to 50% of sphingomyelin synthesis.
We did not observe any major disruption of the Golgi morphology in CERT- or VAPA/B-depleted cells (Supplementary Figure S9) and 60% of the inclusions had detectable Golgi ministacks around them in both control and CERT depleted cells, suggesting that the Golgi fragmentation necessary for optimal sphingomyelin acquisition was not impaired. When infected cells were labeled with fluorescent ceramide, which is subsequently metabolized into sphingomyelin in the Golgi, and incorporated into the inclusion [12], [48], [49], we observed the accumulation of fluorescent lipids at the Golgi and in C. trachomatis inclusion from control or CERT- or VAPA/B-depleted cells (Supplementary Figure S10, S11, S12 and S13). These results suggested that lipids were still trafficked to the inclusion in CERT siRNA-treated cells and it is most likely that, in the context of CERT siRNA-treated cells, the CERT-independent pathway of ceramide trafficking from the ER to the Golgi, explain the labeling of the Golgi. Labeling of the inclusion could then occur through vesicular trafficking of sphingomyelin from the Golgi as previously described [12], [48], [49].
If both CERT-dependent and –independent pathways participate to the accumulation of fluorescent lipids in the inclusion, one could expect a reduction in fluorescence in CERT siRNA-treated cells. The rapid bleaching of the fluorescent signal of BODIPY-C5-Ceramide, however did not allow us to accurately quantify the fluorescence intensity of the inclusions. Another limitation of this assay is that it does not determine the nature of the fluorescent lipid(s) that accumulate in the inclusion. Additional experiments involving lipid extraction and analysis will be required to unambiguously determine the nature of the fluorescent lipid(s) that we observed in the inclusion in CERT- VAPA/B siRNA-treated cells.
Altogether, it is possible that a partial overall reduction in host cell sphingomyelin explains the partial reduction in C. trachomatis growth observed in CERT siRNA treated cells, however further experiments, including the uncoupling of CERT-dependent and –independent pathways, are required to unequivocally address this question.
Given the localization of CERT to the inclusion, the described role of CERT in ceramide transfer and the presence of ER tubules (source of ceramide) in close apposition with the inclusion membrane, we favor the idea that CERT transfers ceramide from the ER to the inclusion membrane. Because of the requirement of Chlamydia for sphingomyelin for replication, ceramide could serve as a precursor for sphingomyelin synthesis directly at the inclusion membrane, assuming that a sphingomyelin synthase of host or bacterial origin is also present at the inclusion membrane. Alternately, it is possible that ceramide accumulates at the inclusion without being further modified and serves as a signaling molecule at the inclusion membrane. For example, it has been proposed that the lipid transfer/binding proteins (LT/BPs) Nir2 [52] and OSBP [53], which are respectively involved in the non-vesicular transfer of phosphatidylinositol/phosphatidylcholine (PI/PC) and sterols at ER-Golgi MCSs, act together with CERT to affect the lipid composition of the Golgi membranes and therefore influence the structural and functional identities of these membranes [54]. Similarly, the CERT-dependent transfer of ceramide at ER-Inclusion MCSs could influence the lipid composition of the inclusion membrane and generate specialized metabolic and/or signaling microenvironment favorable for bacterial development. Whether other LT/BPs, such as Nir2 or OSBP, or additional host factors localize to ER-Inclusion MCSs and participate in the formation of these specialized platforms remain to be determined. Finally, we cannot exclude that CERT has a yet to be discovered role at the inclusion membrane that does not involve ceramide transport.
Altogether our results suggest a model in which the C. trachomatis effector protein IncD specifically interacts with the non-vesicular ceramide transfer protein CERT, at MCSs between C. trachomatis inclusion membrane and ER tubules harboring the VAPA/B proteins (Figure 9). We speculate that the IncD-CERT-VAPA/B interaction may be involved in the non-vesicular transfer of ceramide from the ER to the inclusion. We however cannot exclude a more complex role for ER-Inclusion MCSs in supporting bacterial development. Further studies of the ER-Inclusion MCSs, including the identification of additional structural and functional components, may not only reveal the mechanisms underlying C. trachomatis pathogenesis, but may also illuminate the poorly understood cellular mechanisms underlying inter-organelle communication.
HeLa cells and HEK293 cells (ATCC) and HeLa229 cells (Dautry-Varsat Laboratory, Pasteur Institute, Paris, France) were cultured at 37°C with 5% CO2 in DMEM high glucose (Invitrogen) supplemented with 10% heat inactivated FBS (Invitrogen). C. trachomatis Lymphogranuloma venereum, Type II were obtained from ATCC (L2/434/Bu VR-902B). C. caviae was obtained from the Dautry-Varsat Laboratory (Pasteur Institute, Paris, France) and was originally from Roger Rank Laboratory (Little Rock, Arkansas). Chlamydia propagation and infection was performed as previously described [55].
The protocol used for siRNA transfection was adapted from Dharmacon HeLa cells transfection's protocol (www.dharmacon.com) [55]. The sequences of the siRNA duplexes used in this study are described in Supplementary Table S1. DNA transfection was performed using Fugene 6 according to the manufacturer recommendations. DNA transfection of siRNA treated cells was performed 2 days post siRNA transfection.
Total RNA and first-strand cDNA synthesis was performed using the TaqMan Gene Expression Cells-to-Ct Kit (Applied Biosystems) as recommended by the manufacturer with the addition of DNaseI for the removal of unwanted genomic DNA. mRNA levels were determined by quantitative real-time PCR using the LightCycler 480 Master Kit and LightCycler 480 instrument (Roche Biochemicals, Indianapolis, IN). The combination of probes and primers that were used to determine the relative amount of target mRNA by quantitative PCR is described in Supplementary Table S2.
The PH domain of CERT, VAPA and VAPB ORFs were amplified from HeLa cell cDNA. The ORF corresponding to full length CERT was amplified from a myc-CERT construct [56]. IncD, IncE and IncF ORFs were amplified from C. trachomatis genomic DNA. The primers and vectors used in this study are described in Supplementary Table S3. Arf1-GFP [57] and PX-YFP [44] were described previously.
Protein samples were separated by SDS-PAGE and analyzed by immunoblot using HRP-conjugated secondary antibodies and Amersham ECL western blotting detection reagents.
At the indicated time, the cells seeded onto glass coverslips were fixed for 30 min in PBS containing 4% paraformaldehyde. Immunostaining were performed at room temperature. Antibodies were diluted in PBS containing 0.16 µg/ml Hoechst (Molecular Probes), 0.1% BSA and 0.05% Saponin. Samples were washed with PBS containing 0.05% Saponin and a final PBS wash was performed before examination under an epifluorescence microscope. Samples presented in the lower panels of Figure 7 were fixed in ice-cold methanol for 5 min and labeled with antibodies diluted in PBS containing 0.1% BSA. Images presented in Supplementary Figure 2 were acquired using a confocal microscope.
The following primary antibodies were used: rabbit polyclonal anti-C. trachomatis (1∶300 (IF), Virostat), rabbit polyclonal anti-C. trachomatis IncA (1∶200 (IF), kindly provided by T. Hackstadt, Rocky Mountain Laboratories), rabbit polyclonal anti-C. caviae IncA (1∶300 (IF), kindly provided by A. Subtil, Pasteur Institute, Paris, France), mouse polyclonal anti-C. trachomatis IncD (1∶300 (IF), kindly provided by G. Zhong, University of Texas Health Science Center at San Antonio), chicken polyclonal anti-CERT (1∶200 (IF), 1∶1,000 (WB) Sigma), mouse anti-p115 (1∶200 (IF), kindly provided by G. Waters), mouse anti-GM130 (1∶200 (IF), BD Transduction Lab.), sheep anti-TGN46 (1∶200 (IF), Serotec), rabbit polyclonal anti-GFP (1∶200 (EM), 1∶2,000 (WB), Invitrogen), mouse anti-FLAG (1∶20,000 (WB), Sigma), rabbit polyclonal anti-VAPB (1∶200 (WB), Abcam), rabbit polyclonal anti-actin (1∶10,000 (WB), Sigma).
The following secondary antibodies were used: goat anti-rabbit AlexaFluor 594 antibody (1∶500, Molecular Probes), goat anti-mouse AlexaFluor 594 antibody (1∶500, Molecular Probes), donkey anti-sheep AlexaFluor 594 antibody (1∶500, Molecular Probes), Fluorescein (FITC) donkey anti-chicken IgY antibody (1∶300, Jackson ImmunoResearch), peroxidase-conjugated goat anti-rabbit IgG (1∶10,000, Jackson ImmunoResearch), peroxidase-conjugated goat anti-mouse IgG (1∶10,000, Jackson ImmunoResearch), peroxidase-conjugated donkey anti-chicken IgY (1∶10,000, Pierce).
HeLa cells seeded onto glass coverslips and infected with C. trachomatis for 24 h were fixed in 2.5% glutaraldehyde in 0.1 M sodium cacodylate buffer pH 7.4 with 2% sucrose for 1 hour. The samples were rinsed 3 times in sodium cacodylate buffer then were postfixed in 1% osmium tetroxide for 1 hour, en bloc stained in 2% uranyl acetate in maleate buffer pH 5.2 for a further hour then rinsed, dehydrated in an ethanol series and infiltrated with epon resin. The coverslips were then covered with resin filled capsules and baked over night at 60°C. Hardened block enface sections were cut using a Leica UltraCut UCT. 60 nm sections were collected and stained using 2% uranyl acetate and lead citrate. Samples were all viewed FEI Tencai Biotwin TEM at 80 Kv. Images were taken using Morada CCD and iTEM (Olympus) software.
HeLa cells were transfected with a GFP-CERT or GFP-VAPB construct for 18 h and infected with C. trachomatis for an additional 24 h. The cells were washed with PBS, fixed in PBS containing 4% Paraformaldehyde/0.1% Glutaraldehyde for 15 min at room temperature, followed by 45 min at 4°C in 4% paraformaldehyde and re-suspended in 10% gelatin. Trimmed smaller blocks were placed in 2.3 M sucrose overnight on a rotor at 4°C. Then transferred to aluminum pins and frozen rapidly in liquid nitrogen. The frozen block were trimmed on a Leica Cryo-EMUC6 UltraCut and 65 nm thick sections were collected using the Tokoyasu method [58], placed on a nickel formvar/carbon coated grid and floated in a dish of PBS ready for immunolabeling. For immunolabeling of the sections, the grids were placed section side down on drops of 0.1 M ammonium chloride to quench untreated aldehyde groups, then blocked for nonspecific binding on 1% fish skin gelatin in PBS. Grids were incubated on a primary antibody rabbit anti-GFP. The grids were then placed on protein A gold 10 nm (UtrechtUMC). All grids were rinsed in PBS between steps, lightly fixed using 1% glutaraldehyde, rinsed and transferred to a UA/methylcellulose drop, then collected and dried. Grids were viewed in FEI Tencai Biotwin TEM at 80 Kv. Images were taken using Morada CCD and iTEM (Olympus) software.
Lysates from HEK293 transfected with a 3xFLAG-CERT construct for 18 h and infected with C. trachomatis for an additional 24 h were immunoprecipitated using anti-FLAG M2 agarose beads. The bound proteins were separated by SDS-PAGE and analyzed by Silver Nitrate staining. A ∼15 kDa band present in the transfected/infected samples but not in the transfected or infected only samples was analyzed by LC MS-MS at the Yale W.M. Keck laboratory.
5.105 HEK293 cells plated in 6-well tissue culture dishes and transfected for 48 h were washed once with 1× PBS and lyzed for 30 min in 300 µl of lysis buffer (20 mM Tris pH 7.5, 150 mM NaCl, 2 mM EDTA, 1%Triton X-100, 1 mM PMSF and protease inhibitor cocktail (Roche)). The lysates were centrifuge at 13,000 rpm for 10 min. An aliquot of the clarified lysate was collected (Input). The clarified lysates were incubated for 2 h in the presence of 10 µl of anti-FLAG M2 agarose beads (Sigma). The beads were washed 3 times (20 mM Tris pH 7.5, 150 mM NaCl, 2 mM EDTA, 1%Triton X-100) and the bound proteins were eluted with 15 µl of elution buffer (20 mM Tris pH 7.5, 150 mM NaCl, 2 mM EDTA, 100 µg/ml 3XFLAG peptide (Sigma)). 10 µl of the eluted fraction was collected (IP). All steps were conducted at 4°C.
MBP and MBP-CERT PH were expressed in E. coli (BL2lDE3). The bacteria were resuspended in lysis buffer (20 mM Tris pH 7.5, 300 mM NaCl, 2 mM EDTA, 1 mM MgCl2, 1% Triton X-100, 1 mM DTT and 1 mM PMSF) and lysed by sonication. The clarified lysates were incubated for 2 h in the presence of amylose resin (NEB). The resin was washed 3 times with 20 mM Tris pH 7.5, 300 mM NaCl, 1 mM MgCl2. The bound proteins were eluted by incubation with 20 mM Tris pH 7.5, 100 mM NaCl, 1 mM MgCl2, 30 mM Gluthathione, dialysed over night in 20 mM Tris pH 7.5, 100 mM NaCl, 1 mM MgCl2 and stored in small aliquots at −20°C. GST and GST-IncD were prepared as described above and bound to glutathione sepharose beads (GE Healthcare). All steps were conducted at 4°C.
Equal amounts of freshly purified GST fusion proteins bound to glutathione beads were incubated with 0.2 µM of the indicated MBP fusion proteins in binding buffer (20 mM Tris pH 7.5, 150 mM NaCl, 1 mM MgCl2, 0.2% Triton X-100) for 2 h at 4°C, washed 3 times with 20 mM Tris pH 7.5, 300 mM NaCl, 1 mM MgCl2, 0.2% Triton X-100 and Laemli buffer, separated by SDS-PAGE and analyzed by immunoblot.
siRNA treated cells were infected with C. trachomatis for 36 hrs. The nuclei and the bacteria were labeled with the DNA dye Hoechst and the inclusions were counter-stained with a rabbit polyclonal antibody against C. trachomatis and a goat anti-rabbit AlexaFluor 594 antibody. The cells were subjected to automated fluorescence microscopy to capture images corresponding to the cell nuclei and the inclusion. Computer-assisted image analysis, using the analytical tools of the Metamorph software, was used to determine the number of nuclei and the surface area of each inclusion.
HeLa cells incubated with the indicated siRNA duplexes for 3 days were collected 48 h post infection, lysed with glass beads and dilutions of the lysate were used to infect fresh HeLa cells. The cells were fixed 24 h post infection and the number of inclusion forming units (IFUs) was determined after assessment of the number of infected cells by immonolabelling.
Control and siRNA treated C. trachomatis-infected HeLa cells seeded onto glass coverslips were washed three times with cold Hank's Balanced Salt Solution (HBSS, Invitrogen) and incubated in DMEM containing 2.5 µM DFBSA/BODIPYFL-C5-Ceramide (Invitrogen) and 5 µg/ml Hoechst for 30 min at 4°C. The cells were then washed three times with cold HBSS and incubated in DMEM containing 0.34% DFBSA (Calbiochem) at 37°C in the presence of 5% CO2. Pictures of the live cells were acquired in the FITC and DAPI channel with a 10× objective every 15 min after the beginning of the chase.
The accession numbers for C. trachomatis (L2/434/Bu) proteins are IncD (CTL0370): CAP03810.1, IncE (CTL0371): CAP03811.1 and IncF (CTL0372): CAP03812.1.
The accession numbers for the mammalian genes are CERT (isoform 2): NM_031361.2, VAPA (isoform 2): NM_194434.2, VAPB: NM_004738.4, ARF1: NM_001024227, PI4KIIIß: NM_002651.1, OCRL1: NM_000276.3 and PI4KIIα: NM_018425.2.
|
10.1371/journal.pntd.0001606 | Genetic and Anatomic Determinants of Enzootic Venezuelan Equine Encephalitis Virus Infection of Culex (Melanoconion) taeniopus | Venezuelan equine encephalitis (VEE) is a re-emerging, mosquito-borne viral disease with the potential to cause fatal encephalitis in both humans and equids. Recently, detection of endemic VEE caused by enzootic strains has escalated in Mexico, Peru, Bolivia, Colombia and Ecuador, emphasizing the importance of understanding the enzootic transmission cycle of the etiologic agent, VEE virus (VEEV). The majority of work examining the viral determinants of vector infection has been performed in the epizootic mosquito vector, Aedes (Ochlerotatus) taeniorhynchus. Based on the fundamental differences between the epizootic and enzootic cycles, we hypothesized that the virus-vector interaction of the enzootic cycle is fundamentally different from that of the epizootic model. We therefore examined the determinants for VEEV IE infection in the enzootic vector, Culex (Melanoconion) taeniopus, and determined the number and susceptibility of midgut epithelial cells initially infected and their distribution compared to the epizootic virus-vector interaction. Using chimeric viruses, we demonstrated that the determinants of infection for the enzootic vector are different than those observed for the epizootic vector. Similarly, we showed that, unlike A. taeniorhynchus infection with subtype IC VEEV, C. taeniopus does not have a limited subpopulation of midgut cells susceptible to subtype IE VEEV. These findings support the hypothesis that the enzootic VEEV relationship with C. taeniopus differs from the epizootic virus-vector interaction in that the determinants appear to be found in both the nonstructural and structural regions, and initial midgut infection is not limited to a small population of susceptible cells.
| Venezuelan equine encephalitis virus (VEEV) is transmitted to humans and horses by mosquitoes in Mexico, Central and South America. These infections can lead to fatal encephalitis in humans as well as horses, donkeys and mules, and there are no licensed vaccines or treatments available for humans. VEEV circulates in two distinct transmission cycles (epizootic and enzootic), which are differentiated by the ecological niche that each virus inhabits. Epizootic strains, those that cause major outbreaks in humans and equids, have been studied extensively and have been used primarily to develop and test several vaccine candidates. In this study, we demonstrate some important differences in the roles of different viral genes between enzootic/endemic versus epizootic VEEV strains that affect mosquito infection as well as differences in the way that enzootic VEEV more efficiently infects the mosquito initially. Our findings have important implications for designing vaccines and for understanding the evolution of VEEV-mosquito interactions.
| Venezuelan equine encephalitis virus (VEEV) has been recognized as an etiologic agent of neurologic disease in humans and equids for nearly 80 years. Closely related to eastern (EEEV) and western equine encephalitis viruses (WEEV), VEEV belongs to the family Togaviridae, genus Alphavirus. First recognized in the 1920s, Venezuelan equine encephalitis (VEE) outbreaks are typically episodic with several years elapsing between outbreaks. However, when outbreaks do occur, they can cause severe and sometimes fatal disease in hundreds-of-thousands of equids and humans. For instance, after an interval of 19 years with no documented cases between 1973 and 1992, clusters of cases emerged in Venezuela [1] and Chiapas, Mexico [2] prior to a major outbreak involving ca. 100,000 people in 1995 [3]. In general, disease manifestations of VEE range from flu-like illness to fatal encephalitis. It is estimated that central nervous system (CNS) involvement occurs in 4–14% of human cases, and children are at the greatest risk to develop encephalitis and to die from infection [4].
Of the four subtypes of VEEV, IC and IAB are considered epizootic as they are known to cause disease in horses, to use these hosts for amplification, and are also capable of utilizing a variety of epizootic mosquito vectors, such as Aedes (Ochlerotatus) taeniorhynchus, A. (Och.) sollicitans, Psorophora confinnis, Culex (Deinocerites) pseudes, Mansonia indubitans, and M. titillans, among others [5]–[10]. Many of these mosquitoes thrive near coastal brackish water, can fly long distances from larval sites, prefer to feed on humans or other large mammals, and can tolerate feeding in sunny areas, although they may rest in shaded sites. In contrast, enzootic VEEV subtypes IE and ID generally cause little or no viremia or disease in equids, but like the epizootic strains, can cause fatal disease in humans [11]–[14]. Mosquito vectors that maintain these enzootic viruses in nature include a variety of species within the Spissipes section of the subgenus Culex (Melanoconion), and subtype IE strains specifically utilize C. (Mel.) taeniopus. The enzootic cycle typically occurs in shaded, intact forests with stable pools of water that are available for larval development. Some larvae also require the presence of a specific aquatic plant (i.e., Pistia spp.) for respiration [15].
Recent identification of extensive endemic disease in Peru, Bolivia, Ecuador, Colombia and Mexico, caused by spillover of enzootic strains in subtypes ID and IE, indicates the importance of VEEV as a continuous public health threat in Central and South America [16], [17]. The recent documentation of widespread endemic disease is likely associated with increased surveillance as well as the clearing of sylvatic forest habitats to accommodate the expansion of agricultural land types in areas of Latin America where enzootic VEEV persists [18]–[20]. The resulting fragmentation of sylvatic habitats results in an increase in ecotones that can support the life cycle of enzootic VEEV mosquito vectors [21], which also increases the likelihood of an enzootic VEEV strain adapting to epizootic transmission [22]. Enzootic ID strains are known to be a source for the emergence of epizootic IC strains and this emergence has occurred on multiple occasions [1], [23]. While IE strains had not been associated with the emergence of epizootic strains before 1993, recent outbreaks of epizootic-like IE strains were found to infect epizootic mosquito vectors and cause disease in equids [2], [24].
Historically, IE VEEV strains have been found in isolated sylvatic transmission cycles between C. taeniopus mosquitoes and rodent hosts, such as cotton rats (Sigmodon spp.), spiny rats (Proechimys spp.) and other rodent species, including Liomys salvini and Oligoryzomys fulvescens [25]–[27]. Phylogenetic studies of IE strains show that they diverged from other subtype I VEEV viruses, including enzootic ID strains [28], indicating that IE strains have long been established and most likely isolated within their enzootic habitats for at least centuries. Examination of the low threshold for infection and specificity of IE strains for C. taeniopus vectors suggests that IE stains have co-adapted to be highly fit for replication in and transmission by this vector [29]–[31].
The stable, enzootic VEEV IE-C. taeniopus relationship is in sharp contrast to the transient interaction that occurs between epizootic virus strains and mosquito vectors during sporadic outbreaks. However, the majority of experimental studies examining VEEV-vector interactions have utilized epizootic vectors as models. We hypothesize that IE viruses are highly adapted to their enzootic vector through a long-term evolutionary relationship such that the dynamics of infection of IE viruses within their vector differ inherently from those observed in epizootic virus-vector interactions. Reverse genetic studies of epizootic IC VEEV indicate that infection determinants reside within the E2 glycoprotein gene [21], [24], [28], [32]. We hypothesized that the transient nature of the epizootic virus limits its infection determinants to a localized region of the genome to allow for rapid adaptation to a competent vector, whereas the enzootic infection determinants are not limited to a single region in the structural portions of the genome due to the long adaptation of the genome to infection and replication within C. taeniopus. To test this hypothesis, we generated four chimeric VEEVs (Fig. 1), using a strain with a known high susceptibility to C. taeniopus (i.e., subtype IE strain 68U201) and a strain known to be poorly infectious for C. taeniopus [i.e., subtype IAB Trinidad donkey (TrD) strain]. These chimeras allowed us to discern the contributions of the structural and nonstructural protein regions as well as the 3′ untranslated region (UTR) in infection and dissemination in C. taeniopus.
We also examined the initial midgut infection dynamics of the enzootic mosquito model as compared to what has been previously observed in the epizootic model with IC VEEV and A. taeniorhynchus. There is only a small population of VEEV-susceptible midgut cells in A. taeniorhynchus, and thus the midgut infection is initiated by a very small number of infected cells and presumably virions [32]. Evolutionary theory would suggest that a bottleneck in the population of replicating viral genomes might deleteriously affect viral fitness through Muller's ratchet [33]–[35]. However, epizootic strains might regain fitness through recombination [32]. While this is a plausible strategy for an epizootic virus, which only interacts transiently with its mosquito vector during an outbreak, the enzootic virus must maintain a certain level of fitness to persist in nature over centuries or longer and repeated bottlenecks would likely be highly detrimental. We therefore hypothesized that most or all midgut epithelial cells in C. taeniopus are susceptible and, therefore, the population of enzootic VEEV that infect the midgut epithelium does not undergo a severe bottleneck during the infection of the midgut. To examine this hypothesis, we utilized viral-like particles (VLP) to establish the number, distribution, and susceptibility of midgut epithelial cells initially infected in the IE enzootic model.
This study was performed in strict accordance with the recommendations in the Guide for the Care and Use of Laboratory Animals of the National Research Council. The protocol was approved by the Institutional Animal Care and Use Committee of the University of Texas Medical Branch (IACUC Protocol # 0209068, approved July 13, 2010).
Plaque, cytopathic effect (CPE) assays, and replication curves were performed on Vero (African green monkey kidney), and BHK-21 (baby hamster kidney) cells were used for electroporation to rescue parental and recombinant viruses as well as replicon particles from transcribed RNA. Both cell types were propagated in Dulbecco's modified eagle medium (DMEM) supplemented with fetal bovine serum (FBS) and penicillin/streptomycin. For CPE assays of mosquito samples, amphotericin B (50 µg/mL) (Sigma-Aldrich, St. Louis, MO) was added to the DMEM. Cells from an A. albopictus mosquito cell line, C6/36, maintained in DMEM media supplemented with 10% FBS, penicillin/streptomycin, and 1% tryptose phosphate broth (Sigma-Aldrich, St. Louis, MO) were utilized for in vitro replicon co-infection experiments and replication curves. Viruses used for this study were derived from infectious cDNA clones V3000 IAB Trinidad Donkey (TrD) (kindly provided by Nancy Davis and Robert Johnston) [36] and IE 68U201 [37]. Prior to the generation of the V3000 clone, this TrD strain had been passaged once in guinea pig brains and 14 times in embryonated eggs. The 68U201 isolate had been passaged once in newborn mice and two times in BHK-21 cells prior to the construction of the clone. From these clones, four chimeric variants were developed: two with matching cis-acting elements and two with mismatched elements (Fig. 1B and C). Two IE replicons, 68UGFP and 68UCFP, were derived from a full length IE 68U201 clone as previously described (Fig. 2) [32]. Replicons are replication deficient VLPs that can be utilized to analyze the initial sites of infection without the complication of cell-to-cell spread. These particles were generated by electroporating two RNA species simultaneously. The replicon, consists of the nonstructural open reading frame expressing a fluorescent reporter and associated cis-acting elements; the helper contains the structural portions of the genome. Co-electroporation of these two RNAs generates deficient particles that are unable to package the structural genes, but continue to express only the nonstructural genes from the replicon packaged into the particle. Replicons and helpers were transcribed using a T7 mMessage mMachine (Ambion, Austin, Texas), electroporated into BHK-21 cells, and harvested after 24 hours.
The first two chimeric clones were derived with mismatched cis-acting elements to directly compare the roles of the structural and nonstructural protein cassettes in mosquito infection and dissemination. Specifically, IAB/IE had the 5′ UTR and nonstructural protein gene region derived from IAB TrD and the structural protein gene region and 3′ UTR derived from IE 68U201. The reciprocal version, IE/IAB, had the 5′ UTR and nonstructural protein gene region of IE 68U201 and structural protein gene region and 3′ UTR derived from IAB TrD. Fusion PCR utilizing Phusion High Fidelity Polymerase (Finnzymes, Lafayette, CO) and designed around the Tth111I restriction enzyme site (Fig. 1) in the 26S UTR for each parental virus was used to generate a PCR fragment joining the two different viral cDNAs. Initially, for each reciprocal chimera, two overlapping fragments that encompassed the fusion site of the two genomes were generated by PCR using a forward primer from within the nsP4 region (7041 F IAB AND 6509 F IE) with a reverse fusion primer (IAB/IE R and IE/IAB R) and reverse primer downstream of the junction site (8007 R IAB and 8312 R) paired with a forward fusion primer for each chimera (IAB/IE F and IE/IAB R) (Table 1). The two individual fragments were joined by a PCR reaction on both templates utilizing the outermost primer sets. The fusion PCR fragment was cleaved with respective restriction enzymes (BssHII and PspOMI for IAB/IE and Bsu36I and NheI for IE/IAB) and ligated to the two other cDNA fragments with T4 DNA ligase (New England Biolabs, Beverly, MA). Ligated fragments were transformed into One Shot OmniMAX cells (Invitrogen, Carlsbad, CA), and resulting colonies were screened and sequenced prior to cesium chloride (CsCl) plasmid DNA purification.
The IAB/IE and IE/IAB constructs were then utilized to generate the infectious clones with matching cis-acting elements: IAB/IE/IAB and IE/IAB/IE. For both IAB/IE/IAB and IE/IAB/IE chimeras, a fusion PCR was designed at the junction at the end of the structural protein gene region and the start of the 3′ UTR. As described above, two PCR amplicons were generated using primers 10191 F IE, IAB/IE 3′ UTR R, IAB/IE 3′ UTR F, and 12030 R for IAB/IE/IAB and 9528 F IAB, IE/IAB 3′ UTR R, IE/IAB 3′ UTR F, and 12030 R for IE/IAB/IE (Table 1). The two fragments were ligated and then cleaved with restriction enzymes (SpeI and SacII for IAB/IE/IAB and SgrAI and EcoRI for IE/IAB/IE) to generate a single cloning fragment. These clones were ligated in 3 fragments, transformed, purified, and sequenced as described above.
Prior to transcription, plasmids were linearized with either NotI (V3000 backbone) or MluI (68U201 backbone) restriction enzymes [37], [38]. RNA was generated using the mMessage T7 RNA Polymerase Kit in the presence of an analog cap (Ambion, Austin, TX). The yield and integrity of transcripts were evaluated by agarose gel electrophoresis directly prior to electroporation. BHK-21 cells were electroporated using previously described conditions [39]. Virus was harvested at 48 hours post-electroporation when CPE was observed in greater than 80% of the cells. Virus titers were determined by plaque assay on Vero cells.
Replication kinetics of each of the two parental strains and four chimera strains were compared on Vero and C6/36 mosquito cells to identify any deficiencies and compare to in vivo infection and dissemination in C. taeniopus. Cells were seeded at a concentration of 106 cells/well in six well plates and allowed to attach and settle for 4 hours. Monolayers were infected in triplicate at a multiplicity of infection (MOI) of 5 PFU/cell and allowed to incubate for one hour at 37°C. Following incubation, cells were washed 3 times with phosphate-buffered saline (PBS), and overlaid with complete DMEM. Media were collected and stored from each well and replaced with the same volume of fresh media at predetermined time points, followed by plaque assays to measure viral yield. To compare the viral replication curves, a two-way ANOVA test and post-hoc multiple comparisons test with a Bonferroni correction was performed using JMP software, version 8.0.2 (SAS Institute Inc., Cary, NC). P-values≤0.05 were considered significant.
Two parental viruses (IAB, IE) and four chimeras (IAB/IE/IAB, IE/IAB/IE, IE/IAB, IAB/IE) were evaluated for their ability to infect and disseminate in C. taeniopus. The C. taeniopus colony was established from mosquitoes collected from Chiapas, Mexico in 2007 as described previously [40]. For all studies, 10-week old female CD1 mice (Charles River Laboratories) were used as viral hosts. To develop natural viremia, mice were infected with 3 log10 PFU of each virus [41] by subcutaneous (SC) inoculation, held for 24 hours, anesthetized by intraperitoneal (IP) administration of sodium pentobarbital (50 mg/kg), and bled via the retro-orbital sinus to determine viremia levels. Since the replicon particles utilized for this study do not replicate beyond the initial cell infected, and C. taeniopus will not feed on artificial bloodmeals, we utilized an artificial system in which we inoculated CD1 mice intravenously (IV) allowing for an immediate nonreplicative viremia. Mice were anesthetized by IP inoculation of sodium pentobarbital and 200 µl of a stock replicon or a 1∶1 mix of replicons was inoculated into the tail vein. Particles were allowed to circulate for 1–2 minutes before blood was collected from the retro-orbital sinus to estimate the artificial viremia level achieved; the animal was then exposed to mosquitoes for ca. one hour, after which blood was collected again from the retro-orbital sinus to detect any changes in the circulating replicon concentration. Following exposure, engorged mosquitoes were sorted and incubated for 14 days at 28°C with 75–80% humidity. A 10% sucrose solution was provided ad libitum. Statistical analysis of rates of infection and dissemination were broadly examined using a contingency analysis, and specific 2×2 comparisons were evaluated using Fisher's exact test with JMP software (SAS Institute Inc., Cary, NC). P-values≤0.05 were considered significant.
Viral titers of rescued viruses and animal sera were determined by plaque assay on Vero cells. Following the 14-day extrinsic incubation period (eip), legs and wings were removed from mosquitoes and stored at −80°C. Samples were triturated, centrifuged at 9500× G for 5 minutes, and used to infect monolayers of Vero cells in CPE assays. Triturated body samples that generated CPE were indicative of an infected mosquito, while legs and wings were used to detect a disseminated infection. Replicon titration was done in a similar manner to the plaque assay; ten-fold serial-dilutions were plated on a monolayer of Vero cells and allowed to incubate for one hour prior to an overlay with DMEM supplemented with FBS and penicillin/streptomycin. After 24 hours, the media were removed and the monolayer was fixed with 4% paraformaldehyde (PFA) (Affymetrix, Santa Clara, CA) for one hour. The number of fluorescent cells per well was counted using an Olympus Is71 inverted fluorescent microscope and reported as fluorescence units (FU).
Mosquito samples were processed 72 hours after blood feeding to minimize chances of damaging the midgut while distended with blood and to allow for clear images of the midgut epithelia. Mosquitoes were cold anesthetized and submerged for 30 seconds to 1 minute in 70% EtOH prior to being transferred to a PBS solution. Midguts were extracted and covered with a drop of 4% PFA on a glass slide for 30 minutes, then was rinsed twice with PBS before the addition of ProLong Gold Antifade with DAPI (Invitrogen).
Mosquito midguts were imaged on an Olympus BX61 fluorescent microscope and high-resolution images were taken on an Olympus FluoView FV1000MPE confocal microscope. In vitro dual infection experiments were visualized on an Olympus DSU-IX81 spinning disk confocal microscope and analyzed with MetaMorph Software (Molecular Devices, Sunnyvale, CA).
Figure 1 shows the genetic composition of the viruses utilized in this study. The two parental viruses included subtype IE strain 68U201 and subtype IAB strain TrD, which share 77% nucleotide and 89.8% amino acid identity. Four chimeric strains were derived from these parental strains: IAB/IE/IAB and IE/IAB/IE were designed with matching 5′ and 3′ cis-acting sequence elements. IAB/IE and IE/IAB were designed with mismatched cis-acting elements, where the 3′ UTR matched the strain used for the structural protein region of the chimera and the 5′ UTR matched the strain used for the nonstructural protein region of the chimera. It has been shown repeatedly that conserved regions in both the 5′ and 3′ UTRs of alphaviruses are essential for proper synthesis of both negative and positive strand RNA species [42]–[45]. Therefore, chimeras with both matching and mismatching cis-acting elements were utilized to compare these regions of interest and their effect on replicative efficiency.
One-step replication curves of the parental and chimeric strains were performed on Vero cell monolayers at an MOI of five PFU/cell to identify any replication deficiencies that could bias experimental findings in the mosquito model. All of the chimeras showed similar replication, with no major deficiencies when compared to the parental strains (Fig. 3A). However, analysis of variance indicated that the replication of the viruses was significantly different (p<0.0001). Multiple comparison tests showed that strain TrD exhibited higher replication levels at multiple time points (not shown), which does not likely correlate to the in vivo mosquito model because this strain is unable to infect and disseminate in C. taeniopus [40], [46].
One-step replication analyses were also performed on monolayers of C6/36 A. albopictus cells to compare to the in vivo mosquito model (Fig. 3B). No major replication deficiencies were observed; however, analysis of variance did indicate differences among the viruses (p<0.0001). Unlike the Vero cell replication curves where just the TrD strain differed from all other viruses, differences were seen between nearly all viruses upon pairwise comparisons (not shown). The only three pairs out of the total 15 comparisons that did not show any statistical differences from one another were between IE versus IE/IAB/IE, IAB/IE/IAB versus IE/IAB, and IE/IAB/IE versus IE/IAB.
Adult female C. taeniopus were exposed to a range of oral doses for each of the parental and chimeric strains of VEEV and tested for infection and dissemination into the hemocoel following a 14-day eip (Table 2). Two pairs of chimeras with matched and mismatched cis-acting elements were utilized to independently evaluate the roles of the nonstructural and structural polyprotein open reading frames as well as the 3′ UTR in mosquito infection and dissemination. As expected, the parental IAB TrD virus was unable to infect C. taeniopus at blood meal titers as high as 6.2 log10 PFU/ml, which is in agreement with previous work [46], [47]. Similarly, as predicted based on previous studies [30], [40], [46], C. taeniopus mosquitoes were highly susceptible to infection with the parental subtype IE 68U201 strain at oral doses as low as 4.2 log10 PFU/ml.
All four chimeras showed an intermediate ability to infect and disseminate in C. taeniopus when compared to the parental IAB and IE strains (Table 2; Fig. 4). The effect of the exposure dose on infection rate was evaluated by contingency analysis for each of the chimeric viruses (IAB/IE/IAB, IE/IAB/IE, IAB/IE, and IE/IAB) and found to be significant for each (p<0.05; p<0.001; p<0.05; p<0.0001, respectively) (Fig. 4A). In order to compare individual virus strains, a Fisher's exact test was utilized to determine differences in infection rates (Table 3). As expected, comparisons between the parental viruses and the chimeric viruses were all highly significant (p<0.0001), with the exception of the comparison between the IE parental virus and IE/IAB chimera, for which the IE strain had a less notable infectious advantage than the chimera (p<0.0071). Interestingly, infection rates did not differ significantly among three of the four chimeras: IAB/IE, IAB/IE/IAB, and IE/IAB/IE. However, IE/IAB showed a significantly higher infection rate when compared to each of the other three chimeras (p<0.0001; p<0.0037; p<0.0025, respectively).
Although each chimera showed the ability to disseminate into the hemocoel after midgut infection, the dissemination rates among the chimeras were low overall (Figs. 4B and 4C); therefore, no transmission experiments were performed. Previous studies of alphaviruses as well as other arboviruses have shown that infected mosquitoes often have virus restricted to the midgut, which is likely explained by a commonly recognized but poorly understood barrier to viral escape of the mosquito midgut [46], [48]–[50]. While dissemination rates increased as the exposure dose was increased, the overall rates of dissemination were too low to perform reliable statistical analysis. Fisher's exact tests comparing the rates of infected mosquitoes with dissemination (Fig. 4C), showed no differences between the four chimeras.
To observe the number of epithelial cells initially infected, the location of the infected cells, and to determine whether there is a subpopulation of cells within the midgut that is more susceptible than other epithelial cells, C. taeniopus mosquitoes were exposed to a range of doses of 68U201 replicon particles expressing fluorescent proteins (Fig. 2). For the single replicon infections, a clear dose-response was observed such that the lowest oral dose (average of pre- and post- exposure titers) of 3.0 log10 FU/ml infected only 11% of examined midguts with only 2 infected cells/midgut, whereas the highest dose of 7.2 log10 FU/ml infected 100% of examined midguts with a range 535–1757 infected cells/midgut (Table 4). Infected cells were not limited to any particular region of the abdominal midgut (Fig. 5), and only a minority of the midguts (9%) were found to have infection focused in the posterior portion, whereas 25% showed a focused infection in the anterior portion of the abdominal midgut. The remaining 66% of infected midguts showed a mixed infection with concentrated infection within the middle portion of abdominal midgut. Infection of the midgut/foregut junction was not observed.
To determine if there was differential susceptibility of midgut cells, C. taeniopus mosquitoes were orally infected with a 1∶1 mixture of 68UGFP and 68UCFP. A total of fifteen mosquitoes was examined for co-infection at two different doses. The low exposure dose achieved by artificial viremia was a mixture of 5.4 log FU/ml 68UGFP and 5.0 log FU/ml of 68UCFP, and the high dose achieved was 6.5 log FU/ml of each replicon. At the low dose, an average of 70 midgut epithelial cells were infected with 68UGFP and an average of 52 cells was infected with 68UCFP. At the high dose, the average number of cells infected with 68UGFP was 896, whereas the average number of 68UCFP infected cells was 866. At the low dose and of the five co-exposed mosquitoes examined, no co-infected cells were observed. At the high dose where 15 midguts were examined, there appeared to be 2–3 cells with co-localization; however, it was determined that these areas of co-localization were a result of either signal bleed-through or overlap. Even so, there was still an average of less than one observed co-infected cell per midgut in the highest dose group.
As human populations continue to expand into rural environments, the incidence of emerging and re-emerging zoonotic pathogens will continue to climb. This has already been observed with other arboviral viruses, such as chikungunya, dengue, yellow fever, and Japanese encephalitis viruses [51]. Similar trends have also been observed with enzootic strains of VEEV that have caused endemic disease as well as outbreaks in Peru, Central America, and Mexico [16], [17]. Historically, studies of VEEV emergence have focused on epidemic strains within subtypes IAB and IC; however, enzootic ID and IE strains can also cause a large burden of endemic disease, which can often be misdiagnosed as dengue fever [52]. Recent studies have also shown that the primary mosquito vector of enzootic IE, C. taeniopus, can be an efficient vector of newly emerged epizootic IE strains in Mexico [40], [53]. Considering the growing risk of enzootic VEEV strains in causing human disease, it is important to understand the determinants and dynamics for viral infection of the primary enzootic vector, which we hypothesize to be different from what is known about the epizootic virus-vector interaction.
We first examined the genetic determinants of infection and dissemination utilizing chimeric viruses to analyze the molecular determinants for VEEV specificity to the enzootic mosquito vector, C. taeniopus. We used two viruses with distinct phenotypes for these chimeras, to help identify the major genome regions that contribute to specific infection of the enzootic vector. However, because these viruses have such a wide genetic divergence (10.2% at the amino acid level), extrapolation of this method to clarify the roles of each gene during enzootic mosquito infection could be prone to bias from incompatibilities between open reading frames within each chimera (Table 5). Therefore, to ensure that chimerization did not result in general attenuation of virus replication, the parental and chimeric strains were evaluated using in vitro replication curves on Vero and C6/36 cell monolayers; no replication deficiencies were observed. It was noted that the replication in mosquito cells was different from what was observed in the in vivo mosquito model in that the parental IAB virus, which showed no deficiencies in vitro, was unable to infect the in vivo model. Similarly, the differences observed between the IE parental and the four chimeras in the in vivo model were not demonstrated in the in vitro model. These results emphasize the importance of using an in vivo mosquito model to detect differences in viral replication, which may not be detected in a mosquito cell line.
We orally exposed C. taeniopus mosquitoes to varying doses of two parental strains, subtype IAB TrD and subtype IE 68U201, as well as four chimeric strains, IAB/IE/IAB, IE/IAB/IE, IAB/IE, and IE/IAB, and evaluated the role of the nonstructural and structural protein genes and the 3′ UTR as determinants of infection and dissemination in C. taeniopus. We hypothesized that, unlike the epizootic virus strains and their vectors, the genetic determinants for enzootic infection include multiple genes and they are not limited to a single region in the structural portion of the genome. Our data supported this hypothesis, as all four chimeras were able to infect and disseminate in C. taeniopus, albeit at rates lower than the wild-type IE parental strain. We included multiple replicates within each viral group to compensate for variations within the mosquito colony. If the E2 or the structural regions were the primary determinants of infection and dissemination, those chimeras with IE-derived structural regions would have infected mosquitoes at a higher rate than those chimeras with IAB-derived structural regions, and this was not observed.
We anticipated that the chimeras with mismatched 3′UTR regions would show diminished infection and dissemination rates based on previous alphavirus studies examining the effects of mismatched cis-acting elements [54], [55]; however, our IE/IAB chimera showed a significantly higher rate of infection than that of the other three chimeras. This suggests that the 3′ UTR plays some role in infection of the enzootic vector. A closer examination of the effect of the 3′ UTR on infection showed that the chimera with IAB structural genes and IAB 3′ UTR (IE/IAB) had the highest infection rate, while the chimeras with a mixed structural-3′ UTR makeup (IE/IAB/IE or IAB/IE/IAB) had intermediate infection abilities, and the chimera with IE in the structural and the 3′ UTR (IAB/IE) actually had the lowest rate of infection. This suggests that 3′ UTR acts in concert with other portions of the genome, although it is unclear which specific regions are important for this cooperative effect. These potential interactions should be further explored with 3′UTR-specific chimeras, such as a IAB virus backbone with a IE derived 3′UTR and a IE backbone with a IAB-derived 3′ UTR. While the role of these regions was not mirrored by in vitro mosquito infections, our C6/36 data were based on cells from A. albopictus, which in laboratory experiments has been shown to be equally susceptible to epizootic IC and enzootic ID VEEV strains [56]. Previous studies examining chimeras between Ross River virus (RRV) and Sindbis virus (SINV), two genetically distant alphaviruses, have shown that mismatched 3′ UTR regions can result in depressed RNA synthesis in vitro, although the effects on replication in vivo have not been examined [55]. However, studies of chimeras between more closely related alphaviruses, such as o'nyong-nyong (ONNV) and chikungunya (CHIKV) viruses, indicate that chimerization does not have a deleterious affect on the infection of the CHIKV mosquito vector, A. aegypti. There was also no indication that mismatched 3′ UTRs altered infection rates [57].
There were no statistical differences in the infection rates between chimeras IAB/IE, IAB/IE/IAB, and IE/IAB/IE, indicating that both the structural and nonstructural protein regions of the enzootic virus play a role in vector infection, as none of the chimeras displayed infection rates as high as the parental IE strain. However, we observed a trend in which the two chimeras with IE-derived nonstructural protein genes showed higher rates of infection at higher doses. Specifically, the chimeras with IE-derived nonstructural protein genes reached 100% infection at the highest doses, while the chimeras with IAB-derived nonstructural protein gene regions never reached 100% infection even at the highest doses. The diminished infection of all chimeras implies that there are multiple determinants of infection that reside in different genome regions and may act synergistically. Our results show that the determinants for infection of the enzootic vector do not reside solely in the structural protein genes, specifically not only in the E2 glycoprotein of the genome, which supports our hypothesis that infection determinants for VEEV in the enzootic mosquito vector relies on both structural or nonstructural protein regions of the genome. Interestingly, our findings suggest that in the enzootic model, the nonstructural elements are stronger determinants of vector infection.
We also hypothesized that the characteristics of initial midgut infection of the enzootic mosquito vector would be inherently different than those used by the epizootic virus in A. taeniorhynchus. To test this hypothesis, we exposed the enzootic vector, C. taeniopus, to replicon particles generated from a subtype IE enzootic strain. Examination of the initial sites of infection in the midgut indicated multiple locations in the abdominal portion with no predilection for either the anterior or the posterior region; we detected no infection of the cardial epithelium at the midgut/foregut junction. Similar to what was observed in A. taeniorhynchus, a clear response was observed between the oral dose and the number of infected midgut cells, although the ID50 for C. taeniopus was lower and the maximum number of infected cells was higher than the 100 susceptible A. taeniorhynchus cells previously estimated [32]. The greater number of infected cells (>1700) in C. taeniopus following high oral doses indicates that a larger number of its midgut epithelial cells is susceptible to VEEV IE infection compared to A. taeniorhynchus and VEEV IC. This observation, in conjunction with our observation of no co-infected midgut cells in the mixed replicon experiments, supports the hypothesis that the population size of enzootic VEEV virions during initial infection of the midgut is not severely restricted by a limited number of susceptible C. taeniopus epithelial cells. The average population of cells infected by the 68UGFP replicon at the highest dose did not differ from the average number of cells infected by the 68UCFP replicon. This suggests that there is no effect of co-exposure on individual particle infection rates. Utilizing the same methods, previous studies in the epizootic VEEV/mosquito model found an average of 26 midgut cells co-infected with two replicons, which was greater than what we observed in the enzootic model. This indicates that the initial infection of the enzootic vector differs from that of the epizootic VEEV strain. Using the Poisson distribution and given the epizootic data (a model with a small population of susceptible cells), we determined the probability of observing less than a single co-infected cell out of our five C. taeniopus midgut replicates to be 5.1×10−12, indicating an extremely low likelihood that there is a subpopulation of midgut epithelial cells with enhanced susceptibility.
Our studies illustrate the contrast in the virus-vector interactions between the enzootic and epizootic VEEV cycles. Not only do these interactions persist in different ecological cycles and infect different species of mosquitoes, but they also behave differently within their respective vectors. This difference may be explained by the dissimilar selective pressures that are exerted on each subtype during transmission. For example, epizootic viruses produce a very high level of viremia in equids, which facilitates VEEV transmission by epizootic vectors even if only a small population of their midgut cells are initially infected. However, the highly susceptible enzootic vector, which appears to have a greater number of susceptible midget cells that can be initially infected even after small oral doses, can transmit efficiently among populations of rodents that develop only moderate viremia titers [25]–[27].
As the growing impact of enzootic VEEV on human health is becoming more apparent, especially after the recent emergence of epizootic-like IE strains, understanding how these viruses interact with vectors is critical to estimating their threat to human health and for refining public health prevention strategies as well as developing vaccines. For instance, the design strategy of a vaccine that is protective against epizootic and enzootic strains that are currently causing human disease must also consider mosquito vectors that could potentially acquire and transmit should a vaccinee become viremic. If the epizootic vector only has a few susceptible midgut cells and is examined for competence for a given vaccine strain, it may appear to be incompetent. However, the same vaccine may be able to establish an infection in the enzootic vector. Considering that the determinants for infection appear to differ between the two vector types, vaccine strains that are derived from epizootic VEEV and depend on the elimination of mosquito infection may not necessarily reflect how infectious these vaccine candidates would be for enzootic vectors. As enzootic habitats are encroached upon and enzootic cycles gain close proximity to epizootic habitats, it is essential to consider the contribution of enzootic vectors to viral emergence and the potential introduction of vaccine strains into natural cycles.
|
10.1371/journal.pntd.0003555 | Serology for Trachoma Surveillance after Cessation of Mass Drug Administration | Trachoma, caused by Chlamydia trachomatis (Ct), is the leading infectious cause of blindness worldwide. Yearly azithromycin mass drug administration (MDA) plays a central role in efforts to eliminate blinding trachoma as a public health problem. Programmatic decision-making is currently based on the prevalence of the clinical sign “trachomatous inflammation-follicular” (TF) in children. We sought to test alternative tools for trachoma surveillance based on serology in the 12-year cohort of Kahe Mpya, Rombo District, Tanzania, where ocular chlamydial infection was eliminated with azithromycin MDA by 2005.
The present study was a community-based cross-sectional survey in Kahe Mpya. Of 989 residents, 571 people aged 6 months to 87 years were enrolled: 58% of the total population and 73% of 1–9 year olds, the key WHO indicator age group. Participants were examined for TF, had conjunctival swabs collected for nucleic acid amplification test (NAAT)-based detection of Ct, and blood collected for analysis of antibodies to the Ct antigens pgp3 and CT694 by multiplex bead-based immunoassay. Seroconversion rate was used to estimate changes in the force of infection in a reversible catalytic model. No conjunctival swabs tested positive for Ct infection by NAAT. Among 1–9 year olds, TF prevalence was 6.5%, whereas only 3.5% were seropositive. Force of infection modelling indicated a 10-fold decrease in seroconversion rate at a time corresponding to MDA commencement. Without baseline serological data, the inferences we can make about antibody status before MDA and the longevity of the antibody response are limited, though our use of catalytic modelling overcomes some of these limitations.
Serologic tests support NAAT findings of very low to zero prevalence of ocular Ct in this community and have potential to provide objective measures of transmission and useful surveillance tools for trachoma elimination programs.
| Trachoma is the leading infectious cause of blindness. The infectious agent, Chlamydia trachomatis, can be treated with a single oral dose of azithromycin. Donated drug is a cornerstone of programs dedicated to the elimination of trachoma as a public health problem. Azithromycin is given to the entire district for 3–5 years when 10% or more of 1–9 year-olds in the district have signs of a defined follicular conjunctivitis in one or both eyes. However, follicles can be difficult to reliably diagnose and can be caused by other pathogens, especially in settings with low trachoma prevalence. More sensitive and specific ways to assess communities for trachoma transmission at program endpoints are needed. Herein we examined antibody responses in children living in a community in Tanzania born after stopping drug treatment 10 years previously. Low antibody levels (3.5% in 1–9 year-olds) reflected the lack of ocular chlamydial infection in these children. We also modelled the data to show that changes in age-specific antibody prevalence occurred when the mass drug treatment stopped. These data suggest that the age-specific prevalence of antibody responses may be of use to programs seeking to demonstrate the impact of interventions against trachoma.
| Trachoma, caused by the bacterium Chlamydia trachomatis (Ct), is the leading infectious cause of blindness worldwide [1]. Infection can manifest clinically in a number of ways, including follicular conjunctivitis, classified as “trachomatous inflammation-follicular” (TF) in the WHO simplified grading system [2] if five or more follicles are present in the central upper tarsal conjunctiva; and/or inflammatory thickening, classified as “trachomatous inflammation-intense” (TI) if more than half of the deep tarsal vessels are obscured. Repeated infections can lead to conjunctival scarring (TS) and trichiasis (TT), in which in-turned eyelashes rub against the globe and may result in visual impairment or blindness caused by corneal opacity (CO) [3]. Azithromycin mass drug administration (MDA), recommended where the prevalence of TF is ≥10% in children aged 1–9 years, is a critical component of the strategy for Global Elimination of Trachoma by 2020 (GET2020) [4]. The current WHO endpoint for cessation of community-based antibiotic treatment is a TF prevalence in 1–9 year-olds of <5%.
Prevalence surveys illustrate that signs of active trachoma, TF and TI, exceed Ct infection rates. Follicular or intense conjunctivitis may be caused by non-chlamydial bacteria, with the relative importance of this phenomenon probably increasing after populations begin to receive azithromycin MDA [5]. Furthermore, the examination process can be difficult to standardize [6–9]; inter-observer agreement is often sub-optimal. The poor correspondence between signs and infection—seen at both individual and community level—is problematic, given that field grading is the basis of public health decision-making [5, 10].
As trachoma elimination efforts are intensified globally and interventions move populations towards trachoma elimination goals, the availability of a post-elimination surveillance methodology with greater reliability than clinical examination will become increasingly important to allow programs to identify and respond to recrudescent infection. Recent efforts to evaluate serology as a viable option for post-MDA surveillance identified tests using two previously-described chlamydial antigens, pgp3 and CT694, as having high sensitivity to detect current ocular infection, and high specificity using non-endemic controls [11]. The age-specific prevalence of serological responses to Ct antigens at community level could provide an informative proxy measure of intensity of transmission and an early indicator of transmission recrudescence. This study therefore examined the use of serological tools for monitoring and evaluation in a post-MDA setting by assessing the age-specific prevalence of signs of trachoma and Ct-specific antibody responses within a community in which MDA ceased in 2002 and ocular Ct infection was subsequently found to have been eliminated in 2005 [12].
This study was conducted in the Tanzanian community of Kahe Mpya, Rombo District. Kahe Mpya consists of approximately 250 households, with a population (in July 2012) of 989. A Kilimanjaro Christian Medical College (KCMC)/London School of Hygiene & Tropical Medicine (LSHTM)/Huruma Hospital collaboration has been conducting trachoma research in this community since 2000 [13, 14]. High coverage azithromycin MDA was delivered in 2000 and 2002, and topical tetracycline ointment treatment was given, at intervals between 2000 and 2005, to individuals with active trachoma; elimination of ocular Ct infection by 2005 was previously documented [12]. Ethical approval to carry out this research was obtained from the ethics committees at LSHTM (UK), Centers for Disease Control and Prevention (USA), KCMC / Tumaini University, and the National Institute for Medical Research (TZ). All adults provided written informed consent, and for children under 18, the consent of a parent or guardian was obtained.
All Kahe Mpya residents were invited by village leaders to a series of central locations, where those consenting to the study underwent examination of both eyes by a trained, highly experienced ophthalmic nurse known to the community, using binocular loupes (magnification ×2·5) and a torch. Signs of trachoma were graded according to the WHO simplified grading system [2]. After examination, swabs were collected from the everted upper eyelid of the right eye using a sterile polyester-tipped-swab by passing the swab across the conjunctiva four times. Swabs were placed into sterile polypropylene tubes and kept at 4°C until frozen (-20°C). Individuals with signs of active trachoma were given a tube of 1% tetracycline eye ointment free of charge and instructed to apply it daily to both eyes for six weeks. Fingerprick blood was collected by Tanzanian registered physicians onto filter paper with six circular extensions, calibrated so that each extension absorbed 10μl of whole blood (TropBio Pty Ltd, Townsville, Queensland, Australia). Each filter paper was air-dried then individually placed in a zip-lock bag and frozen (-20°C). Each sample was affixed with a pre-printed bar-coded label that linked all samples from an individual but had no other patient identifier.
Dried blood spots were shipped to the Centers for Disease Control and Prevention in Atlanta GA, USA, for detection of IgG antibodies against the previously described chlamydial proteins pgp3 and CT694, on the Luminex platform, using previously defined cut-offs for positivity [11]. Briefly, serum eluted from dried blood spots was incubated with microbeads coupled to the antigens of interest, then excess serum washed off and bound antibody detected with an anti-human IgG and anti-human IgG4 biotinylated detection antibody, and finally detected using streptavidin-conjugated to phycoerythrin (PE). The fluorescent signal emitted by bound PE was converted to a median fluorescence intensity (MFI) with background from the blank subtracted out (MFI-BG). For pgp3, a MFI-BG value of 1024 was established as the low-limit value for positivity, with an indeterminate range of 1024 to 5998. For CT694, a MFI-BG value of 232 was established as the low-limit value for positivity, with an indeterminate range of 232 to 1982 [11].
To examine the change in transmission following MDA, we used seroconversion rate (SCR) to estimate the force of infection by fitting a simple reversible catalytic model to the measured seroprevalence, stratified into yearly age-groups, using maximum likelihood methods [15]. For these models only individuals aged one year and over were included to remove the effect of maternally derived antibodies in infants. Evidence for temporal changes in SCR was explored by fitting models in which the SCR was allowed to change at a single time-point. The significance of the change was identified using likelihood ratio tests against models with no change, and profile likelihoods were plotted to determine confidence intervals for the estimated time of the change.
Samples were processed at the LSHTM and tested in pools of five using the Roche CT/NG Amplicor kit (Roche Molecular Systems, Pleasanton, CA, USA), with the intention of re-testing positive pools as individual samples [16–18]. Manufacturer’s instructions were followed except for sample extraction where a previously published protocol was used[14]. Two Ct positive and two Ct negative processing controls were run with each batch of specimens. According to the manufacturer’s directions, the Amplicor test was positive if the optical density read at 450 nm was ≥0·8, negative if the signal was <0·2, and equivocal if in-between. All equivocal tests were re-tested in duplicate, and only graded positive if at least one test was positive.
All samples were analysed in anonymous fashion through the use of non-sequential sample codes linked only to patient records through the data collection sheet. Statistical analysis was carried out using STATA 12 and GraphPad Prism (version 6.0).
The population and study population structure of Kahe Mpya sub-village is summarized in Table 1, based on census data collected in July 2012 for this study. From the total 989 residents of Kahe Mpya sub-village, 575 (58.1% coverage) people aged 0.2–87.6 years (median age 12.6, Table 1) participated in the study.
The overall prevalence of active trachoma (TF,TI or both) in the examined population (n = 571; four individuals refused clinical exams) was 4.6%, with 21.5% exhibiting signs of scarring trachoma (TS/TT/CO, Table 2). There were no WHO simplified grading scheme signs of trachoma in 76·6% of the study group. The prevalence of TF amongst the WHO index age group (ages 1–9 years) was 6·5% (Table 2). Only one individual ≥ 10 years had TF. TS was absent in those <10 years, but was observed in all age groups >10 years (Table 2). TT was present only in individuals >20 years of age, with an overall population prevalence of 1% (Table 2). CO was only diagnosed in 2 individuals (0·4% of study participants), both of whom were over 70 years old (Table 2).
Overall, 33.8% of participants were seropositive against at least one antigen (Fig. 1A). Seropositivity increased with age. By age 40, over 90% of participants tested positive to at least one antigen (pgp3 alone, CT694 alone, or both pgp3 and CT694, black squares, Fig. 1A), and over 60% tested positive to both antigens (Fig. 1A, red squares); this trend continued to the oldest age groups (Fig. 1A). The MFI also increased with age (Fig. 1B). Of 200 children aged 1–9, seven (3.5%) had antibody responses to one antigen, whereas only two (1%) had antibody responses to both antigens (Fig. 1C). Five of the seven samples with pgp3 reactivity fell into the indeterminate range, as did both of the CT694-reactive samples (Fig. 1C). Samples from six of the seven 1–9 year olds testing positive by serology were re-tested with separate pgp3 and CT694 bead sets and data replicated the original results (S1 Table).
None of the ocular swabs tested positive by NAAT.
When a seroconversion model, which allowed for a single change in SCR, was fitted to the data, the best fit was provided by a change in transmission between 10–15 years previously, consistent with the timing of MDA in the years 2000 and 2002 (Fig. 2A for antibody responses to either antigen; responses to individual antigens gave similar profiles). We chose a model in which SCR changed 10 years previously, which had a better fit than the model that assumed the SCR had remained constant (Fig. 2B). The change in SCR before and after this change point is approximately a 10-fold reduction, from a pre-MDA SCR of 0.0448 (95%CI 0.0373–0.0537) to a post-MDA SCR of 0.004 (95%CI 0.0024–0.0093)].
Global efforts toward the elimination of blinding trachoma are being rapidly intensified, thanks to strong donor interest. As programs reduce the prevalence of disease and infection, robust surveillance systems will become crucial to detect any recrudescence in populations living in post-elimination settings. In this study, we examined the use of serological tools for trachoma in a post-MDA setting. The virtual absence of antibody responses in children born after MDA-precipitated elimination of ocular Ct infection reflects the lack of Ct transmission (as suggested by NAAT) and provides the first evidence that serological monitoring of antibody responses could be viable for informing programmatic decisions in the surveillance phase. Force of infection modelling shown here strongly supports the hypothesis that reductions in transmission in this community coincident with the commencement of azithromycin MDA were reflected in changes in Ct seroconversion rate. This suggests that serology could have a very useful programmatic role even in the absence of complete transmission interruption.
Several factors could contribute to the presence of signs of active trachoma in a community with low or no transmission of conjunctival Ct. First, the WHO simplified grading system employs strict criteria for diagnosis, but was designed for simplicity rather than specificity. Our grader was, however, well trained, highly experienced, and internationally certified, and we are confident of the accuracy of his judgements about the presence or absence of TF. Second, the natural histories of infection and disease differ, with signs arising weeks after infection has been acquired and persisting for weeks or months after infection clears. At the population level, the prevalence of infection declines more rapidly than the prevalence of TF following MDA [12, 19, 20] with some studies showing that TF persists at levels >10% within the population for months or years after infection has subsided [21, 22]. Finally, evidence suggests that, in low-trachoma-prevalence settings, the majority of TF is associated with conjunctival infection with non-chlamydial bacteria, including S. pneumoniae and H. influenzae [5, 23]. Non-bacterial causes of conjunctivitis such as adenovirus [24] may also contribute to TF clinical diagnoses in low-trachoma-prevalence settings. The use of photographs to validate field exams is becoming increasingly common but we have not found it to be reliable [25] and did not incorporate it into this study.
Antibodies against the Ct antigens among 1–9 years old in this study were present at very low prevalence and in general at very low densities. This is in stark contrast to areas of active transmission in which seropositivity exceeds rates of clinical disease, as would be expected from long-lived antibody responses.[26] and has high sensitivity for ocular infection,[11, 26] In the present study, antibody responses in 1–9 year olds may be Ct-specific, resulting from ocular or respiratory Ct infection acquired at birth from a mother with genital tract infection [27], or from ocular infection acquired outside the village or in the village itself. Because the target for trachoma programs is not the complete interruption of transmission, it would not be an indication of programmatic failure to find ongoing low-level transmission in a community. However, it should also be noted that the previously determined specificity limits of this serological assay were 96–98%, such that the 3·5% of 1–9 year old samples testing positive may be false positives[11].
Because data were collected from a single community and enrolment was lower than anticipated (primarily due to lack of availability of participants at the time of enrolment, as many adults were working outside of the community at the time of the study), additional studies in post-MDA settings will be needed to confirm the generalizability of our data. While the overall study enrolment was 58.1% of the total population, enrolment of 1–9 year olds, the key WHO indicator age group, was approximately 72.9% (extrapolated from 2010 census data). Without baseline serology data, the inferences we can make about antibody status before MDA, the longevity of antibodies, and how antibody titers change over time in relation to one another are restricted, although our application of catalytic modelling overcomes some of these limitations. While comparing baseline to post-MDA antibody levels would be optimal, programs using serological tests as monitoring tools for intervention impact would need to do so in populations from whom baseline serological data will be absent. Antibody responses will therefore be most useful as surveillance tools by focusing analyses on children born after initiation or cessation of interventions. The data presented in the current study show the power of antibody-based surveillance in children born after cessation of an MDA program, data supported by the historical documentation of interruption of ocular Ct transmission in this community.
Antibody responses represent exposure to infection and, when integrated with age, represent exposure over time; this can be done simply by applying a catalytic conversion model. SCR has been used widely in a range of infectious diseases [28, 29], most recently and extensively for malaria, for which SCR has been shown to correlate with the force of infection [15, 30, 31]. Fitting models with two SCRs enabled the measurement of changes in force of infection. SCR suggests a 10-fold decrease in the force of infection from approximately 5% seroconversion in the population per year prior to MDA, to approximately 0.5% after MDA, which closely approximates the 0% ocular infection prevalence seen in this study. Catalytic models can be refined by using serological data from multiple settings, pre- and post-MDA, to further validate the use of serological testing for programs.
Serological tests for measuring antibodies in children may represent the best option for monitoring transmission because of the potential for greater sensitivity as population-based markers of exposure. Additionally, they provide an objective marker, relatively free of observer bias (unlike examination for clinical signs), and are likely to be lower in cost than NAATs and provide data on cumulative exposure to the bacterium. Programmatically, such an assay could be used in the same way that antigen detection assays are used in surveillance for lymphatic filariasis elimination programs, and seroprevalence has been proposed for malaria control and elimination programs [30, 32]; that is, to document reductions in the force of transmission. With the recent increased emphasis on a more horizontal approach to disease control, given similarities in control methods (particularly periodic MDA) and the geographical overlap between trachoma and other NTDs, integration across NTD programs is the next step [33]. This will provide economic and pragmatic benefits, as a multiplexed serological tool has the potential to map, monitor and evaluate several diseases simultaneously, facilitating efforts to achieve long-term elimination goals.
|
10.1371/journal.pcbi.1006153 | Functional triplet motifs underlie accurate predictions of single-trial responses in populations of tuned and untuned V1 neurons | Visual stimuli evoke activity in visual cortical neuronal populations. Neuronal activity can be selectively modulated by particular visual stimulus parameters, such as the direction of a moving bar of light, resulting in well-defined trial averaged tuning properties. However, given any single stimulus parameter, a large number of neurons in visual cortex remain unmodulated, and the role of this untuned population is not well understood. Here, we use two-photon calcium imaging to record, in an unbiased manner, from large populations of layer 2/3 excitatory neurons in mouse primary visual cortex to describe co-varying activity on single trials in neuronal populations consisting of both tuned and untuned neurons. Specifically, we summarize pairwise covariability with an asymmetric partial correlation coefficient, allowing us to analyze the resultant population correlation structure, or functional network, with graph theory. Using the graph neighbors of a neuron, we find that the local population, including both tuned and untuned neurons, are able to predict individual neuron activity on a moment to moment basis, while also recapitulating tuning properties of tuned neurons. Variance explained in total population activity scales with the number of neurons imaged, demonstrating larger sample sizes are required to fully capture local network interactions. We also find that a specific functional triplet motif in the graph results in the best predictions, suggesting a signature of informative correlations in these populations. In summary, we show that unbiased sampling of the local population can explain single trial response variability as well as trial-averaged tuning properties in V1, and the ability to predict responses is tied to the occurrence of a functional triplet motif.
| V1 populations have historically been characterized by single cell response properties and pairwise co-variability. Many cells, however, do not show obvious dependencies to a given stimulus or behavioral task, and have consequently gone unanalyzed. We densely record from large V1 populations to measure how trial-to-trial response variability relates to these previously understudied neurons. We find that individual neurons, regardless of response properties, are inextricably dependent on the population in which they are embedded. Specifically, patterns of correlations between groups of neurons, allow us to predict moment to moment activity in individual neurons. Only by studying large, local, populations simultaneously were we able to find an emergent property of this information. These results imply that understanding how the visual system operates with substantial trial-to-trial variability will necessitate a network perspective that accounts for both visual stimuli and activity in the local population.
| In the visual system, decades of research have probed stimulus parameters that evoke responses in single neurons and these responses have been generally trial-averaged [1]. These response properties have revealed principles of functional organization in primary visual cortex (V1), such as orientation columns [2], and canonical computations, such as divisive normalization [3]. However, responses are variable across trials [4], making the relationship between perceptual stability and neuronal single-trial stimulus representations unclear [5]. The fluctuations of response strength are not independent across neurons, and this shared variability impacts population level representations of visual stimuli [6,7]. Neurons are highly interconnected and connection likelihood is biased toward spatially proximal neurons [8], suggesting that trial-to-trial response variability may be in part the manifestation of the state of the surrounding neuronal population [9,10]. Pairwise interactions within a population can shape information representation [11,12,13] and can be regulated by top-down influences [14]. Therefore, comprehensive descriptions of stimulus representations in primary sensory cortex require a network perspective. Here, we used two-photon imaging to record from large populations of L2/3 excitatory neurons in mouse V1 to study effects of local population activity on trial-to-trial variability.
Understanding the sources and consequences of response variability is necessary to extend theories of sensory computation from the average case to single trials. Perception and behavior take place in real time, after all, so variable responses must be taken into account to understand stimulus representations in cortex. Shared variability in neural responses is commonly quantified by the set of pairwise correlations between neurons, and the structure of these correlations can have constructive or destructive effects on stimulus encoding in populations of neurons [15,16], highlighting the importance of its characterization. Moreover, complex patterns of population activity in retina can be captured by taking into account only neuron firing rates and pairwise correlations [17]. Covariability can also be shaped by cognitive properties such as attention in order to improve perceptual acuity [14]. Whether trial-to-trial variability is harnessed to improve the fidelity of sensory representations, or accounted for when decoding from noisy signals, the properties of response variability have a large impact on neural function.
Research on the correlation structure of population activity is still incomplete, however, and can be meaningfully expanded by incorporating a more comprehensive sampling of the network [18]. V1 populations consist of neurons whose activity is not modulated by, or is untuned to, a given stimulus. It is still an open question how this subpopulation contributes to neuron correlations. Two-photon imaging results in a relatively unbiased sampling of spatially proximal neurons including both tuned and untuned subpopulations. Neurons unrelated to a behavioral task can help predict activity in neighboring neurons in hippocampal CA1 [9]. In V1, it has been shown that untuned neurons can help to decode the orientation of drifting gratings [19]. We investigate how co-fluctuations in the activity within tuned and untuned neurons interacts with responses to drifting grating stimuli.
We characterize population activity and correlations between tuned and untuned subpopulations in order to understand the relationship between single-cell response properties and recurrent network dynamics. Traditional noise correlation analyses study covariability independent from stimulus-driven activity [20]. However, untuned neurons have no stimulus modulation, so we use an analogous partial-correlation based method that additionally accounts for population-wide covariability. Fluctuations common across a local population are an important determinant of single-trial responses in mouse V1 [21,22]. Capturing this additional variable allows us to study the correlations in the entire unbiased sampling of the population. Additionally, the partial correlation matrices are asymmetric and relatively sparse, and can thus be represented as a weighted, directed graph. Graph theory analysis is used to summarize structure in complex networks and can resolve emergent properties resulting from pairwise relationships. Connectivity patterns, or motifs, in graphs can be characterized within small groups of neurons [23] or across an entire population [24,25]. Motifs have proved to be impactful for understanding complex biological processes including transcription networks [26] and spike propagation [23]. More generally, motifs patterns impact information representation in complex systems [27,28] and have increasingly been a subject of interest among neuroscientific disciplines.
From the graph neighbors of a given neuron, we can accurately predict activity on single trials using a simple, linear model. The local population contains information sufficient to predict trial-to-trial variability and recapitulates average tuning properties. Furthermore, neurons that are well-modeled by the activity of their neighbors have specific signatures of functional connection motifs. Across the entire graph, the most predictive motifs are also the most prevalent, suggesting that this structure is responsible for the overall quality of reconstruction observed. The triplet motif that facilitates predictions of neural activity may have a broad impact on information representation in graphs. Notably, total variance explained in a field of view scales with the number of neurons imaged, suggesting larger sample sizes are required to fully capture local network interactions.
To study interactions and response variability in local, cortical populations (<800μm diameter imaging plane), we imaged L2/3 excitatory neurons (72–347 neurons; 25–33 Hz; n = 8 animals; 23 distinct fields of view; Fig 1A) in mouse V1 during presentation of drifting gratings (Fig 1B and 1C). Square-wave gratings at 12 directions were presented in pseudo-random order for 5 seconds each, interleaved with 3 seconds of mean-luminance matched grey screen. We designed grating stimuli with slightly longer durations than many studies [16,29,30] for two reasons: first, to allow for the slow decay of the calcium indicator to fall to baseline in order to remove any confounds from the previous grating, and second, to study the sustained response in the population, rather than a transient response to stimulus onset [31]. Mice were awake and allowed to freely run on a linear treadmill. The majority of neurons showed significantly increased activity to one or more gratings over grey screen (3023/4535). Of the responsive subpopulation, most neurons were significantly tuned to orientation or direction (2073/3023; 540/3023 respectively). Neuron tuning was measured by fitting an asymmetric circular Gaussian tuning curve to the trial-averaged mean fluorescence in each grating direction (Fig 1D). These numbers of tuned and untuned neurons are in line with other population studies of awake, mouse V1 [16,30]. In subsequent analyses, we pooled all visually responsive neurons without significant direction or orientation tuning into a class of ‘untuned’ neurons, differentiating two distinct subpopulations in V1 by their responsivity to drifting gratings.
Single trial responses to gratings showed a high degree of variability, even in strongly tuned neurons, manifesting as occasional strong responses to null-directions and weak or absent responses to preferred directions (Figs 1C and 2A). Tuning curves described average response strength of tuned neurons well (0.70+/-0.20 R2). Responses across single trials in tuned neurons, however, were not well-described by their tuning curves. The mean fluorescence for each direction only explained a small fraction of the total trial-to-trial variance which we calculated by subtracting mean fluorescence in each direction from fluorescence in each trial (Fig 2B). This finding matches earlier results in awake, mouse V1 [30]. We computed the distribution of response strength (mean fluorescence across a grey or grating presentation) within single trials, z-scoring to account for neurons with different activity levels. Single trial response distributions were skewed, with most responses weaker than the mean (Fig 2C). Tuned neurons showed slightly stronger responses during gratings, as compared to untuned neurons which had nearly identical response distributions in grey and grating trials. Tuned response distributions were strongly overlapping, however, consistent with the hypothesis that individual neuron activity is not solely driven by tuning properties.
To further describe population activity, we computed the time-varying activity during the presentation of a grating and its preceding and following grey presentations. For each trial, we removed neurons that were silent (defined as no fluorescence change 2*S.D. above baseline) and computed the time-varying z-scored fluorescence across neurons (Fig 3A). Despite the long duration of stimulus presentations, adaptation effects were minimal in these L2/3 neurons, as tuned neurons showed sustained activity throughout the 5 second stimulus presentation. Untuned neurons have weak modulation to the onset and offset of the stimulus and are equally active during the grey period. However, these effects are small in comparison to variability across trials, as indicated by strong overlap between the activity of the two subpopulations. In the awake animal, running speed is known to strongly influence spike rates [32], and we similarly observe that periods of high population activity are very likely to occur during periods of running (Fig 3B, anecdotally in 1B). However, mice did not preferentially run during grating or grey presentations (probability of running 9.7+/-7.7% during gratings; 10.0+/-7.4% during greys; p = 0.278 paired t-test). While we used changes in fluorescence for all other analysis, we expanded our comparison of both subpopulations by estimating spike rates using a spike-inference from calcium fluorescence algorithm [33]. We found that untuned neurons exhibited an identical firing rate distribution to tuned neurons (Fig 3C). Therefore, differences between subpopulation dynamics cannot be explained by differences in firing-rates.
Untuned neurons are a large proportion of total neurons, exhibit similar spike rates to tuned neurons, and are likely to contribute to correlation structure in the population. To begin describing how dynamics are affected by stimuli in populations containing tuned and untuned neurons, we first analyzed pairwise correlations during grating and grey presentations. For each pair of neurons, we computed the correlation coefficient between the mean fluorescence (averaged over time) in either grating or grey trials. We did not remove signal-dependent responses, nor did we shuffle responses to eliminate simultaneous cofluctuations, therefore, these correlations are a combination of signal and noise correlations. Overall, within-subpopulation correlations are weak (0.014+/-0.027 tuned; 0.033+/-0.047 untuned), and beween-subpopulation activity is slightly anti-correlated (-0.019+/-0.017). Comparing mean pairwise correlations across all pairs according to their subpopulation, only within-subpopulation correlations are affected by grating stimuli, while between-subpopulation correlations are unchanged between stimulus and grey conditions (Fig 4A). Tuned neurons show a strong decrease in mean correlations during gratings, as seen in macaques [34]. Conversely, untuned neurons are more strongly correlated during gratings, and yet correlations between tuned and untuned neurons do not change in magnitude between stimulus and grey conditions. Untuned neuron activity is not directly modulated by the stimulus, so changes within this subpopulation most likely reflect changes in activity from the tuned subpopulation propagating through local synaptic connectivity. However, this occurs without a change in mean correlation between tuned and untuned neurons, bringing to question the mechanism involved. The pairwise correlations in tuned neurons during the stimulus are a function of their preferred grating directions, as expected for signal correlations (Fig 4B). Similarly tuned neurons show strong correlations, while orthogonally tuned neurons show negative correlations. This structure is not present during activity in grey periods, however. This is surprising, because if local connectivity underlies correlations in the grey condition, one should expect the structure seen during gratings to remain in part because similarly tuned neurons are more likely to be connected [35].
Overall correlations in populations including untuned neurons begins to reveal properties of local population activity, but in order to study the sources and structure of trial-to-trial shared variability, researchers attempt to remove the stimulus-dependent portion of responses leaving only variability, or ‘noise’ [20]. Correlated fluctuations between the remaining responses are therefore often called ‘noise correlations.’ However traditional noise correlation analysis was not appropriate in this case for the following three reasons: 1) the untuned neuronal subpopulation has no stimulus-dependent response; 2) in tuned neurons stimulus-driven response explains only a small portion of overall variability in tuned neurons (Fig 2B), making traditional noise-correlation analysis unsatisfactory for these neurons as well; 3) finally, V1 populations in mice, cat, and macaque are characterized by global cofluctuations common to every neuron [22,42,43] such as covariance driven by running (Figs 1B and 3B). Therefore in order to study pairwise noise correlations within and between subpopulations we used partial-correlation analysis that allowed us to account for stimulus-driven responses and population-wide co-fluctuations in both tuned and untuned neurons.
Visual stimuli were presented in 5-minute blocks. Each block of visual stimuli contained three repetitions of each direction in pseudo-random order and corresponding luminance matched grey periods. While the order of grating stimuli were pseudo random in each block the order was maintained between blocks. For each pair of neurons, the average activity across all remaining blocks, other than the block considered at that time, represented the stimulus-dependent responses capturing tuning properties, when present. Additionally, we accounted for the mean within-block population-wide activity of all remaining neurons. Consequently, we were able to compute a partial correlation coefficient in each block between the activity of every pair of neurons, controlling for stimulus responses and population co-activity (Fig 5A). The mean partial-correlation across blocks is taken as the final correlation strength and entered as an edge weight into the functional connectivity matrix. We also added directionality to the partial-correlation by examining the mean cross-correlogram across blocks for the neuron pair. If the peak value occurs at lag 0 (i.e. within the same imaging frame), the edge was bidirectional, otherwise the edge was in the direction of positive lag. Lags greater than 500ms were thrown out and correlations set to zero. This resulted in a functional network described by a directed weight matrix. Though we interpret these partial-correlations as equivalent to noise correlations, the correlation matrices are different from traditional noise correlations in two important ways. First, many pairs of neuron correlations are exactly zero (51.7+/-7.2%), and second, non-zero correlations are asymmetric (Fig 5B). This allows us to analyze these matrices from a graph-theoretic perspective representing the functional partial correlations as a weighted, directed graph. Graph representations of pairwise edges allow us to analyze population-wide statistical features of the correlation structure. Overall, partial-correlation strengths, synonymous with edge weights, were long-tailed, centered slightly above zero (Fig 5C), similar to noise-correlations observed elsewhere [36]. As expected, tuned edge weights were similar to signal correlations [36,37], with similarly tuned neurons having larger edge weights on average (Fig 5D).
We next analyzed the partial-correlations within and between tuned and untuned subpopulations. The graphs exhibit dense correlations with varying strengths among subpopulations (Fig 6A). To analyze biases in edge strengths, we thresholded the matrices at increasing values, setting all edges below each threshold to zero. Among all edges, within-tuned connections are more likely, while within-untuned and between-subpopulations are less common (Fig 6B; within-tuned 54.2+/-8.6%, within-untuned 44.1+/-6.0% between 43.9+/-6.2%). Between-subpopulation connections remain the least likely at higher thresholds, but among the strongest edges, within-untuned connections are the most likely. We then recomputed partial-correlation matrices, exclusively using frames during the grey condition or during grating condition to see how correlation strengths were affected, despite controlling for mean stimulus-dependent activity. The two resulting matrices did not have significantly different edge strength (stimulus 0.12+/-.02; grey 0.12+/-.03; p = 0.75 one-way ANOVA). However, probability of connection was higher within grey frames (grating 45.3+/-6.5%; grey 58.0+/-8.2%; p = 5.8*10^-7 one-way ANOVA), indicating a higher degree of interconnectivity in the population in the absence of stimuli. When analyzing magnitude of edge strength differences between grey and grating matrices, we found that all connection types changed similarly with a mean near zero (Fig 6C; within-tuned -0.011+/-0.097; within-untuned -0.005+/-0.103; between -0.008+/-0.093).
Since population dynamics are partly constrained by synaptic connectivity [38], we evaluated whether there was a spatial component to the weight matrices. Edge probability fell monotonically with distance between neurons, similar to traditional noise correlations [36] and synaptic connections [8,39]. Notably, the decay in connection probability is slower within tuned neurons compared to within untuned neurons (Fig 6D). Between-population decay lies in the middle. If the spatial structure of untuned correlations is exclusively driven by local connectivity, then bottom-up sensory drive is the most likely source for the longer range of functional correlations among tuned neurons. Furthermore, the mean lag (delay of the cross-correlogram peak used to determine edge directionality) is greater over longer distances and accumulated evenly across subpopulation connection type (Fig 6E). Assuming a linear change in lag over space, these data suggest the speed of functional correlations in this preparation is roughly 25 mm/s. Despite allowing for directional edges, nearly half of all edges were bidirectional (42.5+/-16.8%). As a function of edge strength, bidirectional edges are more prominent within-subpopulations, and are strongly biased toward the strongest weights (Fig 6F). Thresholding at increasing edge strengths sparsifies the matrices, so we normalized the bidirectional edge counts by the probability of bidirectional edges assuming connections are placed randomly. Among all edges, bidirectional edges occur less often than random. The strongest edges, however, are roughly 5 times more likely to be bidirectional than random. Overall, zero-lag connections are less frequent between tuned and untuned neurons, suggesting a transmission or propagation of information, rather than simultaneous representation of the same information between subpopulations.
Because trial-averaged tuning poorly captures trial-to-trial responses we asked whether information in the local population could better explain V1 trial-to-trial response variability. Pairwise correlations have been shown to capture a significant portion of the complexity in population activity [17]. We tested whether the activity of a neuron could be modeled from the activity of its neighbors that had a non-zero correlation. Since correlation coefficients capture the linear relationship between neuron coactivity, we used a simple linear combination of the input neuron activity with partial-correlation coefficients (edge strengths) as weights. This model gave a time-varying prediction of the fluorescence of a given neuron, which was then rescaled by an offset and a gain to account for different numbers of input neurons (Fig 7A). In many cases, this model resulted in highly accurate predictions of activity. Mean squared error of the reconstruction was small, and often near optimal compared to weights estimated by regression (Fig 7B). Tuned neurons were slightly better-modeled on average than untuned neurons (tuned MSE 0.014 median, 0.037 inter-quartile range; untuned MSE 0.017 median, 0.046 inter-quartile range), possibly because of the additional stimulus-dependent information captured by other tuned inputs. The ratio of MSE to mean-squared fluorescence from the true trace, subtracted from one, gives the percent variance explained of the model for each neuron. The optimal reconstructions from regression explained 77.2+/-15.2% of the variance in activity across all neurons (n = 4531). Using partial-correlation coefficients as weights performed near this upper-bound, at 65.8+/-17.1%. We tested the robustness of the partial-correlation based predictions modeling 5 minutes of fluorescence change, corresponding to one block, using a functional network built using 45 min, corresponding to the nine other blocks, of non-overlapping recordings to recompute the weights. We note that one block, i.e. five minutes of stimulus epochs, corresponded to three stimuli at each of the 12 directions and interleaved grey periods. We then tested our predictions on the left-out 5 minute dataset (repeated to leave out each dataset once). On this cross-validation procedure, the average performance on the left out epoch reached 55.4+/-18.9% variance explained.
We selectively removed either tuned or untuned input neurons to evaluate the relative contributions to model predictions. The accuracy of prediction decreased more when removing within-subpopulation inputs as compared to removing between-subpopulation inputs, but the decrease is small and distributions across neurons strongly overlap (Fig 7C). Within-subpopulation inputs are more frequent, however, so removing a larger fraction of inputs should be expected to have a larger impact on our ability to model neuronal fluorescence. On a single neuron basis, correlations between tuned and untuned neurons both contribute to predicting time-varying activity. We next asked whether our model based on local population activity could also predict trial-averaged tuning properties. For tuned neurons, we used the modeled fluorescence traces to recompute the mean fluorescence in each grating direction. The average responses in direction space were added together to obtain a mean tuning vector. The modeled fluorescence had very similar tuning properties to the data, as measured by the cosine similarity between model and data tuning vectors (Fig 7D). The modeled activity was constructed from a neuron’s input edges; we also asked if a neuron’s outgoing edges were related to tuning properties, which may indicate neurons that are good at decoding the stimulus (strong tuning) were also good at decoding activity in its neighboring neurons (strong outgoing edges). However, there was nearly zero correlation between the tuning strength of tuned neurons and its mean outgoing connection strength (r = 0.04, p = 0.02). Local population activity therefore contains information to capture trial-to-trial variability as well as trial-average stimulus response properties, but a neuron’s dependence on the stimulus does not affect its local population correlations.
To investigate how individual neurons inputs contribute to reconstruction of activity, we selectively removed input neurons based on their edge strength and measured the increase in reconstruction error (MSE), normalized by total mean squared fluorescence in the neuron. As expected, the strongest edges contribute the most to activity reconstruction with the strongest 25% of weights containing over half of the reconstruction capability, whereas removing half of the weakest edges had no discernible effect (Fig 8A). Interestingly, randomly removing edges does not linearly reduce prediction performance. This suggests a level of redundancy in the predictive information of input neuron activity. Accounting for the cumulative weight removed, we still found worse performance when removing the strongest edges first, and removing half of the total weight using only the weakest edges has minimal effect on reconstruction error (Fig 8B). Normalizing for the total weight removed reveals a nearly-linear increase in reconstruction error when removing the strongest weights, suggesting that these neurons may hold independent information from the remaining input pool.
To address the possibility of synergistic or redundant information between input neurons, we analyzed connections between triplets of neurons termed ‘motifs’ in graph theory literature. Triplet connection motifs are built up from pairwise connectivity and collectively represent higher-order connectivity patterns that cannot be captured by individual edges, and can have strong implications for computation and information propagation within graphs [27]. We looked at the clustering of triplet motifs for each type of triangle that can be formed with directed edges (Fig 8C) [40]. Clustering is a measure of how many motifs are present among a neuron’s neighbors given its input and output connections. In comparison to Erdos-Reyni graphs, which have uniform, low levels of clustering across motifs, cycles of edges in the data are less clustered on average, with all other motifs showing elevated clustering. The middle-man motif shows the strongest clustering. Similar results have been found in activity generated by simulated and ex vivo neural networks, although fan-in clustering was higher than middle-man [23]. To address the possibility that triplet motifs are responsible for explaining more of the neuronal response than pairwise edges alone, we analyzed the relationship between motif clustering in single neurons and the variance explained by the predicted activity. Because the magnitude of clustering is different across motifs, we first z-scored clustering coefficients, then computed the mean across neurons with different levels of variance explained. Neurons with the best reconstruction showed higher clustering of middle-man motifs and lower cycle clustering, relative to other neurons in the population (Fig 8D). Total clustering, as well as fan-in and fan-out, had weak, positive correlation with variance explained. Together, these relationships map directly onto the overall prevalence of the graph motifs, suggesting that the graph structure has an important function in representing population information.
Because the motif with the highest mean clustering was also most indicative of model performance, we hypothesized that the partial-correlation structure might underlie our ability to predict neuron activity from its local population. Interestingly, across fields of view, the total variance explained in the population increased with number of neurons imaged (Fig 8E; r = 0.58). To compute population variance explained, we took the sum of all squared prediction errors (residuals) across neurons in a field of view, divided by the sum of the squared population fluorescence activity, and subtracted from one. This improvement of variance explained across the population with more recorded neurons suggests that, in addition to motifs, total neurons sampled in the population determines our ability to measure a neurons’ single trial dependence on its local population. Moreover, the linear trend had no discernible plateau, so representing the local population may require recording from more than 300 neurons simultaneously.
We sought to describe the interrelationships within local populations of V1 neurons, including tuned and untuned neurons, as they relate to single-trial responses to grating stimuli. We used two-photon imaging to record from L2/3 excitatory populations constitutively expressing calcium indicator GCaMP6s. Neurons with similar response properties showed stronger co-variability on average, but across the entire population there was a broad distribution of correlations driven, presumably, by a confluence of sensory drive and activity in the local population. The functional correlations in the recorded populations were sufficient to predict activity in individual neurons, far surpassing predictions from tuning characteristics alone. The dependence on local population activity reinforces theories of layer 2/3 acting under strong modulation with sparse activity and weaker dependence on sensory drive than layer 4 [41]. We summarized the structure of correlations as directed, sparse matrices in order to analyze population dynamics from a graph theoretic perspective. We demonstrate that a simple population model capable of predicting single-trial neural responses is also able to accurately predict trial-averaged tuning responses, a key feature of V1 function. We found that the prevalence a specific triplet connectivity motif, built up from pairwise correlations, corresponded with our ability to predict activity on single-trials. This result could not have been observed from only studying pairwise correlations and motivates the continued use of graph theory to study neural population dynamics.
The single-neuron response properties in our data replicate imaging and electrophysiological results in awake recordings of V1 including proportion of tuned neurons [29,30,36], and variance explained by the mean tuning curve [30]. Both tuned and untuned neurons have low firing rates on average, though estimation of firing rates from calcium imaging is not always straightforward. We note that single neuron properties, including firing rate and trial-to-trial variance, can change substantially between animal models and in different states of anesthesia [42,43].
We found that the magnitude of signal correlations between tuned and untuned subpopulations do not change between grey and grating stimulus conditions, while within subpopulation correlations do change. Perhaps ubiquitous changes would be more expected, or exclusive changes among neurons modulated by the stimulus. Within subpopulation correlations change in opposite directions, however, and could serve as a mechanism to balance changes between the subpopulations.
We found that the spatial organization of the network is a strong determinant of correlation structure with correlations decaying over distance, consistent with paired patch clamp recordings [8] and the correlation structure of activity in isolated preparations [44]. Correlation matrices were computed as an asymmetric partial correlation coefficient, accounting for stimulus and population effects, to allow the incorporation of untuned neurons into traditional noise correlation analyses. While this approach differs from standard noise correlation estimates, the magnitude of correlations and dependence on tuning similarity replicate previous results measuring noise correlations [36]. For this reason, we interpret the partial correlations as measuring trial-to-trial covariability, and as being broadly equivalent to noise correlations. Tuned neurons, which combine feedforward sensory inputs with recurrent inputs, show a slower spatial decay of correlational values than untuned neurons, the latter of which presumably are driven more exclusively by recurrent, local inputs. These data show than subdividing the population by their response to grating stimuli can differentiate rates of spatial decay within the network.
Noise correlations are hypothesized to be driven in part by shared synaptic input [45]. Spatial dependencies of feedforward and recurrent connectivity can qualitatively change the spiking activity in network models [46]. The relatively small diameter of the fields of view imaged here (<1mm vs 10mm) does not allow us to differentiate between the two modes of activity predicted by these models. These data suggest, however, that functional recurrent correlations have less spatial extent than feed-forward functional correlations. To our knowledge, analogous synaptic connectivity estimates do not exist for mouse V1.
In addition to spatial structure of correlations, we found an increase in mean temporal delay of correlations over distance. This delay spans roughly 20–50 ms in our field of view and is therefore likely to reflect timescales of functional correlations rather than monosynaptic transmission delays. We have previously found that functional correlations are indicative of synaptic connections, in some cases, when considering synaptic integration rather than synaptic delays [23]. The implied speed of this delay accumulation is much faster than propagating LFP waves observed in macaque M1 [47] after normalizing for total cortical surface area [48]. The propagation of beta-oscillations and temporal delays of functional correlations likely have different underlying mechanisms, which may explain the different speeds.
We were able to extract a substantial amount of predictive information from the local population using a simple, linear model. Other approaches have successfully predicted single-trial responses from ongoing population activity at a larger spatial scale, averaging over many neurons in a population [49]. In contrast, we use a large, unbiased sampling of the local population with single cell resolution to predict variability on single trials. Alternative models incorporating known neuronal nonlinearities [50], or more sophisticated or biologically-relevant predictive models [51], may have improved performance. Consequently, the increase in total variance explained with population size may show different trends with alternative models. Nevertheless, we chose this modeling approach to maintain ease of interpretation and utilize the linear correlation coefficients estimated from the data in a straightforward manner. The linear models had good performance, and allowed us to remove input neurons without laboriously re-training the models. This straightforward modeling approach may not capture all of the information present in the local population, but its performance sets a lower bound on predictive information in the population. A small subset of neurons have high mean-squared error of predictions, and we speculate that we have not sampled enough of the population to predict these neurons given that total variance explained scales with population size. Neuron activity was only predicted from the in-degree neurons, rather than the entire population, reducing the number of parameters by 52+/-22% across neurons, though in either case, parameters scale by order N. These results substantiate previous models showing that the collection of pairwise relationships in large populations can explain complex activity patterns [17]. The predictions of single-trial activity that are obtained from local activity are, on one hand, a description of inter-neuronal dependencies in recurrent networks, and on the other hand, are capable of recapitulating trial-averaged tuning properties, extending these dependencies to stimulus encoding. We further explored this idea by identifying a specific motif of pairwise correlations underlying accurate predictions of neuron activity on single-trials.
The middle-man motif underlying the most accurate predictions is a specific pattern of pairwise correlations and represents a higher order feature of covariability. Local populations have been shown to contain such high-order correlations, but were not seen at larger spatial scales [52,53]; further experimentation is necessary to test whether functional triplet motifs occur across thousands of microns. The finding that middle-man motifs underlie the best predictions of activity may be initially surprising from the connection pattern of the motif. Compared to the fan-in motif, which has two input connections, the middle-man has one input and one output connection, and acts to route connections from its input to its target. However, the motifs were quantified using the clustering coefficient, which normalizes motif count by the total number of possible motifs (i.e. high fan-in clustering doesn’t correspond to high in-degree). The functional significance of any given triangle motif clustering is unknown and is likely dependent on the underlying system represented by the graph. In a neural network, middle-man clustering may indicate a hub-like property common in the brain [54], having a combination of convergent and divergent correlations. The cycle motif similarly has a combination of convergence and divergence, yet its clustering has a negative effect on prediction accuracy. The difference in the two motifs is the direction of connections between neighbors. In these networks, it may be that cycle clustering reflects recurrent, redundant correlations reducing our ability to predict activity from the population. Conversely, the middle-man motif is isometric to fan-in and fan-out motifs, and could allow for transfer of information between motifs, and in turn increase predictive power. Finally, this result demonstrates that in addition to providing insights into synaptic mechanisms underlying dynamics [23], network science can also provide insights into predictions of single trial neural responses as we have demonstrated here.
V1 is the first stage in which visual information is encoded in densely recurrent cortical networks. Thus, in order to understand activity patterns in V1, one must take into account visual drive as well as local network activity. We have provided a quantitative comparison of the relative influences of these two factors in awake, ambulating mice. Local network effects dominate on single trials, highlighting the importance of investigating cortical computation from a population perspective in order to understand how information is encoded in single trials. Populations of neurons exhibit emergent properties beyond the sum of their individual neurons and connections, and we use the analytic framework of graph theory to begin unraveling this emergent structure.
All procedures were performed in accordance and approved by the Institutional Animal Care and Use Committee at the University of Chicago. Data was collected from C57BL/6J mice of either sex (n = 4 female; 4 male) expressing transgene Tg(Thy1-GCaMP6s)GP4.12Dkim (Jackson Laboratory) between ages P84 –P191. After induction of anesthesia with isoflurane (induction at 4%, maintenance at 1–1.5%), a 3mm diameter cranial window was implanted above V1 by stereotaxic coordinates and cemented in place alongside a custom titanium headbar. Mice recovered for at least 8 days before intrinsic signal imaging to identify V1 followed by two-photon data collection.
Boundaries of V1 were identified by intrinsic signal imaging post-surgery [55] (Fig 1A, left]. Mice were anesthetized with isoflurane and head-clamped under a CCD camera (Qimaging Retiga-SRV). A vertical white-bar stimulus (100% contrast, 0.125Hz) was repeatedly presented on an LED monitor (AOC G2460) approximately 20cm from the contralateral eye while capturing cortical reflectance under 625nm illumination. The retinotopic mapping of V1 was then estimated at each pixel from the phase of peak reflectance driven by increases in activity-dependent blood flow.
Two-photon imaging was collected from awake, head-fixed mice on a linear treadmill. Running speed was measured with a rotary encoder attached to the treadmill axle. A L2/3 field of view (roughly 800μm diameter) in V1 was identified with galvanometer-mirror raster scanning (Cambridge Technologies; 6215H). Once a suitable field of view was found, raster scans (1Hz) were continuously acquired for roughly 10 minutes alongside visual stimulation. Neurons (n = 72–347 per field of view) were then automatically identified using custom image processing software for imaging during visual stimulation using Heuristically Optimal Path Scanning [56] at 25–33 Hz). All imaging was performed at 910nm (Coherent; Chameleon Ultra) with a 20X 1.1NA Olympus objective and GaAsP PMT (Hamamatsu; H10770A-40). Field of view size was estimated by fitting circles to a single raster scan of 15μm fluorescent microbeads and used for each dataset, though true field of view size may vary up to 8% across datasets from realignment of laser beam path.
Drifting grating stimuli were presented on an ASUS VG248QE, 20cm from the contralateral eye at 60Hz; 60cd/m2. The mean luminance was measured and gamma correction was performed and confirmed using a luminance meter. Square-wave gratings were shown at 80% contrast, 2Hz, 0.04 cyc/deg at 12 evenly spaced directions. Gratings were presented for 5s, interleaved with 3s mean-luminance grey screen. Three repetitions of each orientation were presented in a pseudo-random order, resulting in a roughly 5min stimulus movie. The grating order was preserved between movie presentations, and mice were shown 8–11 repetitions of the movie (24–33 repetitions of each direction).
Stimulus presentation was monitored with a photodiode (Thorlabs) and synchronized with running speed and imaging frames at 2kHz. For each neuron, baseline fluorescence was estimated from raw fluorescence by thresholding to eliminate spike-induced fluorescence transients and smoothed with a 4th-order, 81-point Savitzky-Golay filter. Fluorescence time-series were then normalized to percent change from baseline (dF/F0) using this time-varying baseline. Spike inference from fluorescence traces was performed using the OASIS algorithm [33] implemented using software made freely available (github.com/j-friedrich/OASIS). Inference outputs probability of spiking at each time point. As commonly done [57], probabilities were thresholded to obtain a binary spike train.
Neurons were classified as visually responsive if the mean response to any grating was significantly greater than the mean response across all grey periods by Dunnett-corrected one-way ANOVA (alpha = 0.01). In these analyses, each trial response is the mean fluorescence across the entire grating presentation (5000ms), or the last half of the grey presentation (1500ms) to allow for fluorescence from grating responses to decay. Responsive neurons were then tested for statistically significant orientation- or direction-tuned responses according to the trial vectors in orientation or direction spaces ([58] for more detailed methods]. For significantly tuned responses, tuning curves were then fit with an asymmetric-circular Gaussian to significantly tuned neurons. Tuning curve parameters (baseline, tuning width, peak amplitudes, and preferred direction) were fit repeatedly using randomized initial conditions. The parameter set that minimized mean-squared error was maintained.
For each neuron, the mean responses in each trial (using the same time windows as for tuning classification) for either stimulus or grey trials were used as response vectors. Typically, these response vectors had 360 elements (12 directions, 30 trials each). The correlation coefficient between each pair of these vectors was used to compute a pairwise correlation matrix for the grating and grey conditions. We did not shuffle responses, therefore these values measure the combination of signal and noise correlations.
For each pair of neurons, pairwise-correlation was computed as the mean partial correlation between their fluorescence across movies while accounting for three variables. This was computed with a built-in MATLAB function (partialcorr.m). We computed these correlations on time-varying traces of activity, rather than time-averaged trial activity as done in previous correlation analysis. Controlling for a single variable can be computed with the following equation, where rx,y is the correlation coefficient between time-varying fluorescence traces x(t), y(t), and rx,y|z denotes the partial correlation between x(t) and y(t), controlling for z(t):
rx,y∨z=rx,y−rx,zry,z1−rx,z21−ry,z2
Here, x(t) is the fluorescence trace of the ‘input’ neuron, y(t) is the fluorescence trace of the ‘output’ neuron, and z(t) represents the three control fluorescence traces. Controlling for more than one variable can be achieved by successive iterations of this procedure. This was computed for each movie, or 5-minute block of grating presentations that was repeated in each experiment. The first two control variables are the mean response of the two neurons in all other movies, accounting for the stimulus-driven responses. This is similar to normalizing by the mean response as in traditional noise correlation estimates. The third control variable is the within-movie mean fluorescence of all other neurons and was included to control for population-wide covariability, for example running speed effects. Furthermore, the cross-correlogram between the two neurons’ fluorescence traces, averaged across movies, was used to compute directionality of the correlation. The time-lag of the cross-correlogram global maximum determined the direction and lag of the edge. If the lag was zero, the correlation was bidirectional. If the lag was greater than 500ms (roughly 14 imaging frames), no edge was included.
The partial-correlation matrix could equivalently be analyzed as a directed, weighted graph. Open source software (Gephi) was used for visualization, with node layout determined by the Yifan-Hu algorithm and tuned by hand. Edge weights less than 0.05 were set to zero for visualization clarity. Erdos-Reyni (ER) null graphs were generated for each graph to match the mean connection probability. The mean directed clustering coefficient across nodes was calculated across 50 ER graphs and averaged for comparison with data. Clustering coefficients were computed with binary matrices (nonzero weights set to one).
To model neuron responses, we used a linear weighting of the fluorescence of every in-degree using the weights in the partial-correlation matrix. At each time point, a weighted sum was calculated, resulting in a time-varying predicted fluorescence trace. Because different numbers of input neurons and varying calcium transient amplitudes, the modeled trace was then fit with a linear offset and a gain to minimize mean-squared error with the true fluorescence trace. These two parameters were not changed when input neurons were removed (7C, 8A, 8B). For tuned neurons, we also recomputed the trial-averaged tuning response of modeled activity. The fluorescence over a grating presentation was averaged, then trials were averaged over directions to obtain a mean direction-response. The sum of these vectors in direction space gave the model-estimated tuning vector, and the cosine similarity with the data-derived tuning vector was used to quantify the reconstruction of the tuning properties. Cosine similarity was computed in direction or orientation space according to each neuron’s tuning properties. To compute total population variance explained, modeled traces were subtracted from the data traces to obtain residuals, and the ratio of total sum of squares across neurons were subtracted from one as
1−∑i∑t(Xi(t)−X˜i(t))2∑i∑t(Xi(t))2
where Xi(t) is the time-varying fluorescence trace of neuron i, and Xi(t) tilde is the predicted trace.
Optimal weights for all incoming edges were computed for each neuron by LASSO regression as
minβ,β0(12N∑i=1N(yi−β0−xiTβ)2+λ∑j=1p|βj|)
For weights β, and offset β0, with modeled neuron fluorescence as y and input neuron activity as x. Weight estimation and 5-fold cross validation to estimate MSE standard error was performed with MATLAB R2016a implementation. The maximum regularization parameter (λ) whose mean-squared error did not exceed the standard error of the minimum MSE was used to find the set of optimal weights.
|
10.1371/journal.ppat.1003755 | Infectious Prions Accumulate to High Levels in Non Proliferative C2C12 Myotubes | Prion diseases are driven by the strain-specific, template-dependent transconformation of the normal cellular prion protein (PrPC) into a disease specific isoform PrPSc. Cell culture models of prion infection generally use replicating cells resulting in lower levels of prion accumulation compared to animals. Using non-replicating cells allows the accumulation of higher levels of PrPSc and, thus, greater amounts of infectivity. Here, we infect non-proliferating muscle fiber myotube cultures prepared from differentiated myoblasts. We demonstrate that prion-infected myotubes generate substantial amounts of PrPSc and that the level of infectivity produced in these post-mitotic cells, 105.5 L.D.50/mg of total protein, approaches that observed in vivo. Exposure of the myotubes to different mouse-adapted agents demonstrates strain-specific replication of infectious agents. Mouse-derived myotubes could not be infected with hamster prions suggesting that the species barrier effect is intact. We suggest that non-proliferating myotubes will be a valuable model system for generating infectious prions and for screening compounds for anti-prion activity.
| This manuscript describes the generation of a new cell culture system to study the replication of infectious prions. While numerous cell lines exist that can replicate prions, these systems are usually based upon proliferating cells. As mammalian cell cultures double approximately every day, prions established in the culture must also, at least, double to be maintained. This is problematic, however, as prions replicate relatively slowly and cell replication may outpace prion replication. In fact, many cell culture systems do not replicate prions and those that do often do not replicate all strains of prions. Here we describe the use of differentiated non-proliferative muscle cells to replicate prions without the interfering effect of cell division. We observed that prions accumulate to very high levels in this muscle cell culture with infectivity approaching that observed in animals.
| Prions are the etiological agents responsible for the transmissible spongiform encephalopathies. These neurodegenerative diseases affect mammals, are inevitably fatal, and are always associated with the accumulation of a specific post-translationally modified isoform of a normal host glycoprotein, PrPC. This abnormal conformation, the PrPSc isoform, differs from PrPC structurally, resulting in dramatic functional consequences. While PrPC is readily degraded by proteases, soluble in detergent and rich in alpha-helical structure, PrPSc is typically characterized by resistance to proteinase K (PK) digestion, detergent insolubility, amyloid formation and β-sheet structure. In vivo the disease-specific isoform of the prion protein (PrPSc) accumulates to high levels, a process that is marked by a progressive neurodegeneration that is always fatal as well as the generation of hundreds of millions of lethal doses of transmissible prions.
Experimentally, prion infections are typically performed in rodents: wild type mice, transgenic mice or hamsters. Incubation periods are generally “short” (compared to prion infections of cervids, sheep or cattle), ranging from two months in hamsters and certain transgenic mouse lines to greater than a year for other mouse strains and agent strain combinations. Brain infectivity levels are extraordinarily high at clinical stage, 109 50% lethal doses (LD50) in end stage hamster brain and 108 LD50 in mouse models.
In vitro models of prion replication have been established by incubating infectious brain homogenates with various susceptible cell lines, most of neuronal origin, and all expressing PrPC, obligatory as a source for PrPSc generation. Typically, dividing cells are exposed to brain homogenates, derived from infected mice, and the cells are then serially passaged until the inoculum is diluted out. PrPSc accumulates to a steady state determined by the accumulative effect of prion replication, the dilutive effect of cell division and subsequent passaging, prion secretion into the media [1], [2] and prion degradation [3]–[6]. In vitro replication of prions has been observed in numerous cell types including scrapie mouse brain cells [7]–[9], fibroblasts [10], [11], epithelia [12], glia [13], [14], microglia [15], PC12 [16], Schwann cells [17], hypothalmic neurons [18] and neuron-like cells [19]. By far, however, the most widely used cells for in vitro replication of prions are mouse Neuro-2a (N2a), neuroblastoma cells [13], [20]–[22]. Prion infection of cell cultures typically results in relatively low levels of PrPSc and infectivity being generated. In the N2a cells, the level of infectivity is very low, ∼3×103 LD50 per 1×107 cells [20], [21]. When infected neuroblastoma cell lines are subcloned and highly susceptible sublines are isolated however, the infectivity can increase to ∼2×104 LD50 per 1×107 cells [23]. Alternatively, highly susceptible N2a sublines can be isolated and subsequently infected [24], allowing for cognate uninfected cells to be propagated as controls.
One difficulty in generating and maintaining in vitro cultures of prion infection is that the infectivity levels are low and some species and strain combinations do not result in infection or stable infection [24]–[27]. Cell lines that divide rapidly tend not to support prion replication, presumably due to the dilutive effect of cell replication [28]. One potential means of overcoming these effects is the use of post-mitotic, differentiated cells for studies of prion replication. Murine-derived C2C12 myoblast cells [29] provide an intriguing possibility as myoblasts are proliferative but following serum deprivation, terminally differentiate into post-mitotic myotubes, a syncytium of fused myoblasts. Muscle expresses relatively high levels of PrPC [30],which promotes muscle regeneration in vivo [31], and can harbour and replicate prion infectivity [32], [33].
We report that differentiated non-proliferative myotube cultures can replicate prions to surprisingly high levels. Previous work with this cell culture system only observed infection, with 22L prions, when C2C12 cells where co-cultured with susceptible neuroblastoma cell lines [34]. Our studies have focussed on the infection of myotubes, not myoblasts, an approach that may prove useful as the terminally differentiated cells do not divide, removing the dilutive effect of passage in assays of scrapie replication. This system more closely mimics the in vivo situation, where a less dynamic population of cells accumulates PrPSc.
Proliferative myoblasts are capable of undergoing terminal differentiation into muscle fiber-like myotubes (Figure 1a). Spontaneous differentiation occurs at high cell density and after serum withdrawal. Fully differentiated myotubes are multinucleated, contain sarcomeres and can contract. Monolayers of myotubes can remain intact for weeks. Importantly, as a cell culture system for PrPSc replication, myotubes express approximately one fifth as much normal prion protein (PrPC) as brain normalized per mg protein and an equivalent amount to N2a neuroblastoma cells (Figure 1b) though N2a cells are known to be variable in their characteristics [35]. Slightly higher levels of PrPC are consistently observed in C2C12 myotubes compared to myoblasts. PrPC expressed in myoblasts is predominately di-glycosylated.
To examine the replicative potential of prions in a terminally differentiated muscle cell line and contrast them with non-differentiated replicative muscle precursor cells, we exposed both C2C12 myoblast and myotube cultures to RML prions (Figure 2). C2C12 cells were propagated and expanded as myoblasts, seeded into individual wells or flasks, grown to confluence and differentiated into a layer of myotubes upon serum withdrawal. Myotubes can easily be infected with brain homogenates by diluting infected brain homogenate into the media. Myoblast cultures were infected as sub-confluent proliferating cultures, splitting as necessary to prevent differentiation. By passage nine, fifteen days after infection, accumulation of PrPSc was not detectable in myoblast cultures (Figure 2b) consistent with the results of Dlakic et al. 2007 [34]. The infection of differentiated myotubes gave a very different result. We routinely observed accumulation of newly generated proteinase K resistant PrPSc by 10 days post-infection in myotubes (Figure 2a). Levels of PrPSc subsequently increased with time.
To determine the titre of myotube-generated PrPSc, aliquots of cell homogenates were used to infect C57Bl/6 mice. To generate material for the bioassay, we infected T75 flasks of confluent myotubes with RML prions and harvested cells at indicated time points (Figure 3). 30 µL of this lysate was inoculated, intracerebrally, into weanling C57Bl/6 mice. A standard curve for infectivity was performed in parallel with mice being inoculated with serial 10-fold dilutions of RML brain homogenate (Figure S1). Mice inoculated with the infected C2C12 cell lysate (15 days post-infection) had an incubation period of 181.7 days. This suggests a level of infectivity approximate to a 0.3% brain homogenate from mice clinically affected with RML. Between 4- and 15- days post infection, the amount of infectivity present in C2C12 myotubes increased ∼10 fold and a statistically significant difference in incubation period was observed (Figure 3). C2C12 myotubes accumulated 105.5 LD50/mg protein (Figure 3B) as calculated based upon comparison of incubation periods from C2C12 derived material compared to standard 10-fold dilutions of RML brain homogenate (Figure 3A). 10% RML brain homogenate contained 107 LD50/mg protein (Figure 3B).
Cell cultures containing prion-infected and mock-infected myotubes were observed daily. Over the time course of the experiments (up to 21 days), no cell loss was observed in prion-infected plates as compared to mock-infected. To examine the cells for molecular responses to prion infection, gene expression profiling was performed (Figure 4). Expression profiles from infected and uninfected cells were similar and no genes were up-regulated greater than 2-fold by prion infection. Two genes were downregulated by >2-fold by prion infection, Carbonic anhydrase 3 and 2310046K23Rik, an un-annotated transcript. No genes were differentially regulated to an 85% confidence interval as assessed by T-test and a FDR based approach [36]. Subsequently, we examined those genes which were up-regulated in C2C12 cells infected with prion disease by submitting the “top 20” genes to the Prion Disease Database (http://prion.systemsbiology.net/) [37], [38] for which annotations were known. The genes were up-regulated in infected tubes between 1.36 and 1.6 fold. Four of the genes in the “top 20 list” (Iigp1, tgfb1, ifit1, Aldh1l2) were found to be up-regulated in brains of prion-infected mice suggesting that the changes observed might contain a prion-specific signature.
To compare PrPSc levels in prion-susceptible mouse cell lines, following exposure to RML agent, total protein (100 µg) collected from infected cell lysates and from infected brain homogenates was subjected to PK digestion (50 µg/mL final concentration). Strong signals were observed in the prion-infected C2C12 samples and the brain homogenates. Weaker PrPSc levels were observed in the N2a and SMB cells (Figure 5). To semi-quantify the differences in signal intensity, we digested and loaded ten times more total protein from infected myotubes than RML brain homogenate and ten times more N2a and SMB cell protein than myotube protein. These data indicate the PrPSc generation in C2C12 cells is considerably more robust than the other cell culture systems and within an order of magnitude of RML-infected brain homogenates.
To examine whether known inhibitors of PrPSc replication could be used to “cure” prion-infected myotubes, we applied pentosan polysulfate to infected C2C12 myotubes. PPS is a sulfated polyanion previously identified to inhibit prion accumulation in cells [39], [40], prevent scrapie following intraperitoneal inoculation [41] and some therapeutic effect has been observed in human prion diseases, reviewed by Rainov et al., 2007 [42]. In our studies, replicate infected myotube cultures were treated with or without PPS (1 µg/mL final concentration) and harvested at specific time points. Equal amounts of total protein were subjected to PK digestion and immunoblotting (Figure 6). Without PPS, PrPSc abundance increased with time, whereas in cells treated with PPS, PrPSc accumulation was inhibited and the PrPSc signal substantially diminished by 16 days post-infection.
To examine the strain selectivity of the C2C12 myotube system, we infected differentiated myotube cultures with 3 different mouse prion strains (RML, 22L and ME7). Myotube cultures were infected in parallel and analyzed by immunoblotting for the presence of PrPSc. All strains examined replicated prions in this myotube system (Figure 7), albeit with apparent strain specific kinetics. Signal at day 4 is carryover from the infection. The absence of signal at day 8 in 22L and ME7 would suggest that these strains replicate slower than RML, and the weak signal at day 14 in 22L would suggest that this strain replicates the slowest in C2C12 myotubes. The characteristic shift in glycosylation pattern to C2C12-like diglycosylated PrP is apparent.
In vivo transmission of prions between different species is typically associated with a species barrier characterized by low penetrance and extended incubation period upon first passage. To examine the C2C12 model as an in vitro surrogate for examining interspecies transmission, we exposed the C2C12 myotubes to the Hyper (HY) strain of hamster-adapted transmissible mink encephalopathy prions at >109 LD50/mL (Figure 8). While hamster PrPHY appeared to persist in C2C12 myotube culture for 5 days, no mouse PrPSc was detected as a result of infection with HY prions. Control experiments using RML prions readily established PrPSc accumulation. In similar experiments using cervid prions, no conversion of mouse PrPC was observed (data not shown).
We have demonstrated that non-replicative myotube cultures can readily be infected with mouse prion agent. PrPSc replication occurs in a short period of time, is robust, and levels of infectivity are relatively high. Input mouse brain homogenate-derived PrPSc, evidenced by dominance of the mono-glycosylated band, routinely became undetectable after 5 days post-exposure (Figures 2, 5, 6, 7, 8). As media was changed daily, it is remains unclear whether the input PrPSc was actively degraded by the cells or diluted out by media changes. Three lines of evidence suggest that the C2C12 myotubes are replicating prions: 1) in scrapie mouse brain homogenate, the inocula used on our cells, the dominant band of PrPSc is monoglycosylated, whereas the C2C12-generated PrPSc is predominantly diglycosylated (Figures 2, 5, 6, 7, 8). The shift in the glycosylation pattern allows us to discriminate between input material and de novo generated material (Figure S2); 2) HY TME prions were only detectable at 5 days post-exposure suggesting agent does not persistence (Figure 8); and 3) bioassay of cell lysates demonstrated an increase in infectivity over the course of the cell infection (Figure 3). Therefore, we conclude that the PrPSc observed is due to replication and not persistence.
Although this is the first study to establish prion infection in C2C12 myotubes, we are not the first to attempt infection of C2C12 cells. Dlakic et al. established infections of C2C12 cells with 22L prions but only when co-cultured with susceptible N2a cells [34]. We were unsuccessful in infecting myoblasts but consistently infected myotubes directly, i.e., without co-infection with N2a cells. The successful infection of myotubes with 22L, ME7 and RML suggests that the differences observed are a result of both prion strain used for infection and/or the differentiation state of the C2C12 cells. Infection of myotubes with RML and ME7 was more robust than with 22L and we were only successful in infecting completely differentiated myotubes.
Typical proliferative cell cultures are constantly dividing and must be split. This means that, for prion infections to be maintained in culture, PrPSc replication must outpace cell replication and degradation [28], [43]. This is especially important considering that cell lines replicate at different rates. We observe that C2C12 myoblasts double every 18 hours. Contrast this cellular replication rate with in vivo accumulation (replication and degradation) of prions, a process that can take years to fulminate in clinical disease. The replication rate of cell lines may explain the differential susceptibility of different cell lines [11], [22], [24], [44]–[47] or clonally-derived sub-lines [24] to infection with different prion strains. In C2C12 cells, we observed replication of 3 different strains of prions, albeit with different levels of PrPSc at 14 days post-exposure. Another issue with proliferating cell lines is the potential genomic instability of these cells. For example, the median chromosome number of the N2a neuroblastoma cell line is 95 and the range is 59–193 (ATCC datasheet, CCL-131). By contrast, C2C12 myoblasts are diploid [48]. It is clear that stochastic genetic drift of chromosome number over generations of cell culture could cause changes in cell replication rate and/or PrPC expression, both of which may affect PrPSc accumulation or molecular responses [35], [49]. Finally, proliferative adherent cell cultures are routinely passaged by trypsinization. Since PrPC is expressed on the cell surface, and cleaved by trypsinization, passaging cells may result in a temporary decrease of PrPC required for conversion. Myotube cultures, by contrast, are relatively static allowing direct comparison between parallel plates where the influence of cell division (replication rate, changes in chromosomal abnormalities, effect of trypsinization, etc) is removed.
Myotubes can accumulate substantial PrPSc (Figure 5) and associated infectivity (Figure 3). We routinely observe robust PrPSc signals from infected myotubes when loading only 10 µg of PK-digested total protein equivalent. Our data suggest that ∼10 fold higher levels of PrPSc are being generated in C2C12 myotube culture than in common N2a cells. While we observed substantially higher levels of PrPSc in C2C12 myotubes compared with chronically infected N2a cells, the heterogeneity of N2a cells with respect to PrPSc replication cautions against concluding with respect to PrPSc levels in general. Many different sublines of N2a cells can be isolated some which fail to replicate PrPSc and others which are quite prolific. Subclones of infected N2a cells accumulate 1 LD50 per 158 cells [50] or per 500 cells [21], suggesting a titre of approximately104 LD50 per 1×107 cells. A T75 flask of C2C12 myotubes at 15 days post infection contains 106 LD50; as C2C12 myotubes form a monolayer of non-proliferative multinucleated myofibers, titre cannot be expressed as LD50 per cell, however, a confluent 75 cm2 flask of non-differentiated C2C12 myoblasts contains ∼1×107 cells. This result is comparable to that observed in differentiated PC12 cells [44], [51] suggesting a similar conclusion, that differentiated non-replicative cells can accumulate prion infectivity to high levels. Importantly, C2C12 myotubes are not of neuronal origin, implying that the non-replicative differentiated state, and not the neuronal state, is the critical factor for allowing high levels of infectivity to accumulate in both C2C12 and PC12 cells.
We did not detect overt effects of prion infection in C2C12 myotubes as assessed by cell morphology or gene expression profiling despite the accumulation of appreciable levels of PrPSc and infectivity. This apparent lack of cell toxicity has been observed in most cell cultures systems [49] for ex-vivo PrPSc replication, excepting GT-1 cells [18]. The accumulation of high levels of infectivity in C2C12 myotubes and PC12 cells absent any observable cytopathic effect [51] suggests that PrPSc itself is generally non-toxic to cultured cells; the cytoxicity appears to be specific to neurons in the central nervous system [52]. This strongly suggests that the apparent lethality of prion disease must be due at least in part to the multiple cell types or complex architecture present in the CNS.
To examine the potential of myotubes to serve as a model system for assay of prion inhibition, we tested pentosan polysulfate (PPS), a well-established molecule with known anti-PrPSc replication properties, for its ability to hinder PrPSc accumulation in the myotubes. PPS inhibited PrPSc replication in the myotubes (Figure 6) but its efficacy was less than that observed in proliferative systems where PPS treatment and cell passage can completely remove infectivity [39], [40]. It is likely that, in cultured replicating cells, the dilutive effect of cell passage coupled with the PPS inhibition of PrPSc replication enhances the apparent effect of PPS. This observation supports the suggestion by Weissmann et al. [6] of that current proliferative cell-based models for assay of inhibition of PrPSc replication are inadequate. In vivo, there is no comparable dilution of PrPSc to that created by the serial passage of proliferative cell cultures. Treatment of infected myotubes with PPS also extended the time that inocula, as indicated by dominance of mono-glycosylated PrPSc, persisted in the culture. A strong signal from carry-over inocula is present at day 4 in PPS treated cells and seems to persist longer (Figure 6b) than in myotubes without PPS (Figure 6a). One possibility is that PPS enhances the binding of PrPSc to cells preventing it being washed out by media changes. Alternatively, PPS may interfere with degradation of PrPSc.
In vivo, transmission of hamster HY or white-tailed deer CWD prions to mice is not efficacious and is associated with a large species barrier[53], [54]. We examined our cell-based assay for prion replication fidelity by challenging C2C12 myotubes with hamster (Figure 8) and deer prions and were unable to replicate prions from either species, consistent with the normal host range of these agents. This suggests that C2C12 myotubes may be a good model system with which to probe species barriers and prion adaptation, at least with regard to mouse prion protein.
In summary, we have developed a novel, non-proliferative cell culture system that replicates prion infectivity to high levels generating substantial amounts of protease-resistant prion protein. This approach may be useful for probing prion strain and species barrier phenomena. Moreover, this system may allow a better assessment of putative anti-scrapie compounds as it removes the confounding effect of cell-replication from that of prion replication. We are currently adapting the myotube system for use with other prion agents.
This study was carried out in accordance with the recommendations and guidelines of the Canadian Council on Animal Care under protocol 647/10/11C and was approved by the Institutional Animal Care Committee at the University of Alberta.
All animal manipulation and care was performed under institutionally approved animal use protocols approved by the University of Alberta Animal Care and Use Committee. Preparations for bioassay were formulated/diluted in sterile water. 30 µL of each preparation was injected into the anterior fontanelle of weanling C57Bl/6 mice. Animals were scored weekly for the onset of clinical disease.
Prion preparations were obtained by homogenization of brain tissue in Dulbecco's phosphate buffer or water. Brains were collected from animals afflicted with end-stage clinical prion disease, infection was confirmed by proteinase K digestion and western blotting. A 10% (w/v) homogenate prepared from a pool of RML affected brains was used for this work. This pooled homogenate was minimally 109 LD50 per gram of brain as determined by bioassay using the Kärber formula [55]. Infectivity of C2C12 samples was calculated from the regression equation (Figure S1) derived from plotting the incubation period observed from inoculation of standard 10-fold dilutions of RML brain homogenates versus infectivity; time interval assay [56]. 22L and ME7 prions were prepared from clinically-affected tga20 mice [57] and contain approximately 108 LD50 per mL [58].
C2C12 (CRL-1722) myoblast cells were purchased from the American Type Culture Collection, expanded and aliquots stored in liquid nitrogen and expanded as needed. Myoblasts were grown in Dulbecco's Modified Eagle Medium, 10% fetal bovine serum (FBS) with penicillin and streptomycin (PS). Myoblasts were seeded onto experimental plates, differentiated when confluent by switching to differentiation medium; DMEM, 10% horse serum (HS), and PS. Three days after myotubes first appeared, infections were initiated by the addition of prion-infected brain homogenates diluted 1∶100 in differentiation media (DMEM, 1% HS, 1% PS). A typical infection experiment involved treatment of ∼1×107 cells with ∼1×107 LD50 infectious prions, 100 µL of 10% brain homogenate per 10 mL of media. See Methods S1. Media was changed daily, washing cells and removing residual inocula.
Cell lysates were prepared by removing media, washing the cells with PBS and then adding RIPA lysis buffer. Total protein concentration was determined by BCA assay. Typically, 100 µg of total protein was digested with 3.5 µg of Proteinase K (PK) (Roche) for 30 minutes at 37°C in a volume of 70 uL (50 mg/ml PK final concentration). Digestion was terminated by addition of 30 ul of 5X SDS sample buffer. Each sample was loaded (10–15 µg based on the pre-digestion concentration) and resolved on 15-well 12% NuPAGE bis-Tris gels (Invitrogen), transferred to PVDF membrane and probed with anti-PrP antibodies 3F4 (a kind gift from Richard Rubenstein) epitope (107–112) [59], 3F10 (a kind gift from Yong-Sun Kim) epitope (137–151) [60] and/or SAF-83 (Cayman Chemical) epitope (126–164) [61]. Relative quantification of western blotting was performed by loading dilutions of samples until quasi-equivalent signals were obtained on western blots as determined by image analysis software (Adobe Photoshop).
Gene expression profiling was performed as described [62]. Briefly, RNA was purified from cell pellets using the QIAshredder and RNeasy mini kit (Qiagen, Valencia, CA) in accordance with the manufacturers' instructions. Total RNA was amplified and labeled in preparation for chemical fragmentation and hybridization with the MessageAmp Premier RNA amplification and labeling kit (Life Technologies, Grand Island, NY). Amplified and labelled cRNAs were hybridized on Affymetrix (Santa Clara, CA) mouse genome 430 2.0 high density oligonucleotide arrays. Raw data were analyzed with Arraystar 5.0 (DNA Star, Madison, WI). Robust multiarray normalization using the quantile approach was used to normalize all microarray data. Data are deposited into the National Center for Biotechnology Information Gene expression omnibus database with accession number GSE44563.
Prion Protein (MGI:97769), 2310046K23Rik (MGI:1924218),Carbonic Anhydrase 3(MGI:88270), Iigp1 (MGI:1926259), tgfb1(MGI:98725), ifit1 (MGI:99450), Aldh1l2(MGI:2444680)
|
10.1371/journal.ppat.1004265 | Pathogenicity of Mycobacterium tuberculosis Is Expressed by Regulating Metabolic Thresholds of the Host Macrophage | The success of Mycobacterium tuberculosis as a pathogen derives from its facile adaptation to the intracellular milieu of human macrophages. To explore this process, we asked whether adaptation also required interference with the metabolic machinery of the host cell. Temporal profiling of the metabolic flux, in cells infected with differently virulent mycobacterial strains, confirmed that this was indeed the case. Subsequent analysis identified the core subset of host reactions that were targeted. It also elucidated that the goal of regulation was to integrate pathways facilitating macrophage survival, with those promoting mycobacterial sustenance. Intriguingly, this synthesis then provided an axis where both host- and pathogen-derived factors converged to define determinants of pathogenicity. Consequently, whereas the requirement for macrophage survival sensitized TB susceptibility to the glycemic status of the individual, mediation by pathogen ensured that the virulence properties of the infecting strain also contributed towards the resulting pathology.
| Mycobacterium tuberculosis (Mtb) is a highly successful human pathogen, representing the leading bacterial cause of death worldwide. Mtb infects macrophages and it adapts to the hostile intracellular milieu of this cell by exploiting the plasticity of its central carbon metabolism machinery. While several studies have detailed the bacterial adaptations that accompany infection, it is still unclear whether this process also involves engagement with host metabolic pathways. We therefore profiled the kinetic flux of host cell metabolites in macrophages that were infected with differently virulent Mtb strains. Interestingly, we found that Mtb pathogenicity was indeed intimately linked to its capacity to regulate host cell metabolism. A unique subset of host pathways was targeted so as to integrate the glycolytic threshold governing macrophage viability with mechanisms ensuring intracellular bacterial survival. Perturbation of macrophage glycolytic flux was enforced through pathogen-induced enhancement in glucose uptake, which in turn was also influenced by the extracellular glucose concentration. This observation rationalizes the increased susceptibility of diabetic individuals to TB infection Interestingly, Mtb strains also differed in their capacities to stimulate macrophage glucose uptake. Consequently, the resulting pathology is likely dictated both by the individual's glycemic status, and the nature of the infecting strain.
| Pathogenicity of Mycobacterium tuberculosis (Mtb) has been attributed to the plasticity of its central carbon metabolism (CCM) machinery, which facilitates ready adaptation of pathogen to the intracellular milieu of the macrophage [1]. Emerging evidence, however, suggests that Mtb pathogenicity is also supported by engagement with metabolic pathways of the host cell. Thus while Mtb adaption to host requires the switch in bacterial CCM towards catabolism of host lipid substrates [2]–[5], optimal exploitation of this switch involves pathogen-induced promotion of lipid body (LB) accumulation by the host macrophage [6]–[8]. This ensures an abundant supply of the lipid substrates.
To investigate how Mtb infection influences CCM of the host macrophage we examined time-dependent modulations in macrophage metabolism, after infection with mycobacterial strains that varied in both genotype and phenotype. The resulting data, describing the temporal effects against a gradient of mycobacterial virulence, confirmed that Mtb pathogenicity was indeed intimately linked to its capacity to regulate host cell metabolism. Further, we also discovered that expression of virulence required the pathogen to engage with a unique subset of host metabolic pathways. Characterization of these pathways revealed that metabolic thresholds governing host cell survival were tightly assimilated with mechanisms regulating intracellular survival of the bacilli. This synthesis then provided the framework for convergence of both host- and pathogen-derived factors, in dictating the pathology of Mtb infection.
For these studies we primarily employed PMA-differentiated macrophage like THP-1 cells. To characterize the CCM of these cells, and also subsequently delineate the effects of Mtb infection, we adopted the procedure of kinetic flux profiling. In this procedure cells are fed with an isotopically labeled carbon source, followed by a determination of the rates at which this label is then incorporated into the downstream metabolites [9]–[11]. Kinetic profiling possesses the advantage of being sensitive to even subtle effects on host cell metabolism. Importantly for our purpose, it also circumvented the potential complication of host metabolites being contaminated with contributions from the pathogen either as a result of export or diffusion, or, simply leakage from bacilli during the extraction procedure. This could be confirmed in initial experiments where cells infected with the Mtb strain H37Rv were labeled with 13C6-glucose, and label incorporation into both host- and bacterial-derived metabolites was compared. We found that the labeling rates of host metabolites were between 40- to >100-fold higher than that of the corresponding bacterial counterparts (Figure S1). Thus, kinetic profiling enabled us to specifically monitor metabolite flux in the host cell, without interference from bacterial products or processes.
Before examining the effects of infection, it was necessary for us to first characterize the metabolic flux in uninfected THP-1 cells. For this we pulse-fed cells with 13C6-glucose for various times, and then employed liquid chromatography-tandem mass spectrometry (LC-MS/MS) to measure rates of label incorporation in a total of thirty-five metabolites derived either from the glycolytic, the pentose phosphate (PP), or nucleotide flux pathways. In addition, the tricarboxylic acid (TCA) cycle and the citrate shuttle pathway leading to lipid biosynthesis were also followed. Rapid incorporation of 13C-label was detected in glucose-6-phosphate (G6P), the first intermediate in glycolysis, with its subsequent utilization then being distributed between the glycolytic and PP pathways. The latter however primarily fed back into glycolysis at either fructose-1,6-bisphosphate (FBP) or glyceraldehyde-3-phosphate (G3P) since, apart from ribose-5-phosphate (R5P), no labeling of downstream intermediates in either nucleotide or tryptophan synthesis was detected.
Labeling of the glycolytic intermediates proceeded rapidly all the way up to pyruvic acid (Pyr), which was then converted into acetyl-CoA (AcCoA) for subsequent condensation with oxaloacetate (OA) to generate the tricarboxylic acid cycle (TCA) intermediate 13C2-citric acid. Consistent with earlier reports for macrophages though [12] , we found activity of the TCA cycle to be markedly dampened in these cells. No detectable labeling was observed in intermediates downstream of citrate, all the way up to OA. Further, contributions from anaplerotic reactions or gluconeogenesis was also not detected. In line with our previous finding [13], there was no significant conversion of AcCoA into the ketone body 3-hydroxybutyric acid (3HB). Further, no evidence for any de novo synthesis of either free fatty acids (FA) or cholesterol (CL) was also found. 13C-Labeling profiles of representative metabolites are shown in Figure 1A, while the results are summarized in Figure S2A&S2B.
Our observation that the TCA cycle activity was attenuated in these cells was intriguing as it suggested that ATP synthesis by the oxidative phosphorylation pathway may also be diminished. This latter aspect could be experimentally confirmed by demonstrating that addition of the electron transport chain inhibitor Rotenone to cells had no significant effect on the intracellular ATP level (Figure 1B). Further, separation of the mitochondrial and cytoplasmic fraction of the cells revealed that over 85% of the total cellular ATP was present in the cytoplasm (Figure S3A). This finding is consistent with the fact that the ATP was derived, at least primarily, from the glycolytic rather than the oxidative phosphorylation pathway.
A significant consequence of the attenuated TCA cycle, and the resultant oxidative phosphorylation, activity was that maintenance of cell viability was now largely dependent upon the ATP generated through glycolysis. This was evident from experiments where depletion of glucose from the culture medium - after supplementing the latter with Pyr as the mitochondrial energy source - caused ∼70% reduction in cellular ATP levels (Figure.1B). Importantly, this then also led to a loss in mitochondrial membrane potential (ΔΨm) (Figure. 1C), and the consequent activation of apoptotic death (Figure. 1D). Similar effects were produced, in Pyr-supplemented media, by the glycolysis inhibitors 5-thioglucose (5TG) and 2-deoxyglucose (2DG), and the GAPDH inhibitor 3-bromopyruvate (3BP) [14], [15]. The latter finding is notable since GAPDH inhibition mitigates the payoff phase of glycolysis that is responsible for generating the net gain in both ATP and NADH (Figure S3B–S3D). These collective results, therefore, establish that THP-1 cells are indeed glycolytic in nature, and that their preservation in a viable state requires maintenance of an active flux through the payoff phase of glycolysis.
To capture the overall metabolic flux distribution in these cells we mathematically modeled the CCM by using a system of coupled ODEs that encompassed the relationship between individual metabolites in the pathway [16]. In this model we incorporated the experimentally determined rates of actual synthesis and consumption of each metabolite, to estimate the parameters that yielded metabolite steady state concentrations that corresponded with the experimentally determined values. For determination of rates of individual reactions we took the results from the experiment described in Figure 1A and first calculated the 13C-label incorporation rates for each metabolite. This was measured as the slope of the plot obtained up to the half-maximal value of 13C-incorporation in each case. Similarly, the slope for disappearance (i.e. consumption) of the corresponding naturally occurring isotopomer was also determined for each metabolite. The 13C-label incorporation rate of a given metabolite (e.g. X) gives its net rate of synthesis, which represents the combined outcome of both synthesis and consumption of that molecule (i.e. 13C-labeling rate of X = actual rate of synthesis – rate of consumption). Therefore, the actual rate of synthesis of ‘X’ (rX) could then be calculated as: rate of 13C-incorporation in X+ rate of decrease in concentration of its 12C-labeled counterpart. Taking the actual rates of synthesis and consumption then, we could also estimate the steady state concentration of each metabolite. The method employed for this is described in Figure S3E.
Figure 2A shows the pathway that was modeled and the flux distribution parameters that were estimated. The resulting model consisted of twenty-one metabolic state variables and sixty-one rate parameters. Of the latter, thirty-eight were those that were experimentally determined thus ensuring the robustness of the model. Fidelity of the model parameters was established by rigorous validation (Methods S1, Table S6–S9), which was also further supported by the fact that the model-derived metabolite steady state concentrations matched well with the experimentally determined values (Figure. 2B).
To study the effects of infection on macrophage CCM, we took a panel of five mycobacterial strains. In addition to Mycobacterium smegmatis (M.smeg), it included the Mtb strains H37Ra, H37Rv, BND433, and JAL2287. While H37Ra and H37Rv are the respective avirulent and virulent counterparts of a laboratory strain, BND433 and JAL2287 are clinical isolates of the CAS lineage [13], [17]. Following infection in THP-1 cells, M.smeg was rapidly cleared by 24 hr post-infection (p.i.), whereas the remaining strains persisted at a bacillary load that was of the order: JAL2287>BND433>H37Rv>>H37Ra (Figure S4A). Further, while H37Rv-, JAL2287-, or BND433-infected cells displayed significant - but variable - extents of necrotic death by 72 hr p-i, H37Ra-infected cells primarily underwent apoptosis whereas cells infected with M.smeg remained viable after bacterial clearance (Figure S4B). Thus, given that the mycobacterial panel encompassed distinct phenotypic properties that ranged from non-pathogenic to attenuated to differing degrees of virulence, a comparative analysis of the strain-specific effects was expected to yield insights into regulatory features that mediated mycobacterial virulence.
We first monitored for any influence of infection on host cellular glucose uptake. For this THP-1 cells were infected with the individual mycobacterial strains, and the rate of uptake of 2-NBDG - a fluorescent analog of glucose – was measured at appropriate time intervals that spanned a total period of 48 hrs. Infection with M.smeg or H37Ra caused only a marginal effect on 2-NBDG uptake rates (Figure. 3A). In contrast, however, cells infected with either of the virulent Mtb (virMtb) strains exhibited a marked increase in the rate of glucose uptake. Depending upon the infecting strain and time p-i, this increase ranged from 2- to 8-fold greater than the rate seen in uninfected cells (Figure 3A). Subsequent experiments revealed that increased glucose uptake by virMtb infected cells was due to pathogen-induced enhancement in cell surface expression of the high affinity glucose transporters Glut1 and Glut3. This effect was seen in both THP-1 (Figure 3B&C) and primary human monocyte-derived macrophages (HuMΦ) (Figure S4C and S4D), but only after virMtb infection. No such response was induced by either H37Ra or M.smeg infection (Figure 3B), thereby explaining the specificity of the effect in Figure 3A.
To determine the downstream consequences of the effects on glucose uptake, we took infected cells and pulse-fed them with 13C6-glucose at the times indicated in Figure 1A. Following this, we profiled the kinetic flux of the 13C label across the individual metabolites. The experimental design is illustrated in Figure 3D, and this strategy enabled us to determine the rates of individual reactions - at the various times p-i - in cells infected with each of the mycobacterial strains (Table S10). This experiment produced a complex data set that, unfortunately, was difficult to interpret. This was because each strain produced a unique glycolytic landscape profile with no distinct pattern segregating the effects of virulent Mtb strain from those of the non-pathogenic ones (Figure 4A).
Therefore, to interrogate these results further, we employed our ODE model for the infected cells across all the time points studied (see Figure 2A). Our aim here was to search for any underlying chokepoint reactions that may distinguish perturbations caused by the virMtb strains. That is we asked whether the glycolytic perturbations seen in Figure 4A were truly incoherent in nature, or whether they instead derived from targeted but variable effects on a small subset of the overall reactions? To address this we performed two related sets of simulation exercises. In the first we took the parameter set obtained for uninfected cells, and determined the minimum number of changes that would be required to simulate the profile of metabolite concentrations generated by each of the virulent Mtb strains. Interestingly, we found that this could be achieved through calibrated enhancements in rates of only two of the steps in the glycolytic pathway. These steps were: (i) uptake of glucose by the host cell and (ii) generation of DHAP and G3P through cleavage of FBP (Figure 4B). Especially significant also was the converse finding that the glycolytic landscape profiles (GP) produced against the virulent Mtb infections could be reset to resemble the H37Ra-specific response simply by appropriately attenuating the rates of these same two steps (Figure 4C). Although the robustness of our model gave confidence in these inferences we, nevertheless, additionally tested for fidelity of the predictions. For this we simulated a progressive reduction in glucose uptake rate in JAL2287-infected cells, and determined the consequent effects on GP, and on synthesis of R5P, AcCoA and FA. We found that the resulting trends compared well with experimentally determined values in cells cultured in medium containing decreasing concentrations of glucose (Figure S5A).
Thus in contrast to at least H37Ra, virMtb isolates were distinguished by their ability to promote glucose uptake by the host macrophage, while at the same time also accelerating the last reaction in the preparatory phase of glycolysis. The magnitude and dynamics of these influences however differed across the dimensions of strain variability and time, and it was these differences that then accounted for the heterogeneity of the profiles in Figure 4A. It is pertinent to note here that the virMtb-induced acceleration of synthesis of DHAP/G3P is especially relevant given that these molecules initiate the downstream steps of the energy generating, or payoff, phase of glycolysis.
Consistent with expectations, the targeted perturbations exercised by virMtb strains markedly enhanced glycolytic synthesis of ATP by the host cell across all the time points studied. This was determined by calculating the net rate of ATP production in each instance, from the rates of the individual reactions in the preparatory and pay-off phases of glycolysis (Figure 5A). In the case of H37Ra infection however, the rate of host cellular ATP production progressively declined with a significant inhibition being evident by 48 hr (Figure 5A). These estimates of the net biosynthesis rates corresponded well with subsequent measurements of the intracellular ATP concentration (Figure 5B), thereby confirming that infection-induced perturbations in glycolysis indeed impacted on the overall energetics of the host cell.
Importantly, in keeping with the dependence of macrophage survival on glycolytically generated ATP, it was the depletion of the latter that caused apoptosis in H37Ra-infected cells. This aspect was verified by experiments where glucose uptake by these cells was potentiated by the exogenous addition of ESAT-6 as described earlier [13]. Thus, whereas ESAT-6 addition rescued H37Ra-infected cells from death, co-addition with 3BP to selective block glycolysis-dependent ATP generation – however – reversed this effect (Figure 5C).
In contrast to the situation with H37Ra, it was the endogenous augmentation of ATP biosynthesis that suppressed apoptosis of virMtb-infected macrophages. This was confirmed by our finding that either suppression of glucose transport with Phloretin [18]–[20], or depletion of glucose in the medium, induced apoptotic death of H37Rv-infected cells (Figure 5D&E). Also in line with the increased bioenergetic demand placed on infected macrophages, we found that maintenance of H37Rv-infected cells in a viable state required a higher threshold of glucose availability in comparison with that for uninfected cells (Figure 5E and Figure S5B). Finally, addition of 3BP to H37Rv-infected cells caused the expected suppression of cellular ATP synthesis (Figure 5F), resulting in mitochondrial membrane depolarization (Figure 5G) and apoptotic death (Figure 5H).
These collective results therefore establish the functional significance of pathogen-induced perturbations in glycolytic flux of the host macrophage. Importantly, by delineating the manner in which this influence impinged on the regulation of host cell apoptosis, these findings also unveil an early checkpoint that likely contributes towards pathogenesis of the infection.
Integration of glycolysis with the TCA cycle is mediated through oxidation of Pyr to AcCoA. Enzymatic condensation of this AcCoA with OA then generates the TCA cycle intermediate, citrate. Interestingly in this context, we found that virMtb-infected cells – but not H37Ra or M.smeg infected ones – also exhibited a pronounced increase in the rate of cellular AcCoA synthesis (Figure 6A). Strain-dependent differences were, however, noted in the kinetics of this process. JAL2287 infection produced an early burst of ∼7-fold increase, which then later subsided to about 3-fold over the basal value. In comparison, AcCoA production gradually increased in H37Rv-infected cells whereas BND433 infection yielded a biphasic response curve (Figure 6A). The mechanism involved, however, is presently unknown and awaits clarification.
Surprisingly, we found that the virMtb-induced augmentation of host cellular AcCoA synthesis did not translate into the expected increase in citrate production (Figure 6A). Instead, AcCoA was primarily utilized for synthesis of the ketone body 3HB (Figure 6A). Ketone body synthesis represents an alternate pathway for mitochondrial AcCoA, which is activated when AcCoA production rates exceed that of its utilization in the TCA cycle [21]. Thus, the observed conversion of accelerated AcCoA synthesis into 3HB generation probably reflects a consequence of the attenuated nature of the TCA cycle in these cells. Indeed we have previously demonstrated that activation of this host pathway constituted a virulence mechanism of Mtb, with 3HB then driving acquisition of the ‘foamy’ phenotype by the host macrophage [13].
Although virMtb infection did not affect citrate production we, nonetheless, observed significant differences at the level of MVA and MalonylCoA (MaCoA). Synthesis of both of these intermediates was heightened in virMtb- but not in H37Ra- or M.smeg-infected cells (Figure 6A). MVA and MaCoA, intermediates in the biosynthesis of CL and FA respectively, are both products of a cataplerotic pathway of the TCA cycle wherein mitochondrial citrate is exported to the cytoplasm. Enzymatic cleavage of this exported citrate by ATP-citrate lyase generates cytoplasmic AcCoA, which then constitutes the substrate for both MVA and MaCoA biosynthesis. To resolve how synthesis of these latter molecules was enhanced in the absence of any effect on citrate production, we pulse-labeled infected cells at 6 and 36 hr p-i. The cytoplasmic and mitochondrial fractions were then separated to determine the relative citrate concentrations present in each fraction. In addition we also monitored the rate of labeling of the cytoplasmic citrate pool, as a measure of the rate of its export from the mitochondria.
Figure 6B shows that the proportion of the mitochondrial citrate was significantly reduced in virMtb-infected cells. This effect was more pronounced at 36 hr, relative to 6 hr, for cells harboring either H37Rv or BND433 whereas the converse was true of those infected with JAL2287 (Figure 6B). These effects could subsequently be rationalized through a comparison of the relative rates of citrate export to the cytoplasm (Figure 6C). Exposure to virMtb uniformly led to an increase in this process, with an initial effect at 6 hr being further amplified by 36 hr in the case of H37Rv and BND433 infection. For JAL2287 however, peak activation of host cellular citrate export was already evident by 6 hr, with a slight decrease at 36 hr (Figure 6C). In contrast to the effects of virMtb, infection with M.smeg failed to exert any influence whereas only a marginal effect was detected in response to H37Ra infection (Figure 6C).
The functional significance of virMtb-specific augmentation in MVA and MaCoA synthesis could next be confirmed by experiments demonstrating that this process also resulted in a corresponding increase in synthesis rates of CL, and palmitic acid as a representative FA (Figure 6D). Interestingly, although for reasons presently unknown, infection either with JAL2287 or BND433 preferentially activated FA synthesis whereas H37Rv infection biased towards CL production (Figure 6D). Furthermore, the kinetics of FA induction also differed between JAL2287 and BND433 infected cells (Figure 6D). Nonetheless, given that intracellular survival and persistence of Mtb depends upon its ability to access host-derived FA and CL as nutrients [4], [22], [23], these results would suggest that activation of host lipid synthesis likely constitutes an obligatory prerequisite for establishment of infection.
Key to the intracellular survival of Mtb is its ability to induce the host macrophage to differentiate into FMs [7], [13], [24]–[26]. This occurs through the accumulation of lipid bodies (LBs) in the cell, which are composed of triglycerides (TAGs) and cholesteryl esters (CEs)[27]. LBs function by serving as a nutrient source, and also by providing a secure niche, for the intracellular bacteria [25]–[27]. Therefore, it seemed probable that the de novo lipid synthesis induced by virMtb represented a precursor step for LB biogenesis. We verified this by treating virMtb-infected cells with either Atorvastatin [28] to suppress CL biosynthesis via HMGCoA reductase inhibition, or with the FA synthase (FAS) inhibitor C-75 [29]. Either of these inhibitors reduced the LB content of the host cell although the effect of C-75 was more pronounced (Figure 7A). Significantly, the intracellular bacillary load was also reduced (Figure 7B). We obtained similar effects upon silencing of the corresponding target enzymes by siRNA (Figure S5C and S5D), which confirmed that compromised mycobacterial survival was a result of suppression of host rather than bacterial lipid biosynthetic pathways. In this connection, an earlier study has reported upregulation of transcripts coding for ATP citrate lyase and HMGCoA reductase in human TB granulomas [30]. This observation would support the physiological relevance of our findings.
Figure 7C superimposes the time course profiles for 3HB and lipid (FA plus CL) synthesis by the virMtb-infected cells. Although the magnitude and kinetics vary, it is evident that the concomitant induction of both lipids and 3HB by the host cell constitutes a recurring theme of virMtb infections. We previously showed that 3HB also played a critical role in FM differentiation [13]. It acted by stimulating the antilipolytic receptor GPR109A to inhibit TAG turnover and, thereby, promoted LB accumulation [13]. Thus the complementary capabilities of provoking host cell lipid synthesis on the one hand, while simultaneously also suppressing TAG turnover on the other, likely enables Mtb to efficiently steer FM differentiation.
TAG accumulation in macrophages exhibit cytotoxicity, causing the cells to undergo necrosis [13], [31], [32]. Therefore, FM differentiation could also potentially account for the necrotic death of virMtb-infected macrophages. To test this we induced FM differentiation in THP-1 cells by stimulating them with 3HB. The consequent enhancement in TAG/LB levels also led to an increase in the population of necrotic cells (Figure 7D). Suppression of either TAG synthesis with Triacin C [33], or of LB biogenesis with CI-976 [13], however, inhibited this process (Figure 7D). Similarly, H37Ra-infected cells could also be induced to switch from apoptotic to necrotic death in the presence of exogenous 3HB. This switch was sensitive to the GPR109A inhibitor Mepenzolate bromide (MPN)[13] (Figure 7E), supporting a mediatory role for LBs.
On the converse side, necrosis of virMtb-infected cells was suppressed by inhibition of pathways mediating TAG accumulation in macrophages (Figure 7F). Pathways that were inhibited included citrate export from mitochondria, its subsequent catabolism, lipid synthesis, and TAG turnover (Figure 7F). Further, suppression of mitochondrial AcCoA generation (Figure 7G) by addition of the mitochondrial pyruvate carrier (MPC) inhibitor UK5099 [34], [35] resulted in concomitant inhibition of both FA and 3HB synthesis by the host cell (Figure 7H). Consequently, LB formation was also inhibited and the virMtb-laden cells were protected from necrotic death (Figure 7H). RNAi-mediated silencing of the human MPC1 and MPC2 genes also produced similar effects (Figure S5E; S5F and S5G), confirming the target specificity of the effects of UK5099. These collective results, therefore, establish that TAG cytotoxicity prevails in infected macrophages, and that initiation of FM differentiation in virMtb-infected cells also constitutes commitment to the eventual death by necrosis.
The mode of death adopted by infected macrophages is relevant to the pathophysiology of Mtb infection. While apoptosis constitutes a bactericidal response that limits disease, necrotic death facilitates escape of bacilli from the host cell to promote the spread of infection [36]–[41]. Our finding that extracellular glucose concentrations influence the extent of apoptosis of Mtb-infected macrophages would then also link glucose availability to the control of infection. This relationship could be demonstrated in both THP-1 and HuMΦ cells, in the context of infection with diverse clinical isolates (Figure 8A). Further, in mouse models for diabetes and hypoglycemia, the glycemic status clearly impacted on Mtb growth (Figure 8B&C). Importantly here, the relatively enhanced bacterial growth in diabetic mice could be suppressed by UK5099 treatment (Figure 8D), although this compound had no effect on extracellular cultures of Mtb or on macrophage viability (Figure S6A and S6D). These results therefore underscore the dependence of in vivo Mtb growth on host cellular metabolic flux.
The capacity of the infecting strain to enhance glucose uptake by the host macrophage (Figure 3A–C) could also additionally superimpose on the constraints of glucose availability, and contribute towards determining survival versus death of the infected macrophage. To investigate this we took another Mtb isolate, JAL2261, which was compromised in its ability to potentiate glucose uptake rates (<25% effect) in both THP-1 and HuMΦ cells (Figure S6B). Its inclusion alongside JAL2287, BND433, and H37Rv therefore, expanded the quantitative range of perturbations in host cell glycolytic flux that could be studied. Notably, virMtb strain-dependent effects on macrophage glucose uptake did not strictly correlate with the corresponding intracellular bacillary density (Figure S6B), suggesting that the former property was at least not entirely dependent on the latter.
The magnitude of Mtb-elicited acceleration of glucose uptake by the host cell clearly influenced the mode of death that resulted. Apoptosis of the host cell predominated when the pathogen exerted only a marginal effect on glucose uptake (JAL2261, Figure 8E). However, as the capacity of the infecting strain to stimulate host cell glucose uptake increased, the death response progressively shifted towards necrosis (Figure 8E). 5TG addition neutralized this shift confirming its dependence on the glycolytic flux (Figure 8E). Consistent with these ex vivo observations, both the extent of bacterial growth (Figure 8F) and the magnitude of the granulomatous response (Figure 8G and Figure S6C) in mice also displayed a trend that correlated with the ‘strength’ of the infecting strain to initiate glycolysis in the host macrophage. Admittedly, however, it is presently difficult to distinguish the relative contribution of metabolic effects from that of other factors such as bacterial burden, to the observed pathology.
The ability to control and/or manipulate the timing and mode of death of infected cells plays an important role in mycobacterial infection [36]–[41]. Control of death by the host generally initiates apoptosis. This then functions as an innate defense mechanism that restricts microbial spread and enhances the induction of adaptive immunity. It is generally believed that virMtb inhibits apoptosis early in infection and, instead, induces necrosis of the host cell at later time points. This outcome provides for release of viable bacteria from the infected cells thereby promoting re-infection and, ultimately, transmission of disease. In addition, necrosis of infected macrophages has also been implicated in granuloma formation and inflammatory tissue damage [30], [42]. Our present study substantially enriches this general view by elucidating the scheme employed by virMtb to regulate death pathways of the host macrophage. Significant in this context is the identification that this process was controlled at the level of host cell metabolism, with regulation also impacting on pathology of the disease.
Resolution of the pathogenetic strategies employed by Mtb was aided by our study design, which specifically incorporated several distinctive features to enable this end. First was the inclusion of a mycobacterial panel that collectively presented a manifold of virulence traits. By complementing this with a time-series analytical protocol, we could distinguish between the macrophage metabolic response profiles that correlated with either the elimination of non-pathogenic bacteria, the curtailment of an attenuated strain, or even with the successful cooption of host cell function by virulent bacilli. Interrogation of the data with ODE models then helped identification of the key interference points that dictated expression of Mtb virulence.
Our discovery that the capacity to modulate glycolytic flux of the host macrophage dictated Mtb pathogenicity provides a new perspective on host-pathogen crosstalk. As demonstrated, this ability facilitated pathogen survival by subverting apoptosis of the infected cell. Subsequent studies clarifying the sole dependence of macrophages on glycolysis for their bioenergetic needs then revealed the basis for this effect. Thus, modulation in glycolytic flux was alone sufficient to determine the balance between survival and apoptotic death of the host cell. Our ODE models subsequently identified that Mtb virulence in fact derived from the targeted perturbation of only two steps in the pathway. One of these involved augmentation of glucose uptake by the macrophage, while the other fostered this effect by increasing substrate availability for the payoff phase of glycolysis. As a result, the rate of ATP synthesis was markedly enhanced in virMtb-infected cells. Thus, whereas the default consequence of infection on the host macrophage was to impose an energy demand that breached the homeostatic barrier, manipulation of ATP synthesis by virMtb circumvented this to ensure that infected cells remained viable.
In conjunction with suppression of host macrophage apoptosis, establishment of infection also requires that the pathogen tend to its own survival needs within the intracellular milieu [43]. Here again our study yielded important insights by delineating how the glycolysis-mediated inhibition of apoptosis was allied with downstream regulation of select mitochondrial pathways to drive FM differentiation. The LBs generated as a result are known to provide a privileged niche for the intracellular bacteria [2], [13], [25], [26], [30]. As shown, the proficiency of pathogen in this regard stemmed from its ability to stimulate the complementary host processes of de novo lipid synthesis to generate TAGs on the one hand, and production of 3HB to inhibit lipolysis on the other. The concerted functioning of these two pathways was confirmed by the fact that pharmacological inhibition of either one of them compromised FM differentiation of virMtb-infected cells. Thus, regulation of LB biogenesis constitutes an indispensable ingredient of Mtb pathogenicity.
While FM differentiation facilitated stabilization of the infection, we also found that continued accumulation of TAG eventually led to necrotic death of the cells. That is, the control over macrophage metabolism by pathogen persisted up to the point of securing a mechanism for eventual escape of the bacteria from these cells. These latter findings add to the earlier results and highlight the fluidity with which the individual stages that contribute to pathogenesis are integrated, and governed, by the pathogen. Noteworthy here was that this was achieved through a parsimonious strategy of targeting only select reactions. In this connection we note that both the temporal and quantitative aspects of virMtb-induced host cell lipid synthesis varied significantly between the individual strains. Therefore, it will be interesting to determine if the end point of TAG-induced necrosis is also differentially regulated depending on the Mtb strain and/or the bacillary load.
Diabetes is an established risk factor for TB [44]–[46], and at least one causal factor suggested is the compromised immunity of hyperglycemic individuals [47], [48]. Our present finding that circulating glucose levels also contribute to pathogenicity by influencing macrophage apoptosis, however, reveals an additional facet to this relationship. That is, the glycemic status of an individual also impacts on the innate defense mechanism that controls Mtb infection. Notably, there is emerging evidence to suggest that the link between glucose availability and fate of the infected macrophage has broader implications that extend beyond the case of Mtb infection alone. Thus, for example, it may explain earlier findings that glucose availability also sustains chronic Brucella abortus infection in macrophages [49]. Similarly, it may also rationalize the recent discovery that control of glucose levels in the airway surface liquid serves as a mechanism for maintaining sterility of the lung airways [50].
Our observation that virMtb strains differ in their capacity to promote uptake of glucose by the host macrophage is also pertinent as it adds a complimentary dimension to the issue of glucose-dependency of Mtb infection. As shown, this property syndicates glucose availability with the cellular glycolytic rate by scaling the rate of glucose uptake by the host cell. Thus, glycemic levels and macrophage glycolysis together constitute an axis that integrates properties derived from both host and pathogen, in deciding the course of infection. That is, the ensuing pathology in any Mtb-infected individual is likely a result of the integrated effects of the individual's glycemic status, and nature of the infecting strain. Further analysis of this relationship may provide insights into factors that determine susceptibility to TB. Additionally, it may also be worthwhile to explore whether necrosis of Mtb-laden FMs contribute to the observed link between TB and atherosclerosis, with the consequent risk potentiation of ischemic stroke [51]–[55].
The facility with which Mtb continues to persist and disseminate in the human population remains a daunting challenge [6]. Our predominant focus on host immunity as a determinant of disease susceptibility is yet to yield results in terms of an effective vaccine against TB, or even a less ambiguous description of the correlates of protection. A case may, therefore, be made for revisiting the existing paradigms for understanding Mtb pathogenicity. By highlighting the extent to which metabolic intervention dictates the pathology of TB, our present report would indeed support such a view. Significant in this context is our identification that chokepoint interactions do exist that may be indifferent to either genotype or phenotype of the Mtb strain. Such a scenario then, also raises the possibility of exploiting such interactions for the development of new therapeutic strategies.
PMA-differentiated THP-1 cells (routinely tested to be Mycoplasma free) were infected with mycobacteria at a multiplicity of infection of 10 over a 4 hr period. This was followed by amikacin treatment for 2 hr to remove any extracellular bacteria as described earlier [13], [17]. Human PBMC-derived monocytes were isolated by centrifugation over Ficoll-Paque from heparinized blood and then selected by adherence. Monocytes were then spontaneously differentiated into macrophages as described in Methods S1. For determination of CFUs, infected cells were lysed in 0.06% SDS and then plated on 7H11 agar plates supplemented with OADC and 5% glycerol. The Mycobacterial strains used in this study have been described earlier [13], [17].
The animal care and use protocol was reviewed and approved by the Institutional Animal Ethics Committee (IAEC) of International Centre of Genetic Engineering and Biotechnology, New Delhi (ICGEB, New Delhi). The animal care and use protocol adhered to the Committee for the Purpose of Control and Supervision of Experiments on Animals (CPSEA), Government of India, guidelines for Laboratory animal Facility. The reference number for the approval (IAEC number) is ICGEB/AH/2013/1/IMM32.
Inhibitor treatment of infected and uninfected controls was initiated after Amikacin removal post infection (6 hrs p-i.). The medium containing the appropriate dose of inhibitor was replenished at 36 hrs and where needed, at 64 hrs p-i. The inhibitors were used at the following final concentrations: Atorvastatin at 10 µM, C75 at 20 µM, UK5099 at 5 µM, Rotenone at 1 µM; Mepenzolate (MPN) at 100 nM; 3BP at 50 µM; Citrate Lyase inhibitor-3,5 dichloro-2-hydroxy-N-(4-methoxy biphenyl-3-yl) benzene sulphonamide (DCBS) at 50 µM; 1,2,3 benzene tricarboxylate at 200 µM; 5TG at 20 mM; 2DG at 12 mM; and Triacin C at 5 µM.
The above inhibitor concentrations were determined by generating dose response curves on H37Rv-infected THP-1 and HuMФ cells following the protocol described above. The concentration showing maximal efficacy, but without any cytotoxic effect on the host cells (as determined in parallel experiments of inhibitor to UI cells), was selected. Inhibitor cytotoxicity was determined by the MTT assay. Importantly, we also confirmed that – at the concentrations tested – none of the inhibitors employed had any effect on growth of cell-free bacterial cultures. (Figure S6A and S6D)
Labeled RPMI medium was prepared by adding 2 g/liter of 13C6-glucose or 0.3 g/liter of 13C5-glutamine to glucose or glutamine free RPMI media respectively followed by sterile filtration. For labeling, 5*106 differentiated THP1 cells were switched to fresh unlabeled media 1 hour before labeling so as to reduce perturbations in metabolite levels at the time of the labeling [56]. Unlabeled media was then completely removed, cells were washed once with glucose free RPMI, and fresh 13C6-glucose or 13C5-glutamine labeled media supplemented with 10% dialyzed fetal calf serum was added for the desired times. Metabolic reactions were quenched by adding chilled (−75°C) mixture of methanol:water (80∶20, v/v) After incubation at 75°C for 10 minutes the cells were scrapped from the culture dish (kept on dry ice) and collected. The scrapped cell suspension was then vortexed, followed by centrifugation at 6000 g for 5 minutes at 4°C, and the supernatant was stored. The pellet was re-extracted two more times with 80% methanol at −80°C. The three extractions were pooled, centrifuged at 13000 g for 5 minutes to remove any debris, and dried under a Nitrogen stream. The dried samples were re-suspended in MS-grade water, and centrifuged at 13000 g, 4°C for 10 minutes. Supernatants were used for the LC-MS/MS analysis.
Metabolites were quantified by employing concentration-dependent standard curves for each metabolite as described in Methods S1. Mitochondrial and cytoplasmic fractions were generated as detailed in Methods S1 and the kinetic labeling data was similarly obtained. For determining the rate of lipid and cholesterol synthesis, cells were incubated in 13C6-glucose containing RPMI for 4 hours and total cellular lipids were extracted. Cellular TAGs were then hydrolyzed with HCl/CH3CN (1∶4, v/v) and Palmitic acid was analyzed by direct infusion on an Applied Biosystems/MDS Sciex Q 4000 TRAP linear ion trap mass spectrometer in a negative polarity mode. Determination of free cholesterol was achieved by converting it to cholesterol-3-sulfate, and then analyzing by mass spectrometry with the sulfate group providing the negative charge. The detailed protocol is provided in Methods S1.
Cells were initially washed with glucose free RPMI followed by addition of glucose free RPMI supplemented with 200 µM 2-NBDG for 0,5,10,20 and 30 minutes. Subsequently, the labeling media was removed cells were washed in glucose free RPMI and lysed in 200 µl of a non-interfering buffer (1% Nonidet P-40, 1% sodium deoxycholate, 40 mM KCl, 20 mM tris pH7.4). The lysate was centrifuged at 13000 g at 4°C for 5 minutes to remove debris. Fluorescence of the internalized glucose was measured on a flourimeter at 535 nm (excitation wavelength 485 nm).
TAG levels were measured after solubilizing cells in 5% (v/v) Triton X-100 with the Triglyceride Quantification Kit (Abcam), and following the instructions of the manufacturer.
To determine the mode of death, the Apoptosis/Necrosis Detection kit (Enzo) was used. At the time of assay the cells were fixed in 2% PFA and the protocol recommended by the manufacturer was followed. Cells were analyzed by confocal microscopy.
Cells were stained with the required reagents as mentioned in Methods S1. Stained cells were observed with a Nikon EclipseTi-E laser scanning confocal microscope equipped with 60X/1.4NA Plan Apochromatic DIC objective lens. DAPI and Lipid Tox/MitoSOX were excited at 488 nm, 408 nm and 543 nm with an argon ion, blue diode and a Helium-Neon laser respectively. The emissions were recorded through emission filters set at 515/30; 450 and 605/75. Images were acquired with a scanning mode format of 512×512 pixels. The transmission and detector gains were set to achieve best signal to noise ratios and the laser powers were tuned to limit bleaching fluorescence. The refractive index of the immersion oil used was 1.515 (Nikon). All settings were rigorously maintained for all the experiments.
Female BALB/c mice, 4–6 weeks of age (8 per group) were infected with H37Rv through aerosol exposure by delivering, during 30 min of exposure, between 100–150 bacteria per lung. The latter was determined by the culture of lung homogenates at 24 hr later. STZ (180 mg/kg) treatment was initiated by administering intra peritoneal injection once, 7 days before aerosol exposure.
In the appropriate groups, bovine insulin (60 IU/kg/day) was administered via a mini osmotic pump (Alzet), implanted subcutaneously on the back of mice. Insulin treatment was started 24 hrs prior to aerosol administration. The mice in the corresponding control group were similarly implanted with saline filled mini osmotic pumps.
In a separate experiment female BALB/c mice, 4–6 weeks of age (8 per group) were treated with UK5099 (12 mg/kg/day), via a mini osmotic pump implanted subcutaneously on the back of mice, 24 hrs prior to aerosol infection with JAL2287 strain of Mtb. Blood sugar levels for all treated mice and controls were monitored every third day. At the appropriate times, these mice were sacrificed and mycobacterial load in lung was monitored. In all animal studies, the author taking CFU counts was not aware of the sample identity, the nature of treatment and thus the counts were taken on an unbiased background.
We determined the statistical significance by using the unpaired two-tailed Student's t test with origin software. p<0.05 was taken as statistically significant.
High-pressure liquid chromatography (HPLC) was performed on an Agilent 1260 infinity Binary HPLC (Agilent technology, Waldbronn, Germany) equipped with a degasser, and an auto sampler. In all, three types of columns were used in the study- i) Aminopropyl, ii) C-18, iii) Cyano column. The auto sampler was maintained at 4°C to ensure sample stability. A flow rate of 200 µl/min and sample injection volume of 20 µl was maintained in the study. Elutes were continuously directed to the mass spectrometer with the help of turbo ion source.
Targeted metabolomic analysis of 25 hydrophilic metabolites was performed on the aminopropyl column using multiple reactions monitoring (MRM). Agilent Polaris 5 NH2 2×150 mm column was used with a non-linear gradient of 85–0% B over 36 min (Table S1). Solvent A was 5% ACN/H2O containing 10 mM ammonium acetate and 10 mM ammonium hydroxide; pH 9.4 and solvent B was 100% ACN.
The HPLC was coupled with a hybrid 4000 QTRAP (AB SCIEX, Foster City, CA, USA) with a Turbo V ESI ionization source interface, and a computer platform equipped with a Solution Analyst software version 1.5 (ABSciex, Foster City, CA, USA), which was used for data acquisition and processing as described in Methods S1. The mass spectrometric parameters for the precursor and product ions selected in MRM for metabolites under study and their corresponding parent and daughter ions parameters are depicted in Table S2. Standards were purchased from Sigma Aldrich and were used to generate the MRM profile and optimizing the source and compound parameters. Additional information on the validation method, the LC-gradient profile (Table S1, Figure S7), isotopomer analysis (Tables S2, S3, S4 and Figures S8, S9, S10) and chromatographic approaches employed for the various metabolites are described in Methods S1.
GAPDH (Glyceraldehyde-3-phosphate dehydrogenase): P04406; GLUT1 (Solute carrier family 2, facilitated glucose transporter member1): P11166; GLUT3 (Solute carrier family 2, facilitated glucose transporter member 3): P11169; HMGCR (3-Hydroxy-3-Methylglutaryl-CoA Reductase):P04035; FAS(Fatty Acid Synthase):P49327; GPR109A (Hydroxycarboxylic acid receptor 2):Q8TDS4; MPC1(Mitochondrial pyruvate carrier 1): Q9Y5U8; MPC2(Mitochondrial pyruvate carrier 2): O95563.
|
10.1371/journal.pgen.1006041 | Fission Yeast SCYL1/2 Homologue Ppk32: A Novel Regulator of TOR Signalling That Governs Survival during Brefeldin A Induced Stress to Protein Trafficking | Target of Rapamycin (TOR) signalling allows eukaryotic cells to adjust cell growth in response to changes in their nutritional and environmental context. The two distinct TOR complexes (TORC1/2) localise to the cell’s internal membrane compartments; the endoplasmic reticulum (ER), Golgi apparatus and lysosomes/vacuoles. Here, we show that Ppk32, a SCYL family pseudo-kinase, is a novel regulator of TOR signalling. The absence of ppk32 expression confers resistance to TOR inhibition. Ppk32 inhibition of TORC1 is critical for cell survival following Brefeldin A (BFA) induced stress. Treatment of wild type cells with either the TORC1 specific inhibitor rapamycin or the general TOR inhibitor Torin1 confirmed that a reduction in TORC1 activity promoted recovery from BFA induced stress. Phosphorylation of Ppk32 on two residues that are conserved within the SCYL pseudo-kinase family are required for this TOR inhibition. Phosphorylation on these sites controls Ppk32 protein levels and sensitivity to BFA. BFA induced ER stress does not account for the response to BFA that we report here, however BFA is also known to induce Golgi stress and impair traffic to lysosomes. In summary, Ppk32 reduce TOR signalling in response to BFA induced stress to support cell survival.
| The Target of Rapamycin (TOR) pathway plays a central role coordinating cell growth and cell division in response to the different cellular environments. This is achieved by TOR controlling various metabolic processes, cell growth and cell division, and in part by the localisation of TOR protein complexes to specific internal endomembranes and compartments. Here, we report a novel role for the SCYL family pseudo-kinase, Ppk32 in restraining TOR signalling when cells experience stresses, which specifically affect endomembranes and compartments where TOR complexes are localised. Cells exposed to endomembrane stress (induced by Brefeldin A), displayed increased cell survival when simultaneously treated with the TOR complex 1 (TORC1) inhibitor, rapamycin, presumably because the reduction in TORC1 signalling slows cellular processes to allow cells sufficient time to recover and adapt to this stress. Importantly cancer, neuro-degeneration and neurological diseases are all associated with stress to the endomembrane protein trafficking system. Our findings suggest that therapeutic rapamycin treatment to reduce TOR signalling and block proliferation of cancer cells, which are inherently experiencing such stress, may have the unintended consequence of enhancing cell survival. It is notable, therefore, that our reported mechanisms to reduce Ppk32 protein levels, likely to be conserved in humans, may represent a way to increase TOR signalling and thus increase cell death of cancer types with inherent stress to these internal membrane systems.
| TOR signalling allows eukaryotic cells to adapt their metabolism, cell growth, stress and survival to meet the demands of the prevailing conditions [1]. TOR kinases form at least two distinct complexes: TOR complex 1 (TORC1) and TORC2 [2–4]. These complexes are defined by the presence of unique binding partners; Raptor interacts with TOR kinase in complex 1, whereas Rictor replaces Raptor in complex 2 [2,4]. The yeasts differ from higher eukaryotes in having two separately encoded TOR kinases. In fission yeast Tor1 is the main kinase that binds Ste20 (rictor) in TORC2, whereas Tor2 is in a complex with Mip1 (raptor) in TORC1 [5–7]. Yeast and mammalian TORC1 responds to changes in the abundance of nutrients and growth factors (mammals) to adjust the cell cycle, cell growth and metabolism accordingly. The roles for TORC2 include modulation of the actin cytoskeleton [3,8], stress responses [9] and chaperone-mediated autophagy [10]. Both TOR complexes localise to membrane-enclosed structures. In nutrient rich conditions, TORC1 is found on lysosomes [11] and the Golgi apparatus [12], whereas TORC2 localises mainly to the endoplasmic reticulum (ER) [13]. TORC2 can also be found on the plasma membrane [13,14] and lysosomes [10].
The endoplasmic reticulum originates from the nuclear envelope to extend throughout the cell. ER membranes are in constant contact with the Golgi apparatus. COP-mediated trafficking, regulates vesicle transport both to and from the ER and Golgi and between Golgi stacks [15]. Once proteins reach the trans-Golgi apparatus, they are directed either to plasma membrane, lysosomes or other vesicle-based compartments. Therefore, these endomembranes function as protein and lipid factories but also as scaffolds for complexes such as TOR signalling modules.
The lactone antibiotic Brefeldin A (BFA) inhibits the GEFs for class II ARFs (ADP-Ribosylation Factor a GTPase) [16,17] to release ARF into the cytosol. This release reversibly blocks traffic between the Golgi and ER and within the Golgi stacks to generate Golgi stress. The BFA induced block to protein traffic can also induce ER stress through the unfolded protein response (UPR) [18]. Furthermore, BFA impairs traffic to the lysosomes [19]. Importantly, BFA has been key to elucidating the mechanisms of trafficking at the Golgi.
Here we report that Ppk32, a SCYL family pseudo-kinase, is a novel regulator of TOR signalling. Reduced TORC1 activity promoted survival upon Brefeldin A (BFA) induced stress. Ppk32 was critical for inhibition of TORC1 signalling and survival during this BFA induced stress. Ppk32 also controlled TORC1 activity during sexual differentiation and ablation of ppk32+ expression advanced differentiation. The response to BFA that we report here is not regulated though UPR induced ER stress. Therefore, cells are likely to sense Golgi stress, or impaired traffic to lysosomes, following BFA treatment. Finally we show that Ppk32 phosphorylation on two sites is important for function because it controls Ppk32 protein levels and determines cellular sensitivity to BFA induced stress and TOR inhibition. This target sites for phosphorylation are conserved through to human members of the SCYL pseudo-kinase family
The ability of rapamycin to specifically inhibit TORC1 has facilitated extensive characterisation of TORC1; in contrast TORC2 regulation is less well defined. Preliminary data suggested that the S. pombe Ppk32 kinase co-immuno-precipitated with Tor1 (the main catalytic component of TORC2) in a large-scale screen to identify TORC2 interacting proteins. Small-scale Tor1 immuno-precipitations validated this interaction with Tor1, through co-precipitation of both Ppk32 and of PK tagged Ppk32 with Tor1 (Fig 1A). Ppk32 is a member of the SCYL pseudo-kinase family that includes the human SCYL1 and SCYL2 kinases. SCYL kinases have an amino-terminal pseudo-kinase domain and carboxy-terminal serine-rich sequences separated by HEAT repeats that are known to facilitate protein-protein interactions. TOR kinases interact with Raptor (TORC1) and Rictor (TORC2). Interestingly, this interaction is mediated, in part, through the HEAT repeats found in all 3 proteins [2]. Fission yeast TORC1 is localized to the surface of vacuoles (yeast lysosomes) and to undefined cytoplasmic foci that are not vacuoles [20]. In contrast, TORC2 localises to cell ends and to the cell equator (sites of the ER and plasma membrane) during late stages of cell division [21]. A GFP-Ppk32 strain (S1B Fig) revealed that Ppk32 was present in the cytoplasm (although signals were faint), and it was excluded from the inside of vacuoles (Fig 1B). In some places cytoplasmic Ppk32 accumulated in small foci (Fig 1B-2), between two to ten foci were seen in each cell. Ppk32 was also concentrated at the cell equator during cell division in more than 60% of cells (Fig 1B). Thus, Ppk32 localizes to structures where TORC1 and/or TORC2 may reside.
In human cells the SCYL1 kinase has been implicated in retrograde traffic between the Golgi and ER [22], a process that is affected by the lactone antibiotic Brefeldin A (BFA). BFA effectively blocks Golgi transport to compromise viability in fission yeast [23]. The fission yeast genome encodes an additional SCYL homologue, Ppk3. Strains from which the genes encoding both homologues had been deleted were more sensitive to BFA than wild type controls (Fig 1C) (and GFP-Ppk32 cells S1C Fig). Importantly, Ppk32 but not Ppk3 deficient mutants were resistant to a potent inhibitor of both TOR complexes, Torin1 [24] (Fig 1C). This Ppk32 dependent resistance to Torin1 was determined by the nutritional context, as ppk32.Δ (ppk32+ deletion) cells were only resistant to Torin1 when they were also auxotrophic for leucine and grown on nutrient rich (YES) media (Fig 1D). Neither prototrophic ppk32.Δ grown on rich YES nor auxotrophic ppk32.Δ mutants grown on minimal media were drug resistant. Therefore, unless otherwise stated, all experiments in this report use leucine autotrophic strains grown on rich YES medium.
Assessment of Ppk32 levels in total protein extracts revealed that Ppk32 protein levels in rich media were twice those of cells grown in minimal EMMG media (Fig 1E). A similar increase of Ppk32.PK total protein levels was seen in YES grown cells (S1E Fig). Addition of the Phos-tag reagent [25] to the resolving gel revealed that Ppk32.PK is phosphorylated, however similar relative levels of phosphorylation were observed in both media (S1E Fig). The increased Ppk32 total protein level in YES may explain the increased Torin1 resistance of ppk32.Δ cells compared to ppk32+ cells when grown in rich YES medium. Elevated TOR signalling or deficient drug uptake by ppk32.Δ mutant could account for the resistance to Torin1. The decline in TORC1 controlled Maf1 phosphorylation [26] after Torin1 treatment followed similar dosage effects, in wt and ppk32.Δ cells (S1D Fig) to indicate that ppk32.Δ mutants are proficient for drug uptake. Fission yeast TORC2 phosphorylates the AGC kinase Gad8 (AKT homolog) at Ser-456 [27]. The phosphorylation profiles of two TOR substrates, Maf1 and Gad8.S546 phosphorylation (Fig 1F), suggests that, while ppk32.Δ mutants cells are resistant to Torin1 (Fig 1C), Ppk32 does not appear to have a major impact on either TORC1 or TORC2 signalling under non-stressed condition.
To gain further insight into TOR and Ppk32 regulation, Ppk32 protein levels were assessed in TOR deficient mutants. Tor2 is the main catalytic kinase component of TORC1 [5–7]. TORC1 is essential for growth of all organisms, including fission yeast, [28,29]. We therefore incubated temperature sensitive tor2.51 [5] mutants at 37°C to inactivate TORC1 activity. Ppk32 levels were maintained at 37°C (the restrictive temperature) when TORC1 was inactivated (Fig 2A). In contrast, Ppk32 levels in wild type cells were significantly reduced 2 hours following a shift to 37°C in wild type cells (Fig 2A) (note: the level of an unknown protein, recognised by anti-ppk32 antibodies in ppk32.Δ mutants, is strongly induced by heat stress). In a similar heat stress experiment of wild type cells, TORC1 dependent Maf1 phosphorylation increased 1 hour upon after shift to 37°C (S2A Fig), suggesting that TORC1 may increase 1 hr after heat stress (S2A Fig). These results suggest that Ppk32 protein turnover requires TORC1 activity. Steady state Ppk32 proteins levels were reduced in strains that were TORC2 deficient because they lacked Ste20 (the fission yeast version of the conserved TORC2 specific component Rictor) (Fig 2B), a similar decrease in Ppk32.PK total protein levels was seen in the ste20.Δ mutant (S2C Fig). No significant reduction in Ppk32 levels was observed in tor1.Δ mutants, that lack the main TORC2 kinase, however Tor2 kinase can also interact with Ste20 and this interaction is enhanced in the absence of Tor1 [30]. Simultaneous inhibition of both TORC1 and TORC2 through treatment with Torin1 did not significantly change Ppk32 levels (Fig 2C). In summary, TORC1 negatively regulates Ppk32 protein levels whilst TORC2 is required to maintain steady state levels.
The resistance to Torin1 induced TOR inhibition in Ppk32 deficient mutant cells suggests that Ppk32 opposes TOR signalling (Fig 1C). Furthermore, TORC1 inhibition in tor2.51 mutants, enhanced Ppk32 levels when compared to wild type cells (Fig 2A). We therefore used colony size as a good proxy for cell fitness [31] to ask whether Ppk32 contributes to the reduced growth rate fitness of tor2.51 cells. Ppk32 removal enhanced tor2.51 fitness (Figs 2D and 4A). Loss of ppk32 also had a less pronounced, positive, impact on the fitness of tor1.Δ cells (Figs 2D and 4C). Thus, Ppk32 reduces the fitness of cells compromised for either TORC1 or TORC2 activity. Ppk32 may therefore constitute a novel inhibitor of TOR signalling. To test this possibility ppk32+ was overexpressed in wild type cells grown in minimal media to facilitate high levels of expression. Ppk32 over-expression dramatically reduced cell growth (S3A Fig) and this block to cell proliferation correlated with a decrease in the TORC1 and TORC2 dependent Maf1.PK and Gad8 S546 phosphorylation respectively (S3B Fig). In summary, elevation of Ppk32 levels reduced both TORC1 and TORC2 signalling, while abolition of ppk32 expression provided Torin1 resistance and stimulated growth of TORC1 and TORC2 mutants.
Brefeldin A (BFA) inhibits Guanine nucleotide Exchange Factors (GEFs) for ARF GTPases [16,17]. ARFs regulate COPI recruitment to Golgi membranes to facilitate retrograde trafficking within the Golgi and from the Golgi to the ER [32,33]. Importantly BFA can also induce ER stress through the unfolded protein response (UPR) and impair traffic to lysosomes [19]. We next determined whether the BFA sensitivity of fission yeast observed here (Fig 1C) arose from ER stress. In fission yeast UPR is regulated by Ire1 also known as Ppk4 [34]. The ire1.Δ mutant was not sensitive to BFA whereas double ppk32.Δ ire1.Δ deletion mutant was (S4A Fig). This result shows that the BFA sensitivity we observe arises from either Golgi stress or impaired traffic to lysosomes (yeast vacuoles). Interestingly, the human Ppk32 homolog the SCYL1 kinase has been implicated in retrograde traffic between the Golgi and ER [22]. SCYL1 is thought to act as a scaffold protein that recruits ARFs to COPI complexes, furthermore, BFA treatment releases ARFs from Golgi membranes in to the cytosol [35,36]
We next assessed the impact of BFA on TOR signalling in wild type and Ppk32 deficient cells. Previous work established that BFA treatment reduced mTORC1 activity [37]. In wild type fission yeast cells both TORC1 and TORC2 activity were reduced by BFA treatment (Fig 3A and 3B). Importantly, in ppk32.Δ mutants, the BFA induced TOR inhibition of both complexes was less pronounced to suggest that the BFA sensitivity of ppk32.Δ mutants (Fig 1C) might arise from increased TOR activity in this mutant. To test this hypothesis, we generated a ppk32.Δ tor2.51 double mutant in which TORC1 signalling would be reduced by the mutation in Tor2. The ppk32.Δ tor2.51 double mutant was less sensitive to BFA than the single ppk32.Δ mutant alone (Fig 4A) to imply that reduced TORC1 signalling may promote survival upon exposure to BFA induced stress [26]. To further test this hypothesis rapamycin was added to BFA treated cells. Importantly although rapamycin treatment does reduce TORC1 activity (S2B Fig), the reduction in activity is not sufficient to block cell growth in fission yeast [38]. Addition of rapamycin to inhibit TORC1 activity in ppk32.Δ mutants abolished their sensitivity to 15μg ml−1 BFA (Fig 4B). Furthermore, inhibition of TORC1 in wild type cells with rapamycin promoted growth at a higher BFA concentration (20μg ml−1) (Fig 4B).
Rapamycin inhibits TORC1 activity when in a complex with FKBP12 [3] (known as Fkh1 in fission yeast [29]). To determine whether the impact of Rapamycin on BFA sensitivity was an indirect one of drug interference in which rapamycin affected BFA uptake, and thus bioavailability, we exposed fkh1.Δ mutants to BFA stress. However, the fact that, in contrast to wild type cells, rapamycin did not enhance growth of BFA treated fkh1.Δ mutants (S5A Fig), excludes the possibility that rapamycin blocked BFA uptake. We therefore conclude that it is the reduction in TORC1 activity that promotes growth upon BFA induced stress.
Addition of Torin1 to inhibit both TORC1 and TORC2, also enhanced cell survival during BFA induced stress (S5B Fig). 20μg ml−1 BFA blocked growth of ppk32.Δ mutants. Addition of Torin1, but not rapamycin, liberated this block to cell growth (Fig 4B, S5B Fig). As mentioned previously, rapamycin reduces but does not completely block TORC1 activity (S2B Fig), whereas a complete block of TORC1 activity is achieved by treatment with Torin1 [24,26]. This may explain the increased potency of Torin1 over rapamycin in combination with BFA. Alternatively a combined reduction of TORC1 and TORC2 activity may be beneficial for survival following BFA induced stress.
We next exposed TORC2 deficient mutants to BFA to assess the role of TORC2 in BFA induced stress. Both tor1.Δ and tor1.Δ ppk32.Δ double mutants were sensitive to BFA and the TORC2 specific mutant ste20.Δ was hyper-sensitive to BFA (Fig 4C and 4D). Furthermore, the constitutively active TORC2 mutant tor1.I1816T [39] was also hyper sensitive to BFA (Fig 4E). This suggests that TORC2 activity is essential for cell fitness during BFA induced stress, but only at moderate levels. Interestingly, BFA resistance was restored in the tor1.I1816T ppk32.Δ double mutant (Fig 4E), suggesting that the hypersensitivity of the constitutively active TORC2 mutant tor1.I1816T is, at least partially, Ppk32 dependent. When considered together these observations indicate that the BFA-induced decrease, but not full inhibition, in TORC2 activity (Fig 3B) that we observed in wild type cells may be important for cell survival. Finally, neither BFA treatment nor the presence of the tor1.I1816T mutation had any impact on Ppk32 protein levels (Fig 4F, S6A Fig).
In summary, our data demonstrate that enhanced TORC1 activity accounts for the BFA sensitivity of ppk32.Δ mutants. Decreased TORC1 activity promoted cell survival and recovery from BFA induced stress.
Human SCYL1 regulates retrograde traffic to modulate Golgi homeostasis [15,22,35], whereas SCYL2 regulates clathrin-mediated endocytosis [40,41]. Despite the high level of similarity between SCYL2 and Ppk32, ppk32.Δ cells were proficient for endocytosis. FM4-64 is a lipophilic dye that incorporates into the plasma and vacuolar membranes via endocytosis [42]. FM4-64 staining of vacuoles in ppk32.Δ mutants (Fig 5A) identified a modest increase in vacuolar size to suggest that an absence of ppk32 expression favours vacuole fusion and/or diminishes vacuolar fission.
In the presence of cells of both mating types nitrogen starvation induces sexual differentiation to promote cell fusion, meiosis and sporulation [43,44]. Vacuoles play an important role during sexual differentiation through their participation in autophagy [45]. Autophagy deficient mutants are sterile but can complete differentiation when nutrients are resupplied [45]. High TORC1 signalling represses sexual differentiation [5,27,46]. The inhibition of TORC1 signalling, arising from nutrient starvation induces differentiation and so promotes cell fusion. Importantly however, TORC1 signalling must be reinstated to support further differentiation because diploid mutants with deficiencies in the fission yeast raptor, Mip1, are unable to execute meiosis and sporulation [47]. Thus, after its initial repression has been triggered by nutrient reduction, TORC1 is reactivated by autophagy-supplied nutrients.
The survival of cells deleted for ppk32 following prolonged periods of nutrient deprivation indicates that ppk32.Δ mutant cells are autophagy proficient (S6B Fig). Nitrogen starvation promoted TORC1 inhibition in both wt and in ppk32.Δ cells (Fig 5B), however TORC1 activity returned more rapidly to ppk32.Δ cells (Fig 5B) and sexual differentiation was increased by a further 10% eight hours after starvation (Fig 5C). Pheromone secretion was also enhanced in ppk32.Δ mutant cells (Fig 5D). A “halo assay” can be used as a proxy for pheromone secretion. It deploys a diploid tester strain that cannot sporulate unless external pheromone (M-factor) is supplied [48]. Sporulating diploid tester strains can be visualized because the starch in their spore walls generates a dark stain when exposed to iodine vapour. Mixing a tester strain with wt or ppk32.Δ mating partners at a ratio of 200:1 revealed increased pheromone secretion from ppk32.Δ cells (Fig 5D and 5E). In summary, the presence of Ppk32 during sexual differentiation retards TORC1 reactivation to delay sexual differentiation.
Global phosphoproteomic approaches have revealed that Ppk32 is phosphorylated on the carboxyl terminal Serine 630 and Serine 632 [49–51]. These phosphorylation sites are conserved throughout the SCYL family of kinases (Fig 6A) but are not present in the other fission yeast SCYL homolog Ppk3 that has no impact upon sensitivity to Torin1 (Fig 1C). To assess the significance of phosphorylation on Ser630 and Ser632 upon the function of S. pombe Ppk32, the endogenous ppk32+ locus was manipulated to simultaneously mutate both sites to either alanine to block signalling or to aspartic acid to mimic constitutive phosphorylation. Ppk32 could not be detected in the ppk32.DD (ppk32.S630D-S632D) mutant whereas levels were modestly enhanced in ppk32.AA (ppk32.S630A-S632A) cells (Fig 6B). Both mutations elevated tolerance to Torin1 (Fig 6C). The low Ppk32 levels in the ppk32.DD mutants are reminiscent of a ppk32.Δ and thereby account for the Torin1 resistance of this mutant. In contrast, the resistance to Torin1 of ppk32.AA is unlikely to arise from changes in protein level. Interestingly, comparisons with wild type controls revealed that the phosphoblocking mutation ppk32.AA reduced the amount of Ppk32 that immuno-precipitated with Tor1 (Fig 6D). Thus, phosphorylation at these residues appears to modulate the affinity of Ppk32 for Tor1. In line with the imposed resistance to Torin1, both phosphorylation site mutants conferred sensitivity to BFA induced stress (Fig 6C), to further support the view that phosphorylation is important for protein function.
Next, phospho-specific antibodies that recognise Ppk32 when phosphorylated on both S630 and S632 were generated by Eurogentec (S7A Fig). Consistent with the inability to detect any Ppk32 in the ppk32.DD mutant (Fig 6B) we were unable to detect Ppk32 phosphorylation in wild type cells (Fig 6E). To facilitate the detection of Ppk32 phosphorylation ppk32+ was over-expressed from the nmt1 promoter [52] (Fig 6E) in wild type cells and TORC1 deficient tor2.ts mutants that had been subjected to heat stress at 37°C (Fig 6F). Comparison of Ppk32 phosphorylation with Ppk32 total proteins levels in these heat stressed cells revealed that phosphorylated Ppk32 was degraded in wild type cells, but not in the TORC1 deficient tor2.ts mutant (Fig 6F and 6G & S7B Fig). When considered alongside the TORC1 dependent turnover of wild type Ppk32 (Fig 2A), these data suggest that TORC1 activity promotes the degradation of phosphorylated Ppk32.
The regions adjacent to the Ser630 and Ser632 phosphorylation sites reside in the part of the protein that is most highly conserved in human SCYL1 (Fig 6A). The carboxy terminal RKXX-COO− motif of SCYL1 binds to coatomer (a COP1 protein complex in an ARF independent manner to regulate retrograde transport and Golgi morphology) [22,36]. This SCYL1 RKXX motif and the sourounding sequence is partly conserved in Ppk32 (Fig 7A). Ppk32 has a mono-basic motif followed by a serine in place of the di-basic motif found in SCYL1 (Fig 7A). Non-conventional mono basic motifs have previously been shown to promote binding to COP proteins [53]. To assess the significance of Ppk32 carboxy-terminal sequence Lys745 and Ser746 they were simultaneously mutated to alanine at the native locus to generate the ppk32.L745A-S746A (ppk32.AAXX) allele (Fig 7A). Ppk32 protein levels were unaffected by the presence of the ppk32.AAXX allele (Fig 7B), however, the mutant gene conferred both resistance to Torin1 and sensitivity to BFA (Fig 7C). Thus, the Ppk32 carboxy-terminal domain of Ppk32 is required for function.
Here we describe a novel control of TORC1 signalling by the SCYL family pseudo-kinase Ppk32 that is essential for survival following BFA induced stress. As the impact of BFA on cell growth that we report here is unlikely to arise from ER Stress (S4A Fig), it probably arises from either Golgi stress or impaired traffic to lysosomes. It is the response to this stress that requires Ppk32 dependent control of TORC1 activity. Interestingly the system appears to involve negative feedback from TOR to Ppk32 as TOR signalling regulates Ppk32 protein levels (summarised in Fig 7D1). At present, it is unclear where in the cell this TOR mediated control of Ppk32 protein levels occurs and where Ppk32 control of TOR signalling takes place.
We show that Ppk32 phosphorylation on two conserved residues is important for protein function. Ppk32 phosphorylation dramatically reduced protein levels and conferred resistance to TOR inhibition. Similarly, despite the increase in protein levels that arises as a consequence of an inability to phosphorylate Ppk32, lack of phosphate on these residues provides resistance to TOR inhibition and highlights the functional significance of these phosphorylation events. Importantly our data suggest that Ppk32 phosphorylation promotes the association of Ppk32 with Tor1.
The impact of Ppk32 on global TOR signalling in non-stressed, steady state conditions is small but measurable, as Ppk32 deficient cells are resistant to Torin1 induced inhibition of TOR signalling (Fig 1C). Our data show how this impact on TOR signalling is dependent on the composition of growth media. A previous report has described an impact of media composition (rather than starvation) on cell growth and chronological lifespan (a process regulated by TOR signalling) to suggest that the steady state level of TOR signalling is regulated by the nutrient composition [54]. We found that Ppk32 deficient cells that were autotrophic for leucine biosynthesis displayed increased resistance to Torin1. Leucine is a well-known activator of TOR signalling in a process that occurs on the endomembranes of vacuoles (yeast lysosomes) [55]. TORC1 concentrates on vacuolar endomembranes [20]. Activation of TORC1 by the addition of leucine may be more efficient in ppk32.Δ cells because these cells lack an apparent inhibitor of TOR signaling. However, whether phosphorylated Ppk32 interacts with vacuolar TORC1 is unclear at present.
Ppk32 shares sequence similarity with both human SCYL family kinases. However, the requirement of the Ppk32 C-terminal for survival following BFA induced stress and the fact that lack of ppk32+ expression has no impact on endocytosis (a SCYL2 function) suggests that, despite slightly higher level of homology to SCYL2, Ppk32 appears to be a functional SCYL1 homologue in the context of BFA induced stress.
It is unclear whether Ppk32 localizes to the Golgi or COPI budding vesicles. In fission yeast Golgi localisation is seen as small cytoplasmic foci [56]. We did observed small foci in the cytoplasm, however the molecular composition of these is unclear at present. A systematic screen of protein localizations in fission yeast, reported the recruitment of an over-expressed Ppk32 protein to Golgi-like structures [57], suggesting that Ppk32 may localize to the Golgi in fission yeast. Interestingly The RSXX carboxy-terminal motif of Ppk32 is important for cells to tolerate BFA induced stress. Because mTORC1 is localized to the Golgi [58] and to lysosomes [55], phosphorylated Ppk32 may facilitate control of TORC1 signalling on either endomembrane system, however this is unclear at present. Localized control of TORC1 activity would explain why a lack of ppk32+ expression had no impact on TORC1 control of Maf1 (cytoplasmic protein) phosphorylation. Future EM based localisation with specific Ppk32 antibodies will determine whether Ppk32 resides on the Golgi and on vacuoles.
SCYL1 regulates both retrograde trafficking of COPI-coated vesicles from Golgi to ER and transport between Golgi stacks [22,35]. Recently SCYL1 was shown to act as a scaffold that supports the interaction between ARFs and COPI complexes [36]. BFA inhibits the GEFs for class II ARFs [16,17] to release ARF into the cytosol. We show that Ppk32 controlled TORC1 inhibition following BFA induced stress was essential for survival. This Ppk32 controlled reduction in TORC1 signalling will slow down cell metabolism and proliferation. This impairment of metabolism is likely to assist recovery from stress to the membrane traffic system and therefore support survival of BFA stressed cells (summarised in Fig 7D2). We show that a simultaneous reduction of TOR activity, by combining treatment with BFA with additional TOR inhibitors, enhanced cell recovery from BFA induced stress, irrespective of Ppk32 status. Interestingly, one hallmark of cancer is an alteration of the glycosylation landscape that arises from modified Golgi function. Our findings suggest that therapeutic rapamycin treatment to reduce TORC1 signalling and proliferation of cancer cells that are experiencing inherent Golgi stress, might actually have the unintended consequence of enhancing cell survival. It is notable therefore that the control of phosphorylation at the sites in SCYL1, that are equivalent to those identified in fission yeast Ppk32, would be predicted to reduce SCYL1 protein levels and so increase cell death and treatment of cancer types with inherent Golgi stress.
Strains used in this study are listed in Table 1. Cells were exponentially grown for 36 hr at 28°C in YES [59] 0.15 mg ml−1 adenine hemisulphate, uracil, L-leucine and histidine (each), or for 48 hr at 28°C either EMM-G or EMM-P [30]. L-leucine was added to a final concentration of 150 μg ml−1. For western blots and growth assays cells were grown to a density of 2×106 cells ml−1. For nutrient starvation cells were grown in YES and at the density of 2×106 cells ml−1 cultures were filtered into sporulation liquid SPL [60]. For mating assays, cells were grown to a density of 2×106 mixed in a ratio of 1:1 h+:h− 1.5x106 cells of each strain, washed with water and spotted onto SPA (SPL with 30 g l−1 agar). Mating efficiency was calculated as [61]. For halo assays, cells were grown to a density of 2×106 and mixed with h− cell in ratio: 200:1, washed with water and spotted onto MSA [48].
The ppk32 open reading frame was amplified by PCR from genomic wt DNA, sequenced and cloned into Nde1 of the pRep1 plasmid to facilitate Ppk32 overexpression from the nmt1 promoter [52]
In order to analyse vacuoles morphology, cells were grown in YES to the density of 1.8 × 106 cells ml−1 before FM4-64 dye 5mg/ml was added to the culture in ratio: 1:10 000 (dye:culture, v/v) and cells were incubated further. After 30 min, stained vacuoles in live cells were analyzed under the microscope. For each culture at least 500 vacuoles were measured.
Total protein extracts were prepared by TCA precipitation [62].
For Tor1 IP, 3x108 cells (culture density 2,4x106 cells ml−1) were harvested and re-suspended in IP buffer (50 mM HEPES pH 7.5, 150 mM NaCl, 0,1% CHAPS, 0,05% Tween20, 50 mM L-arginine, 50 mM L-glutamic acid, 50 mM NaF, 2 mM NaVO4, 60 mM sodium glycerol phosphate, 5 mM NEM, 1 mM PMSF, 1 mM DTT and Sigma protease inhibitor cocktail) and broken in a FastPrep using glass beads. The cell extract was spun (10 000 rpm, 10 min, 4°C) and supernatant was incubated with Invitrogen protein Dynabeads pre-coupled with anti-Tor1, anti-V5 (PK) or anti-myc antibodies for 20 min at 4°C. Beads were then washed three times with IP buffer and heated to 80°C for 10 min to elute the proteins. Samples were loaded on a SDS-PAGE gel and subsequently processed as total protein extracts.
Proteins were detected using the following antibodies: 1:2000 anti-Ppk32 (polyclonal antibodies raised by Eurogentec Ltd to the following peptide PSEARTPSVQPANRR in rabbits), 1:2000 anti-V5 (anti-PK) (AbD Serotec), 1:1000 anti-phospho Ser546 Gad8 [26], 1:100 Gad8 [26], 1:1000 anti-Tor1 [39], 1:2000 anti-PAS (Cell Signaling Technology), 1:2000 anti-Rps6 (Abcam), anti-phospho-Ser630 S632 Ppk32 (generated by Eurogentec). Alkaline phosphatase coupled secondary antibodies were used for all blots, followed by direct detection with NBT/BCIP (VWR International Ltd.) substrates on PVDF membranes (Millipore). Ponceau S staining was used as a loading control [63]. Relative protein levels were quantified in ImageJ and GraphPad Prism was used to calculated significant differences. Asterisks represent statistical significance (p<0.05) as determined by a Student’s t-test.
|
10.1371/journal.pcbi.1000616 | Meta-analysis of Inter-species Liver Co-expression Networks Elucidates Traits Associated with Common Human Diseases | Co-expression networks are routinely used to study human diseases like obesity and diabetes. Systematic comparison of these networks between species has the potential to elucidate common mechanisms that are conserved between human and rodent species, as well as those that are species-specific characterizing evolutionary plasticity. We developed a semi-parametric meta-analysis approach for combining gene-gene co-expression relationships across expression profile datasets from multiple species. The simulation results showed that the semi-parametric method is robust against noise. When applied to human, mouse, and rat liver co-expression networks, our method out-performed existing methods in identifying gene pairs with coherent biological functions. We identified a network conserved across species that highlighted cell-cell signaling, cell-adhesion and sterol biosynthesis as main biological processes represented in genome-wide association study candidate gene sets for blood lipid levels. We further developed a heterogeneity statistic to test for network differences among multiple datasets, and demonstrated that genes with species-specific interactions tend to be under positive selection throughout evolution. Finally, we identified a human-specific sub-network regulated by RXRG, which has been validated to play a different role in hyperlipidemia and Type 2 diabetes between human and mouse. Taken together, our approach represents a novel step forward in integrating gene co-expression networks from multiple large scale datasets to leverage not only common information but also differences that are dataset-specific.
| Two important aspects of drug development are drug target identification and biomarker discovery for early disease detection, disease progression, drug efficacy and drug toxicity, etc. Recently, many single nucleotide polymorphisms (SNPs) associated with human diseases are discovered through large genome-wide association studies (GWAS). However, it is still largely unclear how these candidate SNPs may cause human diseases. The ultimate aim of this paper is to put these GWAS candidate SNPs and their associated genes into a network context to understand their mechanism of action in human diseases. In addition to large-scale human data sets that are often heterogeneous in terms of genetic and environmental factors, many high quality data sets in rodents exist and are frequently used to model human diseases. To leverage such information, we developed a method for combining and contrasting gene networks between human and rodents, specifically to elucidate how GWAS candidate SNPs may contribute to human diseases. By identifying mechanisms that are conserved or divergent between human and rodents, we can also predict which disease causal genes can be studied using rodent models and which ones may not.
| The advent of expression profiling and other high throughput technologies has enabled us to systematically study complex human diseases by simultaneously measuring tens of thousands of molecular species in any given cell-based system [1]. It is now routine to organize such large-scale gene expression data into co-expression networks to shed light on the functional relationships among genes, and between genes and disease traits [2],[3],[4],[5]. Analysis of co-expression networks can be used to study any tissue or organ (such as liver, which plays a key role in the metabolism of glucose, lipids and toxic compounds), as long as the samples from such organs are collected in a population setting. Given that mouse and rat populations are commonly used to study human diseases in this manner, it is important to understand the mechanisms that are conserved between human and the rodent species, especially as we seek better predictions of the efficacy of drug targets identified from mouse or rat in human populations. In addition, identifying mechanisms that differ between humans and rodents can help to improve the design and interpretation of toxicity studies that involve rodent models.
Meta-analysis is the statistical synthesis of data by aggregating results from a set of comparable studies [6]. It can be used to systematically examine similarities and differences between molecular profiling studies carried out in populations from different species [7]. In a gene co-expression network, relationship between gene pairs is usually measured by correlation coefficients of different forms, such as Pearson correlation, Spearman correlation, or Mutual Information. Therefore, the problem of combining or comparing co-expression relationships across multiple datasets can be framed in the context of a meta-analysis of correlation coefficients, for which various methods have already been introduced. One method is Fisher's Inverse test, which computes a combined statistic (S) from the p-values of the correlation coefficients obtained from (k) individual datasets as, . Under fairly general conditions this statistic follows a distribution with degrees of freedom under the joint null hypothesis of no correlation, making it possible to compute p-values of the combined statistic.
Another widely used meta-analysis method involves computing a weighted average of a common metric (i.e. effect size) derived from correlation coefficients in the individual datasets. Such statistic can then be used to test for homogeneity over the individual measures and for statistical significance. Datasets in this type of meta-analysis are typically weighted by the accuracy of the effect size they provide, which is a function of the individual sample sizes. Once the mean effect size is calculated, its statistical significance can be assessed by estimating the pooled variance of the mean effect size. In defining the effect size, Hedges and Olkin [8] and Rosenthal and Rubin [9] both advocated converting the correlation coefficient into a standard normal metric using Fisher's Z-transformation and then calculating a weighted average of these transformed scores. Depending on whether the effect sizes are assumed to be equal or not in the multiple datasets, fixed effect as well as random effect models can be employed. In the fixed effect models, the effect size in the population is a fixed but unknown constant and therefore is assumed to be the same for all datasets included in the meta-analysis. For random effect models, effect sizes may vary from dataset to dataset, and are assumed to be a random sample of all population effect sizes. Hunter and Schmidt [10] introduced a single random-effects method based on untransformed correlation coefficients. One important feature of this type of method is that heterogeneity of the effect sizes can be estimated, which provides a way to assess the difference in correlation coefficients across multiple datasets. Schulze [11] provided a thorough review of these meta-analysis methods and their applications.
For a meta-analysis of co-expression networks from diverse datasets, such as those constructed from different species, one central issue is that it is often unreasonable to assume that every gene pair has a unique, true effect size across evolutionarily diverse species. Although random effect models provide a more realistic way to accommodate cross species variation, it still assumes a parametric distribution on the population effect sizes. To circumvent this problem, a non-parametric meta-analysis method was introduced for the identification of conserved co-expression modules from human, fly, worm and yeast [7]. In this method, Pearson correlation coefficients of expression profiles between every gene pair were computed in each organism and then rank-transformed according to their correlations with all other genes. A probabilistic test based on order statistics was then applied to evaluate the probability of observing a particular configuration of ranks across the different organisms by chance. The advantage of this method is two-fold: 1) because the method is based on non-parametric statistics, it makes no assumption on the underlying distribution of correlation coefficients across multiple datasets; and 2) the effect size (i.e. the rank ratio statistic for every gene pair) is defined in a gene-centric fashion such that for any given gene, correlations with all other genes are considered. However, the method also has several limitations including 1) the loss of power in general given the non-parametric formulization [12],[13], and 2) the meta-analysis results cannot be represented in the same format as the individual datasets given there is no concept of a mean effect size. The details of individual methods are presented in the Methods section. Their pros and cons are summarized in Supplementary Table S1.
In this paper, we develop a method for the meta-analysis of diverse datasets generated across multiple species. Our method is semi-parametric in nature, requiring fewer assumptions on the distribution of the effect size than a purely parametric approach while retaining better statistical power than a fully non-parametric method. It also 1) defines an effect size that is gene centric, 2) allows for the computation of a mean effect size, and 3) leads to a heterogeneity statistic to test for differences in correlation structures among distinct datasets. Unlike most network alignment algorithms [14],[15],[16],[17],[18] (with the exception of [19]) or connectivity-based approaches [20], our method does not rely on the networks inferred a-priori from individual datasets, but instead focuses on the development of rigorous statistics to test directly the relationship between every gene pair. The simulation results showed that our method is robust against noises. When applied to a human, mouse and rat cross species meta-analysis of liver co-expression networks, we demonstrate that our method out-performs existing methods in identifying functionally coherent gene pairs that are conserved among the three species. Our method also leads to the identification of modules of co-expressed genes that represent core functions of the liver that have been conserved throughout evolution. Both highly replicated and less confident genome-wide association study (GWAS) candidate genes for blood lipid levels are found to be enriched in the conserved modules, providing a systematic way to elucidate the mechanisms affecting blood lipid levels. Application of our test for homogeneity leads to the identification of a single sub-network driven by ApoE that distinguishes two nearly identical experimental cross populations whose genetic backgrounds only vary with respect to the gene ApoE. We further demonstrate that genes involved in human- or rodent- specific liver interactions tend to be under positive selection throughout evolution. Finally, we identified a human-specific sub-network regulated by RXRG, which has been validated to play a different role in hyperlipidemia and Type 2 diabetes between human and mouse. Taken together, our approach represents a novel step forward in integrating gene co-expression networks from multiple large scale datasets to leverage not only conserved information but also differences that are dataset-specific.
The intuition behind our meta-analysis approach in the cross-species setting is that, instead of directly comparing the correlation coefficients of a gene pair as an absolute measure of co-expression, which depends on many features such as sample size, expression dynamics, measurement noise, and confounding factors that are usually not well-controlled among the individual datasets, we measure the co-expression relationship as a relative distance with respect to each gene's total relationship to all other genes in each dataset. When the correlation coefficients between a given gene and all other genes were rank-transformed into a uniform distribution, the inter-relationships among the correlations were destroyed. Unlike the previous method [7] we assume the distribution of correlation coefficients of one gene to all other genes follows a normal distribution under the condition that the numbers of samples and genes are large (see Materials and Methods section for details). In fact, for roughly 70–90% of the expression traits in our datasets, the distributions of their correlation coefficients to all other expression traits are well supported as being normal by the Kolmogorov-Smirnov test (Figure S1). Based on this assumption, we define for gene pair (i,j) in dataset , the effect size of its co-expression according to Glass's d score definition [21] as:where is the correlation coefficient between the expression profiles of (i,j) in dataset , and and are the mean and standard deviation of the null distribution, respectively, of the correlation coefficients between gene and all other genes. Essentially, by this definition we transform the correlation measure into a relative distance to the gene-centric mean in terms of standard deviation units. This transformation not only normalizes all effect sizes, but also takes into account the context of each gene in individual datasets. It is of further note that our effect size definition is directional, i.e. is usually different from due to differences in the neighborhoods of gene and . For simplicity, we drop the superscript so that represents the effect size for any gene pair in dataset .
Using a meta-analysis procedure for d score that developed by Hedges and Olkin [8], we can compute the mean effect size as:and the standard deviation of the mean effect size as:The statistical significance of the mean effect size can then be assessed by forming the Z-score statistic:
In addition, heterogeneity of the effect sizes across the datasets can be estimated by the statisticwhich follows a distribution with degree of freedom under the null hypothesis of homogeneous effect sizes.
Given the mean effect size and heterogeneity statistic, a flowchart of our method is summarized in Figure 1. Briefly, the first step begins by computing correlation coefficients for all gene pairs in every dataset. Correlation can be measured by the Pearson or Spearman correlation, depending on the properties of the datasets being analyzed. The method then proceeds by iterating through all gene-pairs one at a time, computing the heterogeneity statistic for every gene-pair. If homogeneity is not rejected at a pre-specified significance level, the mean effect size for the gene-pair is computed and tested for deviation from zero. A statistically significant mean effect size is then considered as a conserved co-expression relationship among the datasets being compared. On the other hand, if the homogeneity of the effect sizes is rejected, the gene-pair is considered as a candidate for change in co-expression relationships, termed differential interactions hereafter, between the datasets. In this case, the direction of change can be determined by examining the actual effect sizes in single datasets.
To compare the performance of our semi-parametric method with the existing parametric and non-parametric methods, we ran several simulations. In each simulation, 3 independent data sets were generated assuming the underlie structure is modular as shown in Figure S2 (see Materials and Methods section for details). There were 150 samples and 2000 genes in each data set. The signal strength is measured by the correlation between the latent regulators and their downstream genes. The signal strengths were different for the 3 simulated data sets, shown in Figure 2A. When there was no systematic noise, the parametric methods (FEM Fisher-Z and combine p-value) performed better than non-parametric method, shown in Figure 2B. It is consistent with other studies' results that there are power losses in general for non-parametric methods [12],[13]. The performance of our semi-parametric method was between the parametric methods and the non-parametric method. It is consistent with the nature that our semi-parametric is a hybrid of parametric and non-parametric methods. It is worth to note that the random effect model (REM Fisher-Z) performed worst among methods tested even though the effect sizes were different as shown in Figure 2A.
When the systematic noises were moderate (measured by the correlation between genes and systematic noises) as shown in Figure 2C, the performances of our semi-parametric method and the parametric methods were similar, shown in Figure 2D. When the systematic noises were stronger (shown in Figure 2E), the performances of parametric methods decreased significantly, and our semi-parametric and non-parametric methods were robust against systematic noises (shown in Figure 2F). Under all conditions, our semi-parametric method performed better than the non-parametric method.
We applied our method to identify conserved co-expression interactions among 6,455 orthologous genes in human, mouse and rat (see Materials and Methods for details about the data, data preparation and orthologous gene identification. The 6,455 genes are listed in Table S2. The 2-D hierarchical clustering views of individual data sets are shown in Figure S3, and ordered sample and gene annotations are listed in Table S3, S4, S5, S6, S7, S8). We used the absolute Spearman correlation coefficient between the expression profiles of a gene pair as the measure of co-expression interaction. By doing this we considered only the magnitude of gene-gene correlation, but not its direction, since the same gene-gene relationship may manifest as either a positively or negatively correlated expression profile due to feedback control [4]. Specifically, our method inferred 20,230 conserved co-expression interactions, covering 4,885 genes, at a p-value cutoff of , corresponding to a Bonferroni corrected false positive rate of 5% (i.e. ) for both effect size and the heterogeneity . The false discovery rate (FDR) of this result is estimated to be based on a permutation test procedure where sample labels were randomly shuffled for each gene independently in every dataset (see Materials and Methods for details). These conserved interactions represent approximately 2.4–15.2% of the co-expression interactions obtained using single species data, given there were 828,031, 334,721 and 132,884 interactions in human, mouse and rat, respectively, at the same statistical significance p-value threshold.
We benchmarked the performance of our method against existing meta-analysis methods in the literature, as well as against the interactions previously reported for single species co-expression networks [22]. The number of predictions (i.e. conserved interactions) inferred by our method lies in between the numbers predicted by existing parametric and non-parametric meta-analysis methods at a common FDR threshold, shown in Table S9, consistent with the semi-parametric nature of our approach. When only considering the same number of top confident predicted pairs, the qualities of the semi-parametric method were better than other methods in terms of coherences with both Gene Ontology (GO) biological processes and curated KEGG pathways (shown in Table S10). To test the full range of predictions, we generated precision vs. coverage curves for each method by varying the statistical significance thresholds and computing 1) the percent of inferred gene pairs that share a common GO biological process annotation, and 2) the percent of inferred gene pairs that share a common curated KEGG pathway (Figure 3). Two conclusions stand out from these results. First, all meta-analysis methods outperform the inference based only on single species datasets, likely due to the increased precision achieved by incorporating evolutionary information and the added power achieved by integrating multiple datasets. Second, our method clearly outperformed all existing meta-analysis methods across the full spectrum of coverage, but most significantly at the stringent p-values. This demonstrates the added value of combining the advantages of existing methods.
We next performed spectral clustering of the orthologous genes based on their interconnectivity in the conserved co-expression network and identified co-expressed gene modules, shown in Figure S4 (see Materials and Methods for the spectral clustering method). Table 1 summarizes the top 13 modules comprised of greater than 20 genes and their enrichment for GO biological process terms. Almost all of the modules are observed to be coherent with respect to some biological processes and many of the indicated processes represent core biological processes in the liver, including immune response (p<2.70×10−43), carboxylic acid metabolic process (p<6.6×10−16), and sterol biosynthetic process (p<1.9×10−27). It is of particular note that these modules differ from modules identified in single species datasets in that the genes in modules of the conserved co-expression network are functionally related based on evolutionary conservation, rather than on correlated gene expression alone.
Recent human genome-wide association studies have identified many candidate genes affecting blood lipid concentrations. However, the mechanisms by which many of these candidate genes contribute to blood lipid concentration remains unclear [23]. In addition, there are potentially many SNPs with weaker associations to lipid concentration that are difficult or impossible to detect or replicate given the lack of power in current GWAS [24]. Therefore, an open question is whether there are many more genes harboring common variation that affect the polygenetic nature of lipid concentration regulation. Because liver is a key tissue for lipid metabolism, we can use the liver networks to interpret the GWAS results and generate hypothesis regarding the mechanisms of the candidate genes. Toward this end, we selected 30 recently identified lipid-associating loci [25] and assessed the ability of our conserved modules to annotate the 45 candidate causal genes nominated from these 30 loci. Of the 45 candidate genes, 26 have orthologs in human, mouse and rat and were therefore included in our study. Nineteen of these genes reside in human, mouse and rat conserved modules (Table 2), where the putative mechanisms with respect to lipid regulation can be annotated based on the module functions. The results suggest that cellular processes such as sterol biosynthetic process and cell-cell communication are involved in regulating blood lipid concentration. Of particular note is SORT1, a gene that resides at the locus most significantly associated with LDL cholesterol [25]. Based on the conserved modules, SORT1 belongs to module 1, a module enriched for genes involved in cell-cell signaling (p-value<6.51×10−23). Other candidate genes at lipid associated loci, such as GALNT2 and NCAN, also reside in module 1, suggesting that cell-cell signaling is important for blood lipid regulation. PCSK9 is clearly annotated as being involved in the sterol biosynthetic process along with FADS1, FADS2, HMGCR and MVK. In contrast, only 14 of 26 candidate genes can be annotated based on modules derived from the human co-expression networks alone (Table 2). The annotations of these genes based on the conserved modules are closer to their known functions than ones based on the human modules (shown in Table S11). For example, MAFB is annotated as “transcription regulation” based on the conserved modules, but as “carboxylic acid metabolic” based on the human-only modules, whereas its annotation in GO is “positive regulation of transcription from RNA polymerase II promoter”. These examples illustrate how the conserved human, mouse and rat modules can enhance the interpretation of GWAS and the annotation of candidate genes identified from these studies.
Blood lipid concentration regulation is a complex process, involving many different cellular pathways. We have recently demonstrated that common variation of complex traits is caused by networks of genes as opposed to single genes [4]. To assess whether GWAS results associate with entire networks of genes, we overlapped blood lipid concentration results from the Framingham heart study [26] and the Broad Institute lipid study [27] with the human, mouse and rat conserved liver network. In this analysis, we consider a gene as associated with the blood lipid trait if any SNP associated with the trait in these studies lies within 50Kb of the gene. Then, at a p-value threshold of 0.001, 22.2% of the genes with human, mouse and rat orthologs are associated with blood lipid concentration in either study. At the same p-value cutoff, 19.7% of all human genes in our dataset were associated with blood lipid concentration, suggesting that the lipid concentration regulation mechanism is conserved globally (∼1.13 fold enrichment, Fisher's Exact Test (FET) p-value = 5.38×10−11, permutation adjusted p-value<0.001, Figure S5A). The distribution of genes associated with blood lipid concentration among the modules is shown in Figure 4A. Seven of the 13 modules were observed to have a higher concentration of genes associated with blood lipids than the background. Modules 1, 7 and 11 were significantly enriched for genes associated with blood lipid levels (1.14, 1.41 and 1.55 fold enrichment with FET p-values of 1.7×10−3, 6.6×10−3, and 7.4×10−3, respectively). These results suggest that cell-cell signaling, cell-adhesion and sterol biosynthesis pathways are associated with variation in blood lipid concentration regulation in the human population. In contrast, a similar test was applied to modules identified from human expression profile data alone. The module with the highest overlap with genes associated with blood lipid traits was not enriched for a coherent biological process and the module enriched for carboxylic acid metabolism were not significantly enriched for genes associated with blood lipid traits (Figure 4B). We have further showed that these results are not sensitive to the window size around the lipid-associating loci for selecting lipid-associating genes. The trends of the global conservation of lipid-associating genes and results in Figure 4 hold true also for window size of 10K, 20K, 30K and 40K (Table S12 and Figure S6).
Genetic loci associating with blood lipid traits from both Framingham and Broad studies may harbor many genes in each of these regions. Dissecting the true causal genes from those irrelevant ones remains a significant challenge. We have previously shown that cis eSNPs – SNPs that are associated with the mRNA levels of genes residing in the same genomic regions – are enriched for functionally relevant genes associating with the trait of interest [28]. In addition to the cis eSNPs, functionally coherent gene modules, representing the cellular processes associated with the trait of interest, can also help pinpoint the true causal genes. By filtering the Framingham and Broad candidate lipid-associating genes with genes that either 1) harbor a cis eSNP in its vicinity, or 2) belongs to any of the three conserved co-expression modules enriched in lipid-associating genes, the overlap between the two studies becomes more significant than the un-filtered sets, demonstrating the utilities of cis eSNP and conserved co-expression modules in teasing out irrelevant candidate genes (shown in Table 3; in this case, the cis eSNP genes we previously identified from a liver expression study were used [28]). There were 395 genes (Table S13) that are associated with a cis eSNP in the human liver, and are also in the three conserved co-expression modules we identified as associated with the blood lipid trait. These genes represent the most likely causal genes controlling the blood lipid concentration by integrating GWAS candidate loci, human cis eSNP genes and conserved co-expression modules between human and rodent species. Among these genes, four of them, SORT1, FADS1, FADS2 and GALNT2, are recently reported as candidate genes at highly replicated genetic loci contributing to polygenic dyslipidemia [25]. This result is statistically significant given there are only 26 such candidate genes in our initial set of 6455 orthologous genes between human and rodents (a 2.51-fold enrichment, FET p-value<0.0189, permutation adjusted p-value<0.015, Figure S5B). These results demonstrate that the combination of multiple types of information can provide an objective way to infer causal genes under the loci of interest.
Many factors contribute to the identification of differential interactions between human, mouse and rat, such as evolution differences, genetic background differences, and perturbation differences in the data sets (such as genetic diversity in human liver data vs. diverse compound treatments in rat liver data), to name just a few. As a proof of concept, we applied our meta-analysis approach to identify differential interactions between the liver co-expression networks from two previously reported F2 intercrosses. The first F2 intercross was constructed between C57BL/6J ApoE null (B6.ApoE−/−) mice and C3H/HeJ ApoE null (C3H.ApoE−/−) mice (referred as BXH/apoe−/−) [29]. The second F2 intercross was constructed between C57BL/6J (B6) wild type mice and C3H/HeJ (C3H) wild type mice (referred as BXH/wt) [30]. These two crosses are essentially identical from the standpoint of genetic background, diet, and rearing, except that in one of the crosses the ApoE gene is knocked out. Given this single gene difference between the crosses, we hypothesized that differentially connected genes would be enriched for genes associated with ApoE related pathways.
Our method identified 500 differentially connected genes involving 1,023 differential interactions between the BXH/wt and BXH/apoe−/− crosses. GO enrichment analysis for this set of genes revealed that the only over represented biological process were those involving ApoE [31], albeit these processes are highly overlapping, including the cholesterol metabolic process (4.5% vs. 0.7% background, p<5.6×10−6), the sterol metabolic process (4.5% vs. 0.9% background, p<1.2×10−4) and the lipid metabolic process (15.2% vs. 7.2% background, p<3.3×10−4). Interestingly, no core biological processes in liver that do not involve ApoE (e.g., immune response) were enriched, which serves as a negative control for our results. To test whether these differential interactions were mainly driven by expression dynamic changes as the result of the ApoE gene knockout, we selected a set of 500 genes with the largest difference in expression variation between the two crosses. GO enrichment analysis revealed no coherent biological functions represented in this set, indicating that the observed network changes could not be explained simply by dynamic differences in gene expression.
We further examined the mouse protein-protein and protein-DNA interaction networks curated from interaction databases and literature, including Ingenuity, GeneGO and HPRD, around the ApoE gene. Of the 22 genes in the immediate neighborhood of ApoE, including ApoE itself, 4 (18.2%) were inferred as differentially connected between the wild type and ApoE−/− crosses, and this proportion was highly significant (∼8.1 fold enrichment, FET p-value<1.1×10−4, permutation adjusted p-value<0.001) (Figure 5 and Figure S5C). Taken together, these results demonstrate the ability of our meta-analysis procedure to dissect differentially regulated pathways around specific molecular perturbations. Although our method is purely expression profile based, it can also recapitulate known physical interactions in the region of the source perturbation, which further supports the validity of our approach.
Differential interactions among diverse organisms can result from true evolutionary differences or from incomplete perturbations in the datasets we examined, leading to reduced expression dynamics in one or both of the interacting genes. Here we assumed that the gene expression system in each species we examined was extensively perturbed, either directly or indirectly (via second or higher order effects). The human samples were collected from more than 400 unrelated individuals, making up an out-bred population comprised of 400 diverse genetic backgrounds. The F2 mice obtained from the BXH crosses represent an in-bred population in which differences in the genetic background of the parental strains are randomly shuffled in each of the individual mice. The rat expression profiles were generated by treating rats with a compendium of drug compounds with various mechanisms of action. Therefore, although liver gene expression in each species is measured under different sets of perturbations, the extensiveness of these diverse perturbations was likely to render that most pathways were perturbed given there are a finite number of pathways.
We carried out the cross-species meta-analysis in a pair-wise fashion to produce human vs. mouse and human vs. rat comparisons. For the human vs. mouse comparison our method identified 8,706 conserved interactions involving 3,205 genes, in addition to 613 differential interactions involving 547 genes. For the human vs. rat comparison, we identified 10,809 conserved interactions among 3,310 genes, as well as 447 differential interactions among 420 genes. All results were obtained using a p-value cutoff of .
We further characterized each orthologous gene considered in the comparisons by classifying each gene's involvement in 1) only conserved interactions, 2) at least one differential interaction. Since it has been shown that genes differentially connected in the co-expression and physical interaction networks tend to evolve at different rate [32],[33], we also attempted to characterize the evolutionary rate for each group by measuring the ratio between the rate of non-synonymous to synonymous substitution (Ka/Ks) [34] in the protein coding regions of the respective genes. Interestingly, for both comparisons we found that genes involved in a larger number of differential connections tend to have a higher Ka/Ks ratio (Figure S7). These results suggest that stronger positive selection (or relative weaker negative selection) may lead to new advantages for a given gene by increasing or decreasing the number of its co-expression partners. To further illustrate this point, we expanded our analysis to include genes that are non-orthologous between human and rodents, and tested whether genes that were differentially connected among orthologous genes also tended to have more interactions with non-orthologous genes in a given species, compared to genes involved in only conserved interactions. This was indeed the case when we looked at the ratio of interactions to human-specific genes vs. human-rodent orthologs in the liver co-expression network built from human expression profiles (Figure S8). Taken together, these results demonstrate that positive selection may render a gene the ability to rewire its co-expression connections with evolutionarily conserved partners as well as to add new partners that emerge through speciation.
One important aspect of understanding the difference in gene expression regulation between human and rodent species is that rodent species (mouse in particular) are frequently used to elucidate the complexity of human diseases. However, there is no guarantee that discoveries made in mouse regarding causes of disease will translate into human systems, so such results can be misleading [35]. In addition to mice being used as a model for human diseases, rats have been established as a critically important model for human drug metabolism and toxicity trials. However, the extent to which toxicity results in rat are faithfully reproduced in humans has not been well characterized [36]. Among the many species-specific variations between human and rodents that may cause such barriers, differential rewiring of the co-expression networks can be an important contributing factor. Understanding species-specific interactions, especially human-specific interactions, is a necessary step to develop relevant animal models for human diseases.
Again using the same p-value threshold described above, 1,171 differential interactions were identified among the human, mouse and rat liver co-expression networks. An interaction between two genes is considered human-specific if 1) the co-expression relationship between the two genes is significantly different between human and the rodent species based on the heterogeneity test, 2) the correlation p-value of the two genes in human is smaller than , and 3) the correlation p-values for the two genes in both mouse and rat are larger than . Of the 1,171 differential interactions identified, 163 were human-specific. The top 20 genes with most human-specific interactions are listed in Table S14. These genes are inter-connected to form three sub-networks (Figure 6). The largest sub-network consists of 11 genes, three of which (PIP5K1B, RXRG and ACSBG1) are well known to be involved in lipid metabolism. RXRG (retinoid X receptor gamma) emerges as a key regulator of this human-specific sub-network. It is one of the genes with the most predicted human-specific interactions, and 7 out of 8 of its interactions involve other genes also with the most human-specific interactions (PIP5K1B, TFAP2E, SLC22A13, DAPK3, RPS27, FAT2 and ACSBG1). RXRG homozygous mutant mice are normal [37], suggesting that it may not exert any essential function in mouse. However, there are many evidences suggesting that RXRG variations in humans are associated with lipid metabolism [38], as well as with glucose and Type 2 diabetes [39]. RXRG mutations are the most frequent variations in familial combined hyperlipidemia and are associated with triglycerides and HDL cholesterol [40]. These differences in RXRG's role between human and mouse are consistent with our prediction that there are differences between human and rodents networks around RXRG. In addition to RXRG's 8 predicted human-specific interactions with genes having a rodent ortholog, it is also known to be an upstream regulator of CETP [41] which has no corresponding ortholog in either mouse or rat. CETP encodes a cholesteryl ester transfer protein that plays a key role in regulating HDL cholesterol. Thus it may partially explain RXRG's contribution to lipid metabolism in humans. These results suggest that attention should be paid to retinoid X receptor activities when CETP transgenic rodent models are studied.
There are a number of systematic efforts for studying complex human diseases using human samples or animal models. Co-expression networks represent a powerful system-level tool for dissecting the architecture of gene expression, and the complex relationships between genes and disease associated traits. Combining co-expression networks across multiple datasets, especially those measured in common tissues from evolutionarily distant species, has the potential to greatly enhance the power to distinguish true associations among gene expression traits from those spurious interactions picked up by guilt-by-association techniques in single datasets. We presented a novel semi-parametric meta-analysis method to combine multiple high dimensional datasets from different species. When applied to the human, mouse, and rat liver co-expression networks, our method out-performed all existing methods with respect to the degree of biological coherence reflected by the identified gene pairs. Using the co-expression network conserved across human, mouse and rat, we identified cell-cell signaling, cell-adhesion and sterol biosynthesis processes as the primary mechanisms represented by GWAS gene candidates associated with blood lipid levels.
In comparing human and rodent co-expression networks we found that ∼10% of the gene-gene co-expression relationships were conserved, in accordance with a recently published comparative analysis of human and mouse gene expression patterns [42]. The conserved interactions could be organized into gene modules that corresponded to core pathways that are critical to normal cellular functions, and therefore are likely to lead to disease if disrupted. Knowledge of the conserved interactions between human and rodent species has the potential to facilitate studies of human disease using rodent models. When we combined conserved liver modules with cis-eSNP information and GWAS results, we identified a list of 395 candidate genes regulating blood lipid levels. Six of these genes (MTHFR, PEX5L, CPE, LIPA, UCP3 and PLIN) have previously been shown to have mutant phenotypes in mouse that involve abnormal lipid levels. Systematic testing of the genes in this set using experimental techniques such as siRNA in cell-based systems could provide further confirmation of their involvement in regulating blood lipid concentrations.
Under a unified framework, our method also allows the identification of gene-gene relationships that differ significantly between datasets. The sensitivity of our method to identify dataset-specific biological perturbations was well highlighted by the identification of a single sub-network driven by ApoE that was able to distinguish two nearly identical experimental cross populations whose genetic backgrounds were identical with the exception of ApoE (knocked out in one of the crosses). This type of network comparisons can help characterize network plasticity due to evolution. We have shown that genes involved in such differential interactions between human and rodents are likely to be under positive selection for gaining or losing co-expression partners. Given that only ∼10% of gene-gene relationships are conserved between these diverse species, divergence in gene expression are likely to be more extensive than genome sequences. It has shown through a chip-chip study that the overlap of transcription factor binding sites is only about 20% across 3 different yeast species where sequence differences are about 0.05% [43]. In some cases, the promoter regions are identical across genomes of 3 yeast species, transcription factors only bound in one species but not others. Thus, variation in transcription regulation is much larger than sequence variation. There could be other factors affecting conversation of pairwise relationship in different data sets, such as 1) inadequate expression dynamics in those parts of the system that lack targeted perturbations, and 2) experimental and technological noise that subdue the real changes in co-expression.
In addition to the meta-analysis methods we compared, there are graphic model-based meta-analysis or Bayesian meta-analysis methods which have been applied to gene expression data in several studies [44],[45]. The performance of Bayesian meta-analysis depends on priors tuning. If noninformative priors are used, then the Bayesian meta-analysis is close to the random effect model. Even through effect sizes are clearly different in our simulated data and empirical data, the mixed effect model performed worse than the fixed effect model. On the other hand, our meta-analysis method is robust across multiple conditions without any tuning of parameters. In addition, the Bayesian meta-analysis is away more computation intensive than the method we proposed so that we did not include it in our comparison.
Meta analysis of co-expression networks we proposed here allow us to compare co-expression networks constructed from data sets of heterogeneous experimental settings. If experimental settings are similar, then direct comparison of signature sets can also provide insights of conserved mechanisms at system levels. For example, a set of periodically expressed genes in H. sapiens, S. ceravisiae, S. pombe and A. thaliana was defined and then orthologs of these genes were compared to see whether they peaked during the same phase of cell cycle [46]. However, in our datasets, experimental conditions were different - the variances of human and mouse liver expression data were due to naturally occurred genetic variation, whereas those in the rat liver expression data were due to diverse compound treatment. Therefore, there is no common way to define gene signatures across different data sets that can be compared directly.
Gene expression is one type of high throughput data that can be leveraged to systematically study human diseases. There are many other types of high-dimensional data to which our method could be applied, including protein-protein interaction, protein expression, metabolite expression, and Chip-on-chip data. Further developments are needed to combine these different types of data across different species. Nevertheless, even at its current stage, our method has been successful in identifying mechanisms that are common between and distinct to human and rodent species, which provides the potential to aid in the drug development process.
We profiled 423 human liver samples [28], 382 liver samples of rats treated with different classes of drugs [47], 300 mouse liver samples from an F2 murine intercross between C57BL/6J ApoE null (B6.ApoE−/−) and C3H/HeJ ApoE null (C3H.Apo E−/−) (referred as BXH/apoe−/−) [29], and 321 mouse liver samples from an F2 intercross between C57BL/6J (B6) wild type mice and C3H/HeJ (C3H) wild type mice (referred as BXH/wt) [30]. For every gene in each expression dataset, the expression values were mean-subtracted and then divided by the standard deviation. Missing values were imputed by the robust regression based the expression of the gene most correlated to the query gene expression.
Orthologous gene pairs between human, mouse and rat represented on microarrays were identified by taking the reciprocal best hit using BLASTN with an E-value cutoff of . This resulted in 8,767 orthologous pairs identified between human and mouse, 6,934 between human and rat, and 10,185 between mouse and rat. There were 6,455 orthologous genes common to all three species, which were selected for subsequent analysis (Table S2).
To estimate the significance of a correlation coefficient , we generally convert to which follows a student t-distribution with . When the sample size is large enough, is approximately normally distributed [11]. However, convergence of the distribution is very slow and it is said to be unwise to assume its normality for n<500 [48].
The assumption for estimating the Pearson correlation coefficient distribution is that all vector pairs are independently and identically distributed. However, this may not hold true in practice such as microarray experiments due to the facts that 1) probes for two genes on the same chip may be correlated because that are subjected to many common noises and biases, and 2) two unrelated genes in a biological network are still remotely connected so that they can not be completely independent. As a result, not all gene pairs are independent, thus their expected correlation coefficient is not necessary zero. In this case, an empirical null distribution is needed. We note that empirical null distributions are different for each gene/probe so that there are 6455 null distributions instead of one global null distribution. We assume the empirical null distribution of all pair-wise correlation coefficients as a normal distribution based on the central limit theorem, which states that the mean of sufficiently large number of independent random variables will be approximately normally distributed [49].
In summary, we assume the empirical null distribution of pair-wise correlation coefficients as a normal distribution under two conditions: (1) the sample size is large so that the variation of is small; (2) the number of genes under study is large so that the central limit theorem can be applied.
We note that our sample sizes are in the range of 300–500, which are out of the recommended range for normal assumption. However, our normal assumption for correlation distributions of our data is supported by the Kolmogorov-Smirnov test of normality. The sample sizes of the data sets we simulated are 150. We checked the distributions of correlation coefficients of each individual gene, and found that correlation coefficients for over 98% of genes are normally distributed. For the empirical data sets, correlation coefficients for over 70% of genes are normally distributed, which is shown in Figure S1.
We assume the underlie system consists of 2000 genes which are divided into10 functional modules and 1 null module, as shown in Figure S2. Each functional module consists of 100 genes and the null module consists of 1000. For simplicity, we assume each gene in a functional module is linearly related to a latent regulator and is simulated as , where is a vector (, the sample size) representing the expression of gene (which belongs to the functional module ) in data set . is a vector for the latent variable in data set . is a vector representing systematic noise in the data set . is the random noise. Genes in the null model are not related and are simulated as , where is a random vector. and are regression coefficients representing the strengths of the signal and the systematic noise, respectively. The latent variables , the random signals , systematic noise and random noise are all assumed to be normally distributed with mean and different variances. The coefficients in are constrained by the strength of correlation . The sign of the coefficient was randomly assigned. It is similar for . We assume are jointly normally distributed, we can write their covariance as , where is of size , is of size , is of size , and is of size . The regression coefficients, and are then given by , and the error term, , is normally distributed with mean 0 and variance .
To assess the goodness of the reconstructed coexpression networks derived from different meta-analysis methods, they were compared to the true network, which was formed by linking all genes in the same functional module as defined in the simulation process. We define the “goodness” of the reconstructed network in terms of its accuracy, which is measured by two parameters. The first parameter is defined as the precision of the network: , which is the proportion of detected interactions that actually exist in the true network. Precision corresponds to specificity and is equal to one minus the false positive rate (). The second parameter is defined as the recall of the network: , which is the proportion of total interactions in the true network that are detected in the reconstructed network. Recall corresponds to sensitivity and is equal to one minus the false negative rate (), which is also known as the true positive rate (). The recall and precision for a perfectly reconstructed network are equal to 1.
The central figure of merit used to evaluate and compare the coexpression networks derived from different meta-analysis methods (with respect to the true network) is the recall vs. precision curve, which can be considered as a variation of the traditional Receiver Operator Characteristic (ROC) curve. ROC curves are generated by plotting the true positive rate (TPR) against the false positive rate (FPR). The area under the ROC curve (AUC) is then a measure of how the constructed network compares to the true network. The larger the AUC, the better the constructed network compares to the true network, where the maximum AUC is 1, indicating that the constructed network perfectly matches the true network. Qualitatively the recall vs. precision curve is equivalent to the ROC curve in that if the AUC for one network is greater than (or less than) the AUC of a second network with respect to one of plot types, that same relationship will hold for the other plot type. We opted to use the recall vs. precision plots over the ROC plots as the figure of merit because recall and precision are the more standard measures used in the network reconstruction community.
The false discovery rate (FDR) of our meta-analysis results was estimated using permutation test procedures. For the conserved interactions, null datasets were created by randomly and independently shuffling the expression values of all genes in each dataset, thus breaking the inter-gene relationships while keeping intact the expression mean and standard deviation of the genes in every dataset. For the differential interaction, we generated the null datasets by shuffling the dataset membership of the samples, so that the permuted datasets are essentially random subsets of the total original samples. The same meta-analysis procedure was applied to both the original datasets as well as the permuted ones. The FDR was then computed as the ratio between the number of inferences made from the permuted datasets (i.e. false discoveries) over the number of inferences made from the original datasets (i.e. total predictions).
Only GO biological process categories with fewer than 1,500 genes (according to human annotations) were included for analysis, precluding non-specific categories, such as metabolic process, from entering the analysis. All GO enrichment analyses were performed using the Fisher's exact test, with all 6,455 orthologous genes forming the background gene set for the human, mouse and rat comparisons. For the BXH/wt vs. BXH/apoe−/− analysis, the background set was comprised of all genes represented on the microarray used in the study.
To partition the network of genes obtained from our procedures into modules of genes, we employed the divide-and-merge methodology of clustering [27], where a top-down divide phase based on a theoretical spectral algorithm [50] was used to obtain a clustering tree, and a bottom-up merge phase was used to parse the clustering tree to obtain a partition of the genes (gene modules) that optimized a certain objective function. We used the modularity function [51] to identify modules in the human-mouse-rat conserved network. The definition of modularity from the cited references is provided here for completeness. Let be a partition of the genes in a network into clusters . Then,where is the number of edges between two genes that both belong to , is the sum of the number of neighbors of all genes in , and m is the number of edges in the whole network.
|
10.1371/journal.ppat.1001321 | Chemokine Binding Protein M3 of Murine Gammaherpesvirus 68 Modulates the Host Response to Infection in a Natural Host | Murine γ-herpesvirus 68 (MHV-68) infection of Mus musculus-derived strains of mice is an attractive model of γ-herpesvirus infection. Surprisingly, however, ablation of expression of MHV-68 M3, a secreted protein with broad chemokine-binding properties in vitro, has no discernable effect during experimental infection via the respiratory tract. Here we demonstrate that M3 indeed contributes significantly to MHV-68 infection, but only in the context of a natural host, the wood mouse (Apodemus sylvaticus). Specifically, M3 was essential for two features unique to the wood mouse: virus-dependent inducible bronchus-associated lymphoid tissue (iBALT) in the lung and highly organized secondary follicles in the spleen, both predominant sites of latency in these organs. Consequently, lack of M3 resulted in substantially reduced latency in the spleen and lung. In the absence of M3, splenic germinal centers appeared as previously described for MHV-68-infected laboratory strains of mice, further evidence that M3 is not fully functional in the established model host. Finally, analyses of M3's influence on chemokine and cytokine levels within the lungs of infected wood mice were consistent with the known chemokine-binding profile of M3, and revealed additional influences that provide further insight into its role in MHV-68 biology.
| Infection of inbred strains of laboratory mice (Mus musculus) with the rodent γ-herpesvirus MHV-68 continues to be developed as an attractive experimental model of γ-herpesvirus infection. In this regard, the MHV-68 protein M3 has been shown to selectively bind and inhibit chemokines involved in the antiviral immune response, a property expected to contribute significantly to virus infection and host colonization. However, inactivation of the M3 gene has no discernable consequence on infection in this animal host. Prompted by recent evidence that natural hosts of MHV-68 are members of the genus Apodemus, and that MHV-68 infection in laboratory-bred wood mice (Apodemus sylvaticus) differs significantly from that which has been described in standard strains of laboratory mice, we addressed whether M3 functions in a host-specific manner. Indeed, we find that M3 is responsible for host-specific differences observed for MHV-68 infection, that its influence on infection within wood mice is consistent with its chemokine-binding properties, and that in its absence, persistent latent infection - a hallmark of herpesvirus infections - is attenuated. This highlights the importance of host selection when investigating specific roles of pathogenesis-related viral genes, and advances our understanding of this model and its potential application to human γ-herpesvirus infections.
| The human γ-herpesviruses - Epstein-Barr virus (EBV) and Kaposi's sarcoma-associated herpesvirus (KSHV; alternatively human herpesvirus 8 [HHV-8]) - possess significant oncogenic potential, particularly in the setting of immune deficiency. Both establish lifelong latent infections, primarily within B lymphocytes, through the actions of a limited repertoire of their approximately 90 genes. While the majority of these have a role in virus production, it is principally the actions of the latency-associated genes of these viruses that contribute to their oncogenic potential. Strict host preferences of EBV and KSHV, unfortunately, severely limit assessment of the mechanisms that contribute to their persistence and pathogenesis. Consequently, there has been considerable effort to develop experimental infection of laboratory mice (Mus musculus) with the murine γ-herpesvirus 68 (MHV-68 or γHV68; officially murid herpesvirus 4 [MuHV-4]) as a model of γ-herpesvirus infection [1], [2], [3], [4], [5], [6], [7].
As a member of the γ2 subfamily of herpesviruses, MHV-68 is closer genetically to KSHV/HHV-8 than to EBV, a γ1 herpesvirus [8], [9]. Regardless, each γ-herpesvirus contains a unique set of genes that contributes to its distinct biology and pathogenic properties. For MHV-68, this is primarily a cluster of latent- and lytic-infection-associated genes at the extreme left end of the viral genome that encodes for four novel proteins, M1–M4, and interspersed among these are eight RNA polymerase III-transcribed genes that encode abundant viral tRNA-like (vtRNA) transcripts [8], [10]. Much of the effort to define the biology of MHV-68 infection and its applicability as a model of human γ-herpesvirus infections, has therefore focused on the roles of these genes in the context of infection within inbred strains of laboratory mice. Of the proteins encoded by this locus, the biochemical function of M3 is the best understood.
A secreted 44-kDa protein, M3 is highly expressed during lytic infection, and probably to a lesser extent during latency [11], [12], [13], [14]. In vitro, M3 selectively binds chemokines associated with the antiviral inflammatory response [15], [16]. Surprisingly, inactivation of M3 expression (by insertion of a translational stop codon) has no apparent consequence on MHV-68 infection following intranasal inoculation of C57BL/6 mice [17]. By contrast, intracerebral injection of the same M3 mutant virus does lead to an altered inflammatory response, with higher numbers of infiltrating lymphocytes and macrophages than observed following inoculation with wild-type virus [17]. Thus, M3 does appear capable of functioning as a chemokine-binding protein in vivo, though it is perplexing why ablation of M3 expression has no apparent impact on pathogenesis or on virus replication and the establishment of latency following intranasal inoculation, clearly more representative of a natural route of infection. One possible explanation for this may relate to the experimental host.
MHV-68 was originally isolated from a bank vole (Myodes glareolus) [18] although this appears to be only an occasional host [19]. In spite of what has been suggested recently [20], we have shown conclusively using sequence analysis that the natural hosts of MHV-68, at least in continental Europe, are in fact members of the genus Apodemus [21]. Specifically, Apodemus flavicollis, Apodemus agrarius, and Apodemus sylvaticus (wood mice) [21]. Significantly, our recent comparative analysis of experimental MHV-68 infection of BALB/c (M. musculus) and laboratory-bred wood mice revealed markedly different findings [22]. In wood mice, virus replication in the lung was substantially muted, and latency within the spleen was established without the dramatic leukocytosis and splenomegaly that are the hallmark pathogenic properties of MHV-68 latency within inbred laboratory strains of mice. In addition, the associated histological changes were significantly different. Notably, in wood mice, viral replication was restricted to scattered alveolar epithelial cells and macrophages within focal granulomatous infiltrations. Latently-infected lymphocytes were also abundant in focal perivascular/peribronchiolar infiltrations and in inducible bronchus-associated lymphoid tissue (iBALT). In addition, while well-delineated secondary follicles with classical germinal center formation were seen in the wood mouse spleens, only poorly-delineated follicles without distinct germinal centers were seen in BALB/c mice.
Given the unlikelihood of an insignificant role for M3, we asked whether M3 might contribute to the vastly different response of wood mice to MHV-68 infection. Here we demonstrate that upon intranasal inoculation of wood mice, M3 does indeed modulate the host inflammatory response in a manner consistent with its chemokine-binding properties, and that it is responsible for the MHV-68-dependent iBALT observed in this species. Additionally, we show that M3 is critical for the organization of splenic follicles, and that in the absence of M3, latent MHV-68 infection is significantly attenuated in both lung and spleen. These results highlight the importance of utilizing a natural host in this small-animal model of γ-herpesvirus infection, and provide substantial new insight into the biology of MHV-68 that should contribute to future use of this model and its applicability to understanding human γ-herpesvirus infections and pathogenesis.
Following intranasal inoculation of mice, e.g., BALB/c and C57BL/6, a burst of MHV-68 replication occurs within lung epithelial cells [23] prior to the establishment of latent infection within lung epithelium [24] and ultimately the hematopoetic system [25], [26]. This replication peaks at approximately 7 days p.i. and is largely resolved by day 10 p.i.. The titer of virus produced in the lungs of wood mice, however, is substantially lower (by ∼3 log10 plaque forming units), though the long-term viral DNA loads established within the lung of wood and BALB/c mice are equivalent [22]. In contrast to BALB/c mice, virus productive replication in the lungs of wood mice appears confined within granulomatous infiltrates, and in separate lesions, numerous lymphocytes within perivascular and peribronchial accumulations harbor latent virus (primarily within B cells). We reasoned, therefore, that the immune-modulatory function of M3 might be particularly critical during acute infection within the lungs of wood mice. To address this, we first examined M3 expression in lung by quantitative reverse transcription PCR (qRT-PCR). To put our results in a more meaningful perspective, we determined M3 mRNA levels relative to those for the other 3 genes in this locus (M1, M2 and M4), as well as to the early lytic-cycle gene ORF50, which served as a general indicator of lytic infection. This experiment was performed twice with comparable results.
As illustrated in Fig. 1A, M3 expression was detected at all four time points evaluated (7, 10, 12, and 14 days p.i.). There were, however, two unexpected findings with respect to M3. First, whereas M3 encodes one of the most highly expressed MHV-68 mRNAs during lytic infection in vitro, especially relative to the other lytic-cycle mRNAs encoded by this locus (M1 and M4) [27], this was clearly not the case in vivo here. Second, there was an obvious spike in M3 expression between 12 and 14 days p.i. (Fig. 1A), a time when virus replication is believed to have subsided. Clearly, the level of M3 expression at 14 days p.i. was significantly higher than that for M1, M2 and M4 (P<0.01). Although this may be indicative of the onset of latency-associated M3 expression, detection of a parallel spike in ORF50 expression suggests that this is lytic cycle-associated expression. However, we cannot exclude the possibility that at this point in infection, M3 transcripts originate from latently infected cells, whereas ORF50 expression is occurring in separate cells still supporting full or an abortive virus replication, probably within granulomatous infiltrates that support productive infection in wood mice lungs [22].
There were several additional observations of note with respect to this gene locus. Whereas M1 mRNAs are relatively low in abundance during MHV-68 replication in vitro and in BALB/c mice spleens [28], M1 expression was significantly higher (P<0.05) at 7 days p.i. than the other genes within the locus, and still one of the most highly expressed genes tested at 10 days p.i. (Fig. 1A). By contrast, M4 transcript levels were nominal, suggesting that M4 either performs a function at times or anatomical sites other than those analyzed here, or that a comparatively lower level of M4 transcript is required for M4 expression. Finally, it was somewhat surprising that expression of M2, believed to be a strict latency-associated gene [29], was readily detectable early in the infection, peaked at 10 days p.i. and decreased through day 14 p.i. where numerous latently-infected B cells are known to be present [22]. MHV-68 latency has been detected as early as 3 days p.i. [30] and B cell infiltrations containing MHV-68 are present as early as day 7 p.i. in the wood mouse (Fig. 2), [22]. Further, it has been shown that the pattern of MHV-68 latent gene expression is differentially regulated in B cells depending on cellular differentiation state [31] and thus the observed pattern of M2 expression is likely a reflection of this.
To assess the relative level and timing of M3 expression between wood and BALB/c mice we determined M3 and ORF50 mRNA levels in the lungs of infected wood and BALB/c mice at 7 and 14 days p.i. by qRT-PCR as above. The results (Fig. 1B) showed that M3 expression in BALB/c mice was similar to wood mice at day 7, but significantly lower (P<0.01) at day 14, showing that the timing of M3 expression differs between the two species of host.
We next sought to localize the site of M3 transcription within lung and spleen by RNA in situ hybridization. Within lung, at 7 days p.i. M3-positive lymphocytes were detected in B cell-dominated perivascular/peribronchial infiltrates and, together with positive macrophages, within granulomatous infiltrates and occasionally within blood vessels, rolling along/attached to vascular endothelial cells (Fig. 2A). This M3 RNA expression pattern was seen also on days 10, 12 and 14. On days 12 and 14 p.i., there were many M3 positive lymphocytes in the progressively prominent perivascular/peribronchial lymphocyte accumulations (Fig. 2C), and some were also seen disseminated in the parenchyma. They were also present in the follicle-like B-cell accumulations that were first seen on day 12 (Fig. 2D) and which had developed germinal centers by day 14 (data not shown). Previous work has shown that these follicle-like infiltrations with germinal centers are inducible bronchus-associated lymphoid tissue (iBALT). Consistent with a trafficking of latently infected cells from the site of primary replication within the respiratory tract to spleen, we observed variable numbers of M3-positive lymphocytes mainly in follicles within bronchial and mandibular lymph nodes on day 7 p.i. (Fig. 2E). Within spleen, in which latently infected cells peak approximately 12–16 days p.i., M3 expression was prominent within follicle centers from 10 days p.i. onward (Fig. 2F). No hybridization was detected in any tissue with sense-strand probes (e.g., Fig. 2B).
Thus, a high level of M3 expression was observed at d14 p.i. in lymphocytes within iBALT and splenic follicles.
We next asked if the development of B cell-dominated, M3 RNA-positive perivascular/peribronchial infiltrates (Fig. 2A) and iBALT, which are not features of the lungs of BALB/c mice infected with MHV-68, was a result of M3 expression. To address this question, we infected cohorts of wood mice with a previously characterized recombinant MHV-68 that has a targeted disruption of the M3 gene (M3.stop, gift of S.H. Speck and H.W. Virgin) [17]. The M3 gene in this virus contains three translational stop codons inserted into the 5′ end of the M3 ORF. The marker-rescue version of this virus, M3.MR, containing a fully restored M3 gene [17], was used as wild-type virus for comparison. The histopathological analyses are shown in Fig. 3 and the quantification of these in Fig. 4). As expected, the histological changes in the lung tissue from wood mice infected with M3.MR on day 7 and 12 p.i. were similar to those observed with MHV-68 infection as follows [22]. There was a marked increase in the amount of interstitial lymphocytes based on the significant (P<0.01), ca 3 fold increase in T cells as compared with uninfected animals (Fig. 4A, B). There was also a moderate perivascular and peribronchial infiltration which contained a higher proportion of B cells (B∶T ratio of 1.77∶1, P<0.001; Fig. 4C, D) with B cell rolling and emigration (data not shown). Multifocal granulomatous infiltrates containing viral antigen were also observed. By 14 days p.i., two types of lymphocyte-dominated perivascular and peribronchial infiltrations had developed multifocally in association with larger arteries and bronchi. One type contained T and B cells in approximately equal proportions (Fig. 4C, D), and the other was B-cell dominated and follicle-like with germinal center formation, i.e., iBALT (Fig. 3A, C; 4E). B lymphocytes made up ca 75% of the cells in iBALT (P<0.001), whereas T cells were present in much smaller numbers (Fig. 3C,E; 4E). While evidence of iBALT formation was already seen on day 12 p.i., the perivascular/peribronchial infiltrates that contained approximately equal proportions of T and B cells were only seen after 14 days p.i. and the latter perhaps represent the physiologic immune response as the acute phase of MHV-68 infection in lung dwindles. Lymphocytes that expressed vtRNA, indicative of a latent infection, were found within iBALT, and very occasionally intravascularly (Fig. 3G).
By contrast, infection with M3.stop virus led to markedly different histological findings. At 7 days p.i., a statistically-significant (P<0.01) increase in interstitial lymphocytes that consisted of predominantly T cells was observed (Fig. 4A, B). Multifocal granulomatous infiltrates containing viral antigen were also observed. Mild perivascular/peribronchial lymphocyte accumulations were obvious, and immunohistological staining showed that this was B-cell dominated (B∶T ratio of 1.5∶1, P<0.001; Fig. 4C, D). T cells were also seen rolling along and emigrating from blood vessels, an observation not seen when M3 was expressed. After 14 days, the perivascular and peribronchial lymphocyte infiltrations were still evident (Fig. 3B). However, these were far less intense than in the lungs of the M3.MR-infected wood mice (compare to Fig. 3A) and consisted of both B cells (Fig. 3D) and T cells (Fig. 3F) in a ratio of 1.3∶1 (Fig. 4C, D). T cells (but not B cells) were also found rolling along arterial walls and emigrating from vessels of M3.stop-infected animals (Fig. 3F, inset). vtRNA-positive lymphocytes were observed in perivascular infiltrates of M3.stop-infected wood mice, but there were fewer of these, possibly due to the lower proportion of B cells and the smaller size of the infiltrates (Fig. 3H) than seen for infection with M3.MR virus (compare to Fig. 3G). Notably, while granulomatous infiltrates were seen in both groups of mice, iBALT was absent in M3.stop-infected mice. Thus, while M3 is not essential for infection, the host response to infection is clearly altered in its absence.
A major organ of MHV-68 persistence is the spleen, in which the number of latently infected cells - primarily B cells but also dendritic cells and macrophages - peaks approximately 2 weeks p.i. [5], [32]. As M3 is expressed within spleen (Fig. 2F), we examined the effect that M3 loss has on MHV-68 infection there. As shown in Fig. 5A, at 14 days p.i. the spleens of wood mice infected with M3.MR virus contained moderately sized follicles with distinct germinal centers. By contrast, the spleens of M3.stop-infected animals displayed expanded follicles without distinct germinal centers, and a slight increase in cellularity of the red pulp (Fig. 5B). Interestingly, splenic architecture observed in mice infected with M3-stop virus was very similar to that observed in the spleens of BALB/c mice infected with MHV-68, but without the marked increase in the number of leukocytes within the red pulp [22]. Further, identification of vtRNA-positive cells by in situ hybridization indicated that the well-delineated splenic follicles of M3.MR-infected mice were heavily populated with latently infected cells, and that these cells were rare outside follicles (Fig. 5C), comparable to what we had observed upon infection with wild-type MHV-68 [22]. Although vtRNA-positive cells were detected within the poorly-defined follicles of M3.stop-infected mice, the number of these cells was notably lower, and they were occasionally present as well within the red pulp (Fig. 5D). However, consistent with the inability of MHV-68 to induce significant leukocytosis and splenomegaly in wood mice (unlike in BALB/c and C57BL/6 mice), we noted no significant change in total spleen cell numbers after infection with either M3.MR or M3.stop (data not shown). Thus, the M3 gene clearly influences MHV-68 infection within the spleen of wood mice and upon its inactivation, splenic architecture resembles that in BALB/c mice infected with wild-type MHV-68.
Given the dramatic histological differences that we observed in the lung and spleen as a consequence of disrupting M3 expression, we next asked how inactivation of M3 expression affected MHV-68 infection itself within these organs. Because MHV-68 replication within the lungs of wood mice does not yield the high titers of virus seen in BALB/c mice that can be readily quantified by plaque assay [22], we chose to indirectly measure levels of virus by qPCR. At 7 days p.i., the level of viral DNA detected within the lungs of mice infected with M3.MR was not significantly higher than that within the lungs of mice infected with M3.stop (Fig. 6A). At day 14, however, significantly reduced levels of viral DNA (P<0.05) were detected in the lungs of M3.stop-infected wood mice (Fig. 6A). A similar result was seen at day 40 p.i. when MHV-68 DNA was detected at low levels in M3.MR-infected mice, but at a significantly lower level in those infected with M3.stop (P<0.05).
To assess the effect of M3 loss in the spleen, infective center assays were performed at 14 days p.i. to measure the number of latently infected cells, which are normally at their peak level at this time. Similar to our observation in the lungs, the number of spleen cells that harbored reactivatable virus was significantly lower (P<0.005), though still detectable, in wood mice that had been infected with M3.stop (Fig. 6B). A parallel effect was seen when viral DNA was measured by qPCR (Fig. 6C), confirming that the disparity in infective centers was not due to an inability of M3.stop virus to reactivate ex vivo. Viral DNA was still detectable in both groups of wood mice at day 40 p.i., but at a significantly reduced level in the animals infected with M3.stop (Fig. 6C). Hence, in both lung and spleen, the lack of M3 significantly reduced the ability of M3.stop to establish a normal level of infection that, at least in spleen, reflected a nearly ten-fold lower number of latently infected cells.
Because M3 is not required for replication of MHV-68 in vitro [17], [33], we reasoned that deficiencies of the M3.stop virus apparent in wood mice were more likely due to a loss of the chemokine-binding properties of M3, rather than to a direct defect in virus replication per se. To determine if loss of M3 expression results in a change in the chemokine profile, we measured the relative levels of a panel of chemokines and cytokines within the lungs of mice at 14 days p.i. (the peak of M3 expression during acute infection; Fig. 1A) with either M3.MR or M3.stop virus. To accomplish this we performed cytokine antibody array analyses (RayBio Mouse Cytokine Antibody Array 3.1), a proven method of comparing cytokine/chemokine levels in tissues [34]. The results (Fig. 7) showed that in a number of cases, the amount of these molecules was notably higher (>2 fold positive fold change) in the lungs of mice infected with M3.stop relative to M3.MR virus. Specifically, we observed relative increases in RANTES/CCL5 (2.5-fold), MIP-1α/CCL3 (2.2-fold), fractalkine/CX3CL1(3-fold) KC/CXCL1 (12.9-fold), MIP-2/CXCL2 (3.1-fold) and MIG/CXCL9 (2.0-fold) in the absence of M3. By contrast, we observed relative decreases in the B-cell associated chemokines BLC/CXCL13 (2.2-fold) and SDF-1α/CXCL12 (2.3-fold), as well as CD30L (2.7-fold), in infections lacking M3 (Fig. 7).
To confirm the above array results, the concentrations of selected chemokines were measured in the lungs of infected mice at 7 and 14 days p.i. by ELISA. The results (Fig. 8) showed that, in agreement with the array results, at day 14 p.i. the concentrations of RANTES/CCL5 and fractalkine/CX3CL1 were significantly higher and the concentrations of SDF-1α/CXCL12 and CD30L/CD153 were significantly lower in M3.stop-infected mice. At day 7 p.i., the only significant difference in the concentrations of chemokines between the groups was a lower level of CD30L/CD153 in M3.stop-infected mice. Of note, also, the levels of KC/CXCL1 and BLC/CXCL13 did not vary significantly between M3.stop and M3.MR-infected mice. Thus, the difference in levels of these chemokines that was seen between the groups in the array experiment above was not substantiated.
Thus, our results are consistent with the notion that M3 functions primarily through its direct interactions with cellular chemokines, and that these interactions are critical for the efficient establishment of persistent infection in the wood mouse.
Here we have shown that the MHV-68 M3 gene, which encodes a highly expressed chemokine-binding protein [15], [16], contributes substantially to infection in the lung and spleen of wood mice (Apodemus sylvaticus), which we have conclusively shown to be a natural host of MHV-68 [21]. This work was prompted by our finding that experimental MHV-68 infection of wood mice differs in several key respects from infection of BALB/c mice (Mus musculus) [22], and an earlier demonstration that, surprisingly, inactivation of the M3 gene has little consequence in the context of comparable (intra-nasal) MHV-68 infection of inbred strains of laboratory mice [17], a now widely utilized small-animal model of γ-herpesvirus infection. Specifically, M3 contributes to the formation of iBALT, and the spike in latently infected cells within spleen that occurs at approximately 2 weeks p.i. and the level of long-term latency. While iBALT is not evident in the lungs of infected BALB/c mice [22], inactivation of M3 expression from the same mutant virus (M3.stop) did not have a comparable effect in CD1 and C57BL/6 mice on this transient rise in latently infected splenocytes [17], a feature common to infection in both strains. Thus, the contributions of M3 to MHV-68 infection are largely species specific, though an attenuation of MHV-68 infection in brain as a result of an altered inflammatory (predominantly neutrophilic) response has been observed in CD1 mice injected intracerebrally with M3.stop relative to M3.MR virus [17], suggesting that M3 is not fully inactive in Mus musculus.
In contrast to the lack of an attenuation of either lytic or latent MHV-68 infection previously observed in CD1 and C57BL/6 mice following intranasal inoculation with M3.stop virus [17], in an earlier report BALB/c mice infected via the same route with a virus in which the M3 gene had been replaced with a LacZ expression cassette exhibited a reduced viral latent load in the spleen [33]. Further, depletion of CD8+ T cells partially precluded this effect, suggesting a role for M3 in the inactivation of chemokines involved in the T-cell response [33], which peaks between 10 and 20 days p.i. [6], i.e., the point at which the consequences of ablation of M3 expression were most evident in spleen. This phenotype, however, is similar to that observed in BALB/c and C57BL/6 mice in three independent reports of infection with MHV-68 M2 mutants [35], [36], [37]. Because of this, and that the 5′ regulatory region of the M2 gene extends into the adjacent M3 ORF [29], [38], we believe it is very likely that the apparent effects of disrupting M3 expression by insertion of a CMV promoter-LacZ cassette in this earlier study may have been due instead to a combination of removing the M3 ORF and/or an unintended disruption of M2 expression or an immune response to LacZ [39]. Since we also observed a reduction in latent virus load in the spleens of infected wood mice here as a consequence of specifically targeting M3 expression (Fig. 6), it will be interesting to determine if loss of M2 contributes also to this phenomenon within this host, as it does in Mus musculus.
In addition to species-associated differences seen within lungs upon MHV-68 infection, the spleens of infected wood mice exhibit clearly defined secondary follicles with highly organized germinal centers, whereas the follicles in BALB/c mice are notably larger and poorly organized [22]. Interestingly, follicles containing infected splenocytes in wood mice that had been inoculated with M3.stop virus (Fig. 5) appeared very similar morphologically to those that we observed in the spleens of BALB/c mice infected with MHV-68[22], indicating that this additional difference between mouse species may also be due to the presence or not of M3.
It was surprising, given that M3 modulates the action of a number of macrophage-specific chemokines that granulomatous infiltrations were present in similar number and size in both M3.stop and M3.MR-infected wood mice. These are most prominent at day 7 p.i., are macrophage rich and are the focus of MHV-68 replication in the lung [22]. This is perhaps due to the location of M3 expression, which is in B cells in perivascular/peribronchiolar infiltrates and iBALT but conspicuously not in granulomatous infiltrates. This suggests that the effect of M3 on chemokines is localized predominantly to areas where M3 is expressed.
Perhaps the most significant observation is that iBALT in the lungs of acutely infected wood mice is dependent on M3. iBALT is an example of tertiary or ectopic lymphoid tissue that develops at any sub-epithelial site in response to inflammation or infection. The organization of tertiary lymphoid tissue is remarkably similar to that of secondary lymphoid tissues with separate B and T cell areas, a network of specialized dendritic cells, and the presence of high endothelial venules [reviewed in [40]]. Additionally, their organization is dependent on the same chemokines that are required in lymph nodes [41]. Although the purpose of iBALT is not completely understood, it has been proposed that it participates in generation of protective immune responses along with secondary lymphoid tissue. For instance, in the absence of secondary lymphatic tissues (using Lta−/− mice), iBALT has been shown to provide protective immunity to influenza virus infection, as it is able to generate isotype-switched B cells via germinal center reactions and specific CD8+ T cells [42]. In contrast, in MHV-68 infection iBALT does not appear to play a protective role as in its absence productive infection is not greater, and in fact latency is attenuated. Instead, we hypothesize that MHV-68, via M3 functions, utilizes iBALT as a means to augment virus persistence by promoting B cell proliferation, as numerous latently infected cells can be found in these B cell-dominated accumulations (Fig. 3) that we have shown are devoid of viral structural antigens, and thus presumably virus replication, which occurs primarily within pulmonary granulomatous infiltrates in wood mice [22]. At the present time, the phenotype of the T cells (CD3+) present in the iBALT is not known, but it is plausible that these are either CD4+ T cells that would promote the activation of B cells by providing the necessary CD40, or a subset of CD8+ T cells (IFN-γ-secreting, CD40L+, perforin negative) that are necessary for ectopic lymphoid follicle formation [43].
Generation of iBALT and highly organized germinal centers are events that rely heavily on coordinated cell migration and organization, for which chemokines are critical. Given the chemokine-binding properties of M3 that have been demonstrated in vitro [15], [16], and the altered inflammatory response to MHV-68 infection in brain as a consequence of eliminating M3 expression [17], we asked whether these events associated with MHV-68 infection in wood mice reflect an M3-dependent change in the pulmonary chemokine/cytokine profile (Figs. 7, 8). Our array-based analysis of chemokine and cytokine levels revealed that numerous T cell, monocyte/macrophage, and neutrophil associated chemokines were present in higher levels within the lungs of wood mice infected with M3.stop relative to M3.MR. For example, levels of the chemokines RANTES/CCL5, MIP-1α/CCL3, MIP-1γ/CCL9, MIG/CXCL9, MIP-2/CXCL2, MIP-3β/CCL19 and fractalkine/CX3CL1 were lower in the wood mice infected with wild-type virus. In contrast, analysis revealed that the levels of two factors, SDF-1α/CXCL12 and CD30L/CD153 were higher after infection with M3.MR in both the array and ELISA assays (Figs. 7, 8). With respect specifically to iBALT formation, stromal chemokines such as SDF-1α/CXCL12 (levels enhanced by M3) have been implicated in the cellular recruitment required for iBALT formation [44]. Additionally, CD30L has a role in the segregation of B and T cells within the murine spleen [45] and so may have an as yet uncharacterized role in iBALT formation. MIP-3β/CCL19 is involved in lymphocyte recruitment, and inhibition by M3 has been proposed as a survival advantage for MHV-68 [46]. Our observations are in agreement with this hypothesis. Moreover, BALT is spontaneously-produced in mice that are deficient in the receptor for CCL19 (CCR7−/−), a phenomenon that is related to a defect in homing of regulatory T cells [47]. In contrast, MIP-3β/CCL19 has been implicated in iBALT formation in lymph-node and spleen-deficient laboratory mice [41], which is at odds with our results. Thus, iBALT formation is complex and modulation in the levels of factors such as SDF-1α/CXCL12, CD30L/CD153 and MIP-3β/CCL19 by M3 may contribute to the formation of iBALT in context of MHV-68 infection in wood mice.
As noted above, a number of T cell, monocyte/macrophage, and neutrophil associated chemokines were present in higher levels within the lungs of wood mice infected with M3.stop relative to M3.MR. RANTES/CCL5 is an important proinflammatory chemokine that induces the recruitment of T cells (including CTLs), monocytes and eosinophils to the sites of virus infection. Other studies have shown that blocking RANTES/CCL5 in vivo significantly increases the titers of respiratory syncytial virus in the lungs of infected mice, and this is associated with reduced T cell recruitment [48] and heightened lung disease. Additionally, influenza virus infection of MIP-1α/CCL3−/− mice leads to a reduced inflammatory response and increased virus titers [49]. MIP-2/CXCL2 induces neutrophil recruitment [50]. Hence, inhibition of such chemokines by M3 conceivably favors MHV-68, not necessarily to increase virus replication, but to promote establishment of latent infection and virus persistence, a hallmark property of all herpesviruses.
Leptin receptor and its ligand (an IL-6 family member) were expressed at elevated levels in the lungs of M3.stop infected wood mice (22-fold and 10-fold respectively). Leptin is an adipocyte-derived cytokine that regulates energy intake and expenditure. However, leptin promotes Th1 immune responses as well as inducing cytokine secretion and increasing phagocytosis by macrophages (reviewed in [51]). Deficiency in leptin production has also been associated with susceptibility to pulmonary disease in a mouse model [52]. Thus, modulation of leptin by the indirect action of M3 may confer a survival advantage for MHV-68.
The cellular and biochemical consequences of M3 expression are clearly complex. M3 is a chemokine-binding protein, and as such is thought to disrupt chemokine gradients, modulating the response of cells in vivo. Thus, a lack of M3 should increase recruitment of cells that respond to the chemokines bound by M3. Thus, the changes in cytokine and chemokine profiles that we observe may be due to modulation of the composition of infiltrating cell types and the activation status of these cells. Nonetheless, levels of the chemokines RANTES/CCL5, MIP-1α/CCL3, MIP-3β/CCL19 and fractalkine/CX3CL1 that are known to be bound or functionally impaired by M3 [15], [16], [46] were lower in the wood mice infected with wild-type virus, and thus a direct effect of M3 on chemokine levels could also play a role.
At this juncture, it is unclear what the basis is for the lack of an apparent influence of M3 in the context of MHV-68 intra-nasal infection in laboratory strains of mice [17]. Given the relatively close genetic relationship between M. musculus and A. sylvaticus, that there is such a notable difference in the role of M3 is surprising, particularly since there is a change in the inflammatory response (predominantly neutrophilic) within brain to MHV-68 infection following intracerebral inoculation of CD1 mice with M3.stop virus [17]. This response is distinct from that seen after intranasal infection of wood mice where few neutrophils are present, but suggests that M3 is indeed capable of functioning within M. musculus, and that the absence of an apparent influence of M3 in the lung and spleen in this host, therefore, may be due to relatively subtle differences between this species and the natural host. One possibility is that M3 expression in the lung and spleen of a M. musculus host is below a critical threshold. However, comparative analyses of MHV-68 mRNA expression in M. musculus-derived cells, albeit within infected cells in vitro, have revealed that M3 is one of the most highly expressed MHV-68 genes during the virus lytic cycle [27]. When we assessed M3 mRNA levels in the lung of infected wood mice and BALB/c mice at 7 and 14 days p.i., M3 expression in BALB/c mice was similar to wood mice at day 7, but significantly lower at day 14 (Fig. 1B). Our observed kinetics of M3 expression in BALB/c mice fits with a previous study [53]. Additionally, high levels of M3 mRNA were detected much later (14 days p.i.) in wood mouse lungs than mRNAs from the other genes in this locus (Fig. 1A), suggesting that the timing of M3 expression may be important, and that in M. musculus the lower level (approx. 10 fold) of M3 at day 14 p.i. may be critical. In this context, previous work has shown that the peak of chemokine expression in MHV-68–infected BALB/c mice occurs after the peak of M3 expression [53], [54] and that deletion of the M3 locus does not affect chemokine levels [53]. Alternatively, the cellular source and location of M3 may play a role. While this may reflect latency-associated M3 expression, at this time we also detected equivalent levels of mRNA from ORF50, a key gene of the lytic cycle. Finally, a possibility worthy of consideration is that minor species differences in cytokine(s) protein structure have combined with coding changes in M3 that have occurred during passage of MHV-68 in vitro to render M3 less effective within M. musculus. Such changes in M3 would be possible due to a reduction in selective pressure to retain M3 integrity in vitro, where it neither contributes directly to nor is it essential for MHV-68 replication [17], [33].
In summary, the results from this study demonstrate that M3 is important for MHV-68 infection by facilitating an environment in which proliferating B cells would accumulate, both during iBALT formation in the lungs and the germinal center reaction in the spleen. These responses ultimately lead to efficient establishment and augmentation of MHV-68 latency, in both the lungs and spleens of its natural host. Significantly, this work also highlights the importance of using the natural host for studying the role of virus genes, particularly those involved in modulating the innate and adaptive host antiviral response, whose functions have no doubt intricately evolved within the context of a specific host.
All animal work was performed under strict accordance with UK Home Office guidelines and approved by the UK Home Office under Project Licence numbers 40/2483 and 40/3403 and Personal Licence number 60/6501.
Wood mice (Apodemus sylvaticus) were obtained from an out-bred colony established at the University of Liverpool, Faculty of Veterinary Science [55], [56]. This colony was obtained from Dr. J. Clarke in 1995, and derived from captive-bred colonies maintained for several decades in the Department of Zoology, University of Oxford, UK with only occasional introductions of new stock from the wild. Their general housing and maintenance has been described elsewhere [57], and at Liverpool they are maintained under semi-barrier conditions. The Liverpool colony has suffered no clinical disease, and, although not specified pathogen free (SPF) in the sense used for most laboratory rodents, samples are tested routinely on a monthly basis for the major infections of laboratory rodents have so far been negative. Of particular relevance to this study, no evidence of MHV-68 infection has been found in the colony by serology and PCR analysis [19]. Animals were anesthetized with isoflurane and inoculated with 4×105 plaque forming units (PFU) in 40 µl of sterile phosphate buffered saline (PBS). At various times between day 3 and 40 p.i., animals were euthanized and tissues were harvested.
Stocks of MHV-68, clone g2.4 [58], and previously published mutant MHV-68 viruses M3.stop and M3.MR [16] were grown and titrated by infection of baby hamster kidney cells (BHK-21), as previously described [23]. BHK-21 cells were maintained in Glasgow's Modified Minimal Essential Medium with 10% newborn calf serum and 10% tryptose-phosphate broth, 2 mM L-glutamine, 70 µg/ml penicillin and 10 µg/ml streptomycin. NIH3T3 cells were maintained in Dulbecco's Modified Eagles Medium (DMEM) supplemented with 10% fetal bovine serum, 2 mM L-glutamine, 70 µg/ml penicillin and 10 µg/ml streptomycin.
Total RNA was purified from lung tissue using the RNeasy Mini Kit (Qiagen) and DNA contamination removed by treating RNA with amplification grade DNase I (Invitrogen) according to the manufacturers' recommendations. Reverse transcription was performed at 50 °C for 30 min with 2 µg RNA in a 20-µl reaction containing 200 U Superscript III reverse transcriptase (Invitrogen), 500 ng oligo(dT)15 primer (Roche), 0.5 mM dNTP mix (Promega), 5 mM DTT, 40 U RNase inhibitor (RNaseOUT; Invitrogen), and First-Strand buffer (50 mM Tris-HCl [pH 8.3], 75 mM KCl, 3 mM MgCl2; Invitrogen). Afterwards, 2 µl was used as template for qRT-PCR in 20-µl reaction volumes. Quantification of cDNA was done using an Opticon Monitor 2 real-time PCR machine (MJ Research) with DyNamo SYBR Green kit (Finnzymes) and 0.5 µM of each oligodeoxynucleotide primer (the oligodeoxynucleotide primers used for PCR and qRT- PCR are provided in Table 1). The cycling parameters were initially 95 °C for 10 min, and then for each cycle: 94 °C for 10 s, 60 °C for 20 s, and 72 °C for 15 s. Melting curve analysis was carried out between 65–95 °C with 0.2 °C increments to confirm product specificity. For each individual experiment, amplification of cDNA from the murine ribosomal protein L8 mRNA (RPL8; accession # AF091511) was used to normalize for input cDNA between samples using exon-spanning primers to control for contaminating cellular DNA. Each sample was amplified in triplicate, and mean cDNA copy numbers were determined from three individual mice and expressed relative to the copy number of RPL8 cDNA.
Quantification of viral DNA copy number (per 200 ng DNA) was determined as previously described [59] using PCR primers specific for the MHV-68 gp150 gene. The RPL8 gene was used to normalize for input DNA between samples. Mean viral genome copy numbers were determined from three or four individual infected animals. Splenocytes isolated from intact spleens were examined for latent virus by an infective center assay using NIH3T3 cells, as previously described [23].
Lung, spleen, and lymph node tissue were fixed in 4% buffered paraformaldehyde and routinely embedded into paraffin wax. Sections (3–5 µm) were either stained with haematoxylin and eosin, or used for immunohistology or RNA in situ hybridization. Immunohistology was performed using the peroxidase anti- peroxidase and the avidin biotin peroxidase complex method as previously described [60]. T cells were detected using rabbit anti-human CD3 antibody (DAKO Cytomation). B cells were identified using rat anti-mouse CD45R (clone RA3-6B2; SouthernBiotech). Quantification of B and T cells was performed by counting the number of cells in corresponding sequential sections identified by the above antibodies. Five randomised areas of interstitial and perivascular/peribronchiolar infiltration and all areas of iBALT in the lung sections were analysed, using images captured with Nikon NIS-Elements Basic Research v3.0 software at 20× magnification. The proportion of B and T cells was then calculated for each area of peribronchiolar/perivascular infiltration and iBALT, and the number of cells per unit area for the interstitial infiltrate. These data are shown as the mean values±SEM and compared between groups using a two sample t-test. Detection of MHV-68 M3 RNA and vtRNAs by RNA in situ hybridization followed a previously described protocol [61]. Briefly, digoxigenin (DIG)-labeled sense and antisense probes were generated by in vitro transcription, using the DIG RNA labeling kit (Roche), of either the entire M3 ORF that was amplified from MHV-68 DNA (see Table I for primers used) and cloned into pCRII (Invitrogen) or transcripts to the MHV-68 tRNA genes 1-4 within plasmid pEH1.4 as described previously [10]. Briefly, sections were treated with proteinase K (1 µg/ml; Roche) at 37 °C for 15 min, and hybridization performed overnight at 52 °C. Hybridized probe was detected with alkaline phosphatase-conjugated anti-DIG Fab fragments (Roche) and BCIP/NBT (Sigma). Slides were counterstained with Papanicolaoùs hematoxylin.
Lungs were screened for expression of 61 cytokines/chemokines using a RayBio Mouse Cytokine Antibody Array Kit (Array 3.1.; Ray Biotech Inc., Norcross, GA), performed according to the manufacturer's instructions. Lung tissue (20–30 mg) taken from wood mice 14 days p.i. with either M3.stop or M3.MR was homogenized in 500 µl lysis buffer (RayBiotech) containing 1% (w/v) sodium deoxycholate, 2% (v/v) NP-40, 0.2% (w/v) SDS, 1 µg/ml each of aprotinin, leupeptin, pepstatin, and 1 mM phenylmethylsulfonyl fluride (PMSF) on ice. Protein concentrations were determined using a BioRad DC-Protein Assay Kit according to the manufacturer's instructions. As an extra control, 25 µg protein from each sample was analyzed by western blot to detect actin to ensure analysis of equal starting material (data not shown). Cell lysates were sent to RayBiotech (RayBiotech, Inc. 3607 Parkway Lane, Suite 200, Norcross GA 30092, U.S.A.) for analysis of chemokine and cytokine levels using the RayBio Mouse Cytokine Antibody Array 3.1 kit (RayBiotech), using 500 µg protein per membrane. Signals were detected and quantified by chemiluminescence.
Lungs were screened for expression of specific chemokines by ELISA. Lung tissue (20–30 mg) taken from mice was homogenized in 1 ml of ice-cold T-PER Tissue Protein Extraction Reagent (Pierce) in the presence of protease inhibitor cocktail (Sigma-Aldrich) before being clarified by centrifugation (10,000 g for 5 minutes at 4 °C). Total protein concentrations were determined by using DC-Protein Assay Kit (BioRad) according to the manufacturer's instructions. Chemokine concentrations were measured using DuoSet ELISA Development systems for RANTES/CCL5 (DY478), KC/CXCL1 (DY453), fractalkine/CX3CL1 (DY472), SDF-1α/CXCL12 (DY460), BLC/CXCL13 (DY470) and CD30 Ligand (CD153) (DY732) in accordance with manufacturer's instructions (R&D Systems Europe Ltd., Abingdon, UK). Lung tissues lysates were investigated in duplicate and diluted as appropriate to ensure protein concentrations were within the linear range of the standard curve. Optical densities were determined at 450 nm using a Thermo Labsystems Opsys MR ELISA plate reader (Thermo Life Sciences, Basingstoke, UK).
|
10.1371/journal.ppat.1006173 | IL-27 Limits Type 2 Immunopathology Following Parainfluenza Virus Infection | Respiratory paramyxoviruses are important causes of morbidity and mortality, particularly of infants and the elderly. In humans, a T helper (Th)2-biased immune response to these infections is associated with increased disease severity; however, little is known about the endogenous regulators of these responses that may be manipulated to ameliorate pathology. IL-27, a cytokine that regulates Th2 responses, is produced in the lungs during parainfluenza infection, but its role in disease pathogenesis is unknown. To determine whether IL-27 limits the development of pathogenic Th2 responses during paramyxovirus infection, IL-27-deficient or control mice were infected with the murine parainfluenza virus Sendai virus (SeV). Infected IL-27-deficient mice experienced increased weight loss, more severe lung lesions, and decreased survival compared to controls. IL-27 deficiency led to increased pulmonary eosinophils, alternatively activated macrophages (AAMs), and the emergence of Th2 responses. In control mice, IL-27 induced a population of IFN-γ+/IL-10+ CD4+ T cells that was replaced by IFN-γ+/IL-17+ and IFN-γ+/IL-13+ CD4+ T cells in IL-27-deficient mice. CD4+ T cell depletion in IL-27-deficient mice attenuated weight loss and decreased AAMs. Elimination of STAT6 signaling in IL-27-deficient mice reduced Th2 responses and decreased disease severity. These data indicate that endogenous IL-27 limits pathology during parainfluenza virus infection by regulating the quality of CD4+ T cell responses and therefore may have therapeutic potential in paramyxovirus infections.
| Respiratory viral infections are important propagators of acute and chronic disease, and a subset of those affected require hospitalization. Type 2 immune responses have a well-established association with increased disease severity; however, these responses have not been manipulated to ameliorate disease severity in clinical practice. Here we show that the cytokine interleukin-27 (IL-27) limits immunopathology after paramyxovirus infection. IL-27 regulates the quality of the inflammatory response, independent of viral replication, by restricting pathologic CD4+ T cell- and type 2 innate immune responses. As such, IL-27 emerges as an endogenous regulator of pathologic inflammation after respiratory viral infection and therefore may have both diagnostic and therapeutic potential in clinical medicine.
| Acute respiratory infections are an important cause of morbidity and mortality in children and adults [1, 2]. Paramyxoviruses including respiratory syncytial virus (RSV), parainfluenza virus, and metapneumovirus are among the most frequent causes of severe illness [1]. Although type 1 immune responses characterized by the generation of Th1 CD4+ T cells are essential for viral clearance and the development of protective immunity during these infections, severity of illness is associated with type 2 polarization of the immune response [3–5]. Despite their clinical importance, the mechanisms that regulate the development of type 2 immune responses during respiratory viral infections are unknown.
IL-27 is a heterodimeric cytokine composed of the Epstein-Barr virus-induced gene 3 (EBi3, also shared with IL-35) and IL-27p28 subunits [6, 7]. It engages a receptor formed by gp130 and the IL-27Rα to activate the Janus kinase (JAK)-signal transducer and activator of transcription (STAT) signaling pathway [6, 8]. Endogenous IL-27 regulates T cell responses in various models of inflammation [9–12]. In addition to antagonizing T cell production of IL-2, IL-27 directly inhibits Th2 and Th17 activities [10, 11, 13] and is a potent inducer of IL-10, a cytokine that antagonizes the function of antigen presenting cells (APC) [14, 15].
The murine parainfluenza virus Sendai (SeV) induces an acute respiratory infection in mice that is cleared by the immune system and that leads to type 2 immune pathology in the lung at chronic time points, mimicking what is observed in humans [16, 17]. IL-27p28 transcripts are increased transiently in the lungs of mice during SeV infection [17], and IL-27 is required for the expansion of Sendai-specific CD8+ T cells [18, 19]; however, the cellular source of IL-27 and its effects on viral control and on the development of immune pathology remain unknown.
Here we show that following infection with SeV, IL-27 deficiency leads to increased lung pathology and disease severity as well as to higher frequencies of eosinophils and alternatively activated macrophages (AAMs), consistent with an increased bias towards a type 2 immune response. SeV infection in control mice induced IFN-γ from CD4+ T cells as well as a subset of IL-10-producing IFN-γ+ CD4+ T cells in the lung. IL-27 deficiency was associated with a loss of the IFN-γ/IL-10 double producers and with the emergence of IFN-γ/IL-13 and IFN-γ/IL-17 double-producing CD4+ T cells. Depletion of CD4+ T cells in IL-27Rα-/- mice led to a decrease in AAMs and reduced weight loss after infection, while eliminating STAT6 signaling in IL-27-deficient mice reduced Th2 responses and decreased mortality. Taken together, these studies support a model in which IL-27 promotes IFN-γ+/IL-10+ CD4+ T cells and restricts the emergence of pathologic CD4+ T cells that contribute to type 2 immune pathology during Sendai virus infection.
Sendai virus replication in the lung has been associated with increased levels of IL-27p28 mRNA [17], but the cellular source was unclear. In mice, IL-27p28 is secreted in excess, and EBi3 is secreted when co-expressed with IL-27p28 [20]. Accordingly, at 10 days post infection (dpi) transcripts of IL-27p28 were detected in higher amounts than EBi3 by qPCR relative to internal control (Fig 1A). Consistent with these findings, analysis of bronchoalveolar lavage (BAL) protein revealed elevated IL-27p28 production at 6 and 10 dpi with no IL-27p28 detected at 4 and 18 dpi (Fig 1B). To identify the sources of IL-27, different populations of APCs in the lungs and secondary lymphoid organs of mock- or SeV-infected C57BL/6 wild type (WT) mice were evaluated by flow cytometry for IL-27p28 expression. Given the high levels of monocyte autofluorescence in the lung, infected IL-27p28-/- mice were also used as a control. Analysis of lung, draining lymph nodes, and spleen revealed that IL-27p28 was not detected in mock controls at local or peripheral sites or in infected mice at 4 or 22 dpi. No IL-27p28 was detected in the spleen or draining lymph nodes (dLN) (Fig 1C) at any time point after infection. In the lung, IL-27p28 was primarily produced by a population of CD11cmidMHC II- cells (Fig 1C). These cells were CD11b+ and expressed high levels of Ly6c and CD64, consistent with monocytes and/or monocyte-derived macrophages (S1 Fig). In the absence of CCR2, a population of CD11cmidMHC II-/CD11b+/CD64+/Ly6c- cells produced IL-27 in response to SeV, suggesting that other lung macrophages retain the ability to make IL-27 even in the absence of recruited monocytes (S1 Fig). No IL-27p28 was detected from CD11chiMHC II+ cells, representative of dendritic cells (DCs) and/or alveolar macrophages (AM). Neither CD4+ nor CD8+ T cells produced IL-27p28. These data indicate that a discrete population of lung macrophages transiently produces IL-27 in the lung after SeV.
To directly assess the impact of endogenous IL-27 on the outcome of SeV infection, WT and IL-27Rα-/- mice were monitored for weight loss as a measure of disease progression. Infected WT mice experienced transient weight loss; however, consistent with enhanced disease severity, infected IL-27Rα-/- mice lost more weight than controls (Fig 2A). The increased weight loss observed in IL-27Rα-/- mice was consistent across multiple experiments using male or female mice, and similar results were observed using IL-27p28-/- and EBi3-/- mice. Male gender is associated with increased severity of respiratory diseases in mice and humans [21], and in fitting with this increased severity of disease half of the infected male mice succumbed to infection between days 9 to 11 (Fig 2B). To determine whether the increased morbidity observed in IL-27Rα-/- mice was due to a reduced ability to clear virus, viral load was evaluated at the peak of SeV replication (3 dpi) and at the time of viral clearance (10 dpi) [22]. Infected WT and IL-27Rα-/- mice had comparable viral titers and levels of viral transcripts (Fig 2C) at these time points. These data indicate that the increased disease severity seen in the IL-27Rα-/- mice was not due to impaired viral clearance.
To determine if the absence of the IL-27Rα was associated with more severe lung pathology, WT and IL-27Rα-/- mice were infected, and lung tissue was analyzed for histopathological differences at 10 dpi. Mice given a mock infection with intranasal PBS did not show any evidence of lung inflammation (Fig 2D). In infected mice, inflammatory lesions were consistent with published reports and included multifocal bronchointerstitial pneumonia, alveolar and bronchiolar epithelial necrosis, and inflammatory cell infiltrate composed of lymphocytes, plasma cells, neutrophils, and macrophages [22]. Although bronchoalveolar hyperplasia was not different between the groups, in comparison to infected WT mice, infected IL-27Rα-/- mice exhibited more severe inflammation in the lung, although to a lesser degree than expected given the differences in weight loss and survival (Fig 2D and 2E). Despite the increased mortality in males, no obvious differences in viral clearance, lung lesions, or inflammatory cell types of interest were observed between males and females when compared at 10 dpi (S2 Fig).
Type 2 immune responses are important mediators of lung pathology at late time points after SeV infection [16, 17]; however, this phenotype has not been described in the acute phase of the immune response to the virus. Flow cytometric analysis of granulocyte populations at 10 dpi in IL-27Rα-/- mice showed an increased frequency and number of eosinophils in the lung (Fig 3A, S4 Fig), which is unexpected after SeV. In addition, there was a corresponding increase in the frequency and number of CD11b+ macrophages expressing the mannose receptor (CD206+) (Fig 3B, S4 Fig), a marker of AAMs [23]. Of note, the increases in eosinophils and AAMs were consistent among mice deficient in each subunit of IL-27 (EBi3-/-, IL-27p28-/-) (S3 Fig). Moreover, the shift to a type 2 immune response in the lung was consistent among infected males and females (S2 Fig). Therefore, while SeV is associated with a Th1-type response in the lungs of WT mice [22, 24], the data presented here suggest that the increased disease severity observed in IL-27Rα-/- mice may be associated with an immune environment more consistent with Th2-like responses.
Previous studies have highlighted that during SeV infection IL-27 promotes the expansion of Sendai-specific CD8+ T cells in the lung and that it induces IL-10 production by CD8+ T cells in vitro and in vivo [15, 18]. We also observed a decreased frequency of virus-specific CD8+ T cells in IL-27Rα-/- mice during SeV infection but did not detect a difference in IL-10 production by the total CD8+ T cell population (S5 Fig). However, CD4+ T cells are the major cytokine-producing T cells in the lung after SeV [22, 25], and the regulatory role of IL-27 on CD4+ T cells has not been explored in this model. In the absence of IL-27, CD4+ T cells from the lungs of mice infected with SeV produced less IL-10 at 10 dpi (Fig 4A and 4B). CD4+ T cells also represent a major source of IFN-γ after SeV infection, with peak production occurring at 10 dpi [22]. Therefore, lymphocytes were isolated from infected lungs at 10 dpi and stimulated ex vivo with phosphomolybdic acid (PMA), ionomycin, B-refeldin, and monensin to evaluate cytokine production by the total CD4+ T cell population. At this time point, total IFN-γ production by all CD4+ T cells was comparable in WT and IL-27Rα-/- mice and therefore independent of IL-27 signaling (Fig 4A and 4B). However, although total levels of IFN-γ were unchanged in infected IL-27Rα-/- mice, a population of the total CD4+ T cells that was double positive for IFN-γ and IL-10 was reduced in frequency (Fig 4A and 4B) and number (S4 Fig), consistent with work in other models of inflammation [11, 15, 26]. IL-27 also directly suppresses IL-4/IL-13 and IL-17 synthesis by CD4+ T cells [10, 27], and although levels of these cytokines in the BAL were not different between groups, analysis of all CD4+ T cells from infected IL-27Rα-/- mice revealed marked increases in IL-17 and IL-13 production (Fig 4C and 4D, S4 Fig). There was also a significant increase in the frequency of IFN-γ+ CD4+ T cells that co-produced IL-13 or IL-17 (Fig 4C and 4D). Of note, these data were consistent among mice deficient in each subunit of IL-27 (EBi3-/-, IL-27p28-/-) as well as in the IL-27Rα-/- mice (S3 Fig). Together, these results suggest that after SeV not only does IL-27 limit Th2 and Th17 responses in the lung but that it also prevents the development of polyfunctional IFN-γ+/IL-13+ and IFN-γ+/IL-17+ CD4+ T cells in this context.
Next, the SPICE data mining software was used to analyze the polyfunctional CD4+ T cell populations and evaluate for phenotypic patterns in infected WT and IL-27Rα-/- mice. As expected, the IFN-γ+/IL-10+ CD4+ T cells were the predominant polyfunctional population in infected WT mice (Fig 4E). In infected IL-27Rα-/- mice, the reduction in IFN-γ+/IL-10+ CD4+ T cells was associated with the emergence of both IFN-γ+/IL-17+ and IFN-γ+/IL-13+ CD4+ T cells. However, these populations were not present in the lung at equal frequencies. Consistent with the type 2 innate immune cells found in the lungs of IL-27Rα-/- mice, IFN-γ+/IL-13+ CD4+ T cells represented the largest population of dual producers (Fig 4E). Of the cytokines examined, neither WT nor IL-27Rα-/- CD4+ T cells were producing more than two cytokines concurrently. From these data, we conclude that during SeV infection IL-27 production induces a population of IFN-γ+/IL-10+ CD4+ T cells and that the loss of IL-27 signaling leads to the emergence of IFN-γ+/IL-17+ and IFN-γ+/IL-13+ polyfunctional CD4+ T cells.
Since the IL-27Rα-/- mice infected with SeV develop more severe disease that correlates with an altered CD4+ T cell phenotype, we next asked whether CD4+ T cells contributed to the enhanced morbidity. In these experiments, infected mice were treated with isotype control or a monoclonal antibody to deplete CD4+ T cells before and during SeV infection. Of note, the depletion of CD4+ T cells in WT mice did not abrogate the infection-induced changes in weight (S6 Fig), suggesting that the CD4+ T cells in WT mice are not pathogenic. In contrast, while infected isotype-treated IL-27Rα-/- mice lost weight as expected, depletion of CD4+ T cells dramatically decreased this weight loss (Fig 5A). CD4+ T cell depletion did not abrogate the pulmonary eosinophilia in IL-27Rα-/- mice (Fig 5B), but the frequency of AAMs in these mice was reduced to baseline levels after CD4+ T cell depletion (Fig 5C). Taken together, these data indicate that the absence of the IL-27Rα after SeV infection results in CD4+ T cells that have pathologic effects associated with an increase in AAMs.
Corresponding to the differences seen in infected IL-27Rα-/- mice, Th2-like responses are associated with enhanced lung damage in murine and human respiratory infection [3, 5]. Thus, experiments were performed to assess the impact of IL-4/IL-13 signaling on disease severity in IL-27Rα-/- mice. Because STAT6 is critical for IL-4- and IL-13-mediated signaling [28], mice deficient in the IL-27Rα and STAT6 (IL-27Rα-/-/STAT6-/-) were generated and then challenged with SeV. As previously shown, WT mice had transient weight loss, but IL-27Rα-/- mice experienced more severe weight loss after SeV. IL-27Rα-/-/STAT6-/- mice also lost more weight than WT mice but did not show any improvement over IL-27Rα-/- mice (Fig 5D). However, the increased mortality seen in infected IL-27Rα-/- male mice was ameliorated by elimination of STAT6 signaling (Fig 5E). As expected, the loss of STAT6 signaling in the IL-27Rα-/- mice resulted in a marked decrease in frequency of pulmonary eosinophils (Fig 5F) and AAMs (Fig 5G), and these mice did not mount Th2 responses after SeV (Fig 5H and 5I, S4 Fig). However, IL-27Rα-/-/STAT6-/- mice did exhibit a persistent loss of IFN-γ+/IL-10+ CD4+ T cells (Fig 5H and 5I), and IL-17 production by IFN-γ- and IFN-γ+ CD4+ T cells was also unchanged (S7 Fig). Therefore, elimination of STAT6 signaling reduced Th2 responses in IL-27Rα-/- mice and mitigated disease severity without restoring the IFN-γ+/IL-10+ CD4+ T cells. Together, these results suggest that IL-27 inhibits the development of Th2 responses during the acute response to SeV infection and that in the context of IL-27 deficiency the emergence of these responses contribute to the T cell-mediated pathology.
The data presented here reveal that IL-27Rα-/- mice challenged with SeV experienced increased disease severity mediated by CD4+ T cells as well as type 2 innate immune responses. In multiple models of infection the loss of the IL-27Rα has been associated with CD4+ T cell-mediated immunopathology linked with the overproduction of IFN-γ [12, 29, 30]. While the ability of CD4+ T cell depletion to ameliorate disease in IL-27Rα-/- mice is similar to other systems [12, 31], the contribution of the STAT6-dependent Th2 response to this pathology is unique and indicates that IL-27 is an important endogenous regulator of Th2 immune responses during parainfluenza virus infection.
There are multiple potential sources of endogenous IL-27, and while macrophages and DCs are considered major contributors [32], a more recent report suggested that during infection with Plasmodium berghei CD4+ T cells also produced IL-27 [33]. While we found no evidence that T cells produced IL-27p28 during this infection, there was a distinct population of monocytes and/or monocyte-derived macrophages in the lungs that produced IL-27p28. The induction of IL-27 during this infection could be a direct response to SeV [32] or may represent a secondary response to the SeV-induced production of type I interferons [32, 34] as during experimental autoimmune encephalitis and multiple sclerosis, the ability of type I IFNs to promote IL-27 is linked to their clinical efficacy [35].
Interestingly, SeV infection is lethal in the majority of male IL-27Rα-/- mice but not in females. Males have been shown to exhibit increased disease severity in both clinical and preclinical studies of respiratory illness [21, 36], and in other inflammatory systems immune responses are influenced by gender [37, 38]. An improved understanding of the impact of gender on immune responses during respiratory viral infection could have important implications for risk stratification and disease management in humans. Although no differences were noted between the immune parameters evaluated in males and females in this study, the increased mortality in male mice deficient in IL-27 signaling presents a unique opportunity for further study.
Based on work from our laboratory and others, abrogating type 2 immunity after SeV infection is not only important for limiting disease severity at early time points but also affects disease progression in the chronic phase [16, 17]. Interestingly, while CD4+ T cell depletion and elimination of STAT6 signaling in infected IL-27Rα-/- mice both led to decreased disease severity and fewer AAMs, these two interventions did not phenocopy each other. Pulmonary eosinophilia was decreased in infected IL-27Rα-/-/STAT6-/- mice but was unchanged by CD4+ T cell depletion. This sustained eosinophilia in the absence of CD4+ T cells may be explained by the ability of IL-27 to inhibit proliferation and cytokine production by type 2 innate lymphoid cells (ILC2s) [39, 40]. It could also be explained by epithelial cells, which express the IL-27Rα [41] and whose production of IL-13 contributes to airway remodeling [42]. Although the emergence of Th2-like innate and adaptive cells was the predominant phenotype in infected IL-27Rα-/- mice, Th17 cells have also been linked to immunopathology in IL-27Rα-/- mice infected with respiratory syncytial virus (RSV) [43], and our data did show persistent Th17 responses in infected IL-27Rα-/-/STAT6-/- mice. This suggests that Th17 cells may contribute to the immunopathology seen during SeV infection in IL-27 deficiency. Ultimately, despite the notable differences, both CD4+ T cell depletion and elimination of STAT6 signaling abrogated disease severity in infected IL-27Rα-/- mice. Therefore, by limiting type 2-mediated pathology after SeV, IL-27 regulates a clinically relevant pathway of the immune response after parainfluenza virus infection.
The production of IL-10 is an important mechanism to limit many types of inflammatory processes [11, 15, 29], and several cytokines including IL-6, IL-12, IL-21, IL-27, and type I interferons have been shown to promote T cell secretion of IL-10 [15, 44]. Many of these cytokines are produced during SeV, but the loss of the IFN-γ+/IL-10+ double producers in the IL-27Rα-/- mice indicates a dominant role for IL-27 in this experimental system. Until recently CD4+ T cell subsets were defined largely on a restricted ability to make unique profiles of cytokines (Th1 making IFN-γ, Th2 making IL-4/IL-13, Th17 making IL-17) after lineage commitment. However, co-expression of these cytokines can occur, and our data are reminiscent of studies that have shown human CD4+ T cells specific for C. albicans and M. tuberculosis that produce IFN-γ also possess the capacity to make IL-4 or IL-17 [45]. Dual cytokine-producing CD4+ T cells also have been identified in models of autoimmunity and viral infection [46–48], and IFN-γ+/IL-17+ CD4+ T cells have been described in pathogenic post viral immune responses [49]. The impact of IL-27 on these dual producers has not been previously described but is consistent with the ability of IL-27 to directly limit Th2 and Th17 differentiation [10, 11] and may represent a mechanism to limit T cell heterogeneity during viral infection.
The experiments presented here suggest that while the T cell response to SeV infection is dominated by the production of IFN-γ, IL-27 acts on CD4+ T cells not only to promote the production of IL-10 but also to limit the emergence of pathologic CD4+ T cell responses associated with severe disease. Identifying an endogenous regulator of type 2 immune responses in the lung during parainfluenza virus has therapeutic implications. Genetic analyses have established connections between IL-27p28 polymorphisms and the severity of acute respiratory infection in premature infants [50] and of chronic lung disease in adults and children [51–53], but additional clinical studies are needed to look at this relationship in the context of respiratory viral infection. Nevertheless, taken together with the data presented here, these studies indicate that endogenous IL-27 may have a broad role in limiting immunopathology in the respiratory system and may be useful as an adjunct therapy to limit disease progression.
EBi3-/-, and IL-27p28-/- mice were provided by Sage Research Labs (Boyertown, PA). IL-27Rα-/- mice were provided by Amgen and bred in our facility. C57BL/6 mice were from Sage Research Labs (Boyertown, PA) or Taconic (Germantown, NY). STAT6-/- mice were obtained from The Jackson Laboratory (Bar Harbor, ME). All mice strains were housed, maintained, and bred under specific pathogen-free conditions at the University of Pennsylvania.
All experimental procedures with mice were approved by the Institutional Animal Care and Use Committee (IACUC) of the University of Pennsylvania in accordance with guidelines of the Association for Assessment and Accreditation of Laboratory Animal Care (Protocol #805045). IACUC uses the NIH Guide for the Care and Use of Laboratory Animals, which is based on the U.S. Government Principles for Utilization and Care of Vertebrate Animals Used in Testing, Research, and Training. Mice were euthanized by administration of CO2 for at least five minutes in accordance with these guidelines.
SeV strain 52 was propagated in the laboratory of C.B.L. as previously described [34]. Mice were infected intranasally with 10 ID50 of SeV strain 52 (10,000 TCID50/mouse) after anesthesia with ketamine and xylazine. Mice were infected after anesthesia with ketamine and xylazine. Mice were weighed on the day of infection (day 0) and every two days following, with weight loss calculated as percent change from original weight. Virus titration was performed by hemagglutination of chicken red blood cells (Lampire Biological Products) as previously described [34]. Briefly, the lungs were extracted, homogenized in PBS-gelatin (1%), and frozen at -80°C for preservation. The presence of infectious particles was evaluated by infecting LLCMK2 cells with 1:10 dilutions of the lung homogenates at 37°C. After 1 hour of infection, 175 μL of medium containing 2 μg/mL trypsin was added and the cells were further incubated for 72h at 37°C. A total of 25μL of medium was then removed from the plate and tested by hemagglutination of chicken red blood cells (RBCs) for the presence of virus particles. Viruses at 1:4 dilutions in 0.5% chicken RBCs were incubated for 30 min at 4°C. The hemagglutination of RBCs indicated the presence of virus particles [24]. For the CD4+ T cell depletion, anti-CD4 monoclonal antibody (clone GK1.5, BioXCell) was administered at 250 μg/mouse on days 0, 4, 7. Rat IgG2b (BioXcell) was used as an isotype control.
Total RNA was purified from lungs using TRIzol (Life Technologies). RNA was reverse transcribed and amplified with specific primers in the presence of Power SYBR Green PCR Master Mixture (Applied Biosystems; Life Technologies). The primers for IL-27p28, EBi3, and β-actin were obtained from Quantitect (QIAGEN), and the SeV Np primer sequence was as follows: (Forward 3’-TGCCCTGGAAGATGAGTTAG Reverse 5’-GCCTGTTGGTTTGTGGTAAG). Normalization was conducted based on levels of beta-actin.
Single cell suspensions were generated from mouse spleens, lungs, and lymph nodes. Mouse spleens and lymph nodes were collected and dissociate through a 70 μm strainer. Red blood cells from spleens were lysed using 0.85% ammonium chloride (Sigma). Red blood cell lysis was not performed on lymph nodes. Lungs were inflated with a solution of 1 mg/ml collagenase A (Roche) and 100 μg/ml DNase (Roche), and then diced and digested in the same solution for 60 min at 37°C to obtain a single-cell suspension. Red blood cells were lysed using 0.86% ammonium chloride (Sigma). Cells from all tissues were counted, washed in flow cytometry buffer (1% BSA (Sigma), 2mM EDTA (Invitrogen) in PBS) and stained for surface markers. For assessment of cytokine production, T cells were restimulated with PMA and ionomycin plus brefeldin A (BGA) and monensin (Sigma) and stained for surface markers, then fixed with 4% paraformaldehyde for ten minutes prior to intracellular staining for relevant cytokines [54]. Cells were blocked with 2.4G2 (BioXCell) and Rat IgG (Invitrogen) before staining with monoclonal antibodies. Samples were acquired using an LSRFortessa flow cytometer (BD Biosciences) and analyzed with FlowJo software (Tree Star, Inc.) and SPICE software (NIAID). Viable cells were identified using the LIVE/DEAD Fixable Aqua Dead Cell Stain Kit for 405nm excitation (Invitrogen). The following mAb against mouse antigens were used for staining: FITC-anti-CD3 (clone 145-2c11), FITC-anti-CD19 (clone 6D5), FITC-anti-NK1.1 (clone PK136), APC-anti-FcεR (clone MAR-1), PE-anti-Siglec F (clone E50-2440), ef780-anti-CD11b (clone M1/70), PerCP-anti-Ly6c (clone HK1.4), Pacific Blue-anti-Ly6G (clone 1A8), ef780-anti-CD3 (clone 145-2c11), PE CF594-anti-CD4 (clone GK1.5), Pacific Blue-anti-CD8α (clone 53–6.7), FITC-anti-Foxp3 (clone 150D/E4), PE Cy7-anti-IFN-γ (clone XMG1.2), PE-anti-IL-13 (clone eBio13A), APC-anti-IL-10 (clone JES5-16E3), PerCP-anti-IL-17A (clone eBio17B7), AF700-anti-CD45 (clone 30-F11), ef780-anti-CD11c (clone N418), PE-anti-F4/80 (clone BM8), PerCP-anti-CD11b (clone M1/70), APC-anti-CD206 (clone C068C2), PE-anti-CD11a (clone H155-78), PE-Cy7-anti-KLRG1 (clone 2F1), Pacific Blue-anti-MHC class II (clone M5/114.15.2). IL-27p28 production was evaluated in the lung by flow cytometry (APC-anti-IL-27p28, clone MM27-7B1) and in the bronchoalveolar lavage fluid using an IL-27p28 specific ELISA (R&D Systems). Antigen-specific cells were identified using a conjugated tetramer to the Sendai virus nucleoprotein fragment FAGPNYPAL, provided by the NIH tetramer core. IL-27p28 production was evaluated in the lung by flow cytometry and in the bronchoalveolar lavage fluid using an IL-27p28 specific ELISA (R&D Systems).
After lavage, the left lobe of the lung was inflated and fixed with 0.5 ml of 10% neutral-buffered formalin solution. Fixed lung tissues were embedded in paraffin, and sections were cut using a standard procedure. Deparaffinized sections from fixed lungs were stained with hematoxylin and eosin. Lung inflammation was scored by a board-certified veterinary pathologist (E.L.B.). Lung inflammation was scored according to the following scale: 0, no manifestation; 1, minimal and peribronchiolar only; 2, mild, in areas of epithelial hyperplasia; 3, moderate; and 4, severe. The severity of bronchiolar/alveolar epithelial hyperplasia was scored as percent of affected 40X area and percent of terminal bronchioles affected, as follows: Percent of affected area: 0, 0%; 1, 1–25%; 2, 26–50%; 3, 51–75%; 4, 76–100%; Percent of terminal bronchioles affected: 0, 0–20%; 1, 21–40%; 2, 41–60%; 3, 61–80%; 4, 81–100%. Bronchiolar smooth muscle hyperplasia was evaluated as follows: 0, none; 1, mild; 2, moderate; 3, severe.
Bar graphs and scatter plots were plotted as means with the SEM in Prism 5 software (GraphPad). All statistics were performed using an unpaired Student t test, except the histologic evaluation between WT, IL-27Rα-/-, and IL-27Rα-/-/STAT6-/- in which the Mann-Whitney test was used and survival analyses, which were evaluated by the log-rank (Mantel-Cox) test.
|
10.1371/journal.pbio.0050177 | Development of the Human Infant Intestinal Microbiota | Almost immediately after a human being is born, so too is a new microbial ecosystem, one that resides in that person's gastrointestinal tract. Although it is a universal and integral part of human biology, the temporal progression of this process, the sources of the microbes that make up the ecosystem, how and why it varies from one infant to another, and how the composition of this ecosystem influences human physiology, development, and disease are still poorly understood. As a step toward systematically investigating these questions, we designed a microarray to detect and quantitate the small subunit ribosomal RNA (SSU rRNA) gene sequences of most currently recognized species and taxonomic groups of bacteria. We used this microarray, along with sequencing of cloned libraries of PCR-amplified SSU rDNA, to profile the microbial communities in an average of 26 stool samples each from 14 healthy, full-term human infants, including a pair of dizygotic twins, beginning with the first stool after birth and continuing at defined intervals throughout the first year of life. To investigate possible origins of the infant microbiota, we also profiled vaginal and milk samples from most of the mothers, and stool samples from all of the mothers, most of the fathers, and two siblings. The composition and temporal patterns of the microbial communities varied widely from baby to baby. Despite considerable temporal variation, the distinct features of each baby's microbial community were recognizable for intervals of weeks to months. The strikingly parallel temporal patterns of the twins suggested that incidental environmental exposures play a major role in determining the distinctive characteristics of the microbial community in each baby. By the end of the first year of life, the idiosyncratic microbial ecosystems in each baby, although still distinct, had converged toward a profile characteristic of the adult gastrointestinal tract.
| It has been recognized for nearly a century that human beings are inhabited by a remarkably dense and diverse microbial ecosystem, yet we are only just beginning to understand and appreciate the many roles that these microbes play in human health and development. Knowing the composition of this ecosystem is a crucial step toward understanding its roles. In this study, we designed and applied a ribosomal DNA microarray-based approach to trace the development of the intestinal flora in 14 healthy, full-term infants over the first year of life. We found that the composition and temporal patterns of the microbial communities varied widely from baby to baby, supporting a broader definition of healthy colonization than previously recognized. By one year of age, the babies retained their uniqueness but had converged toward a profile characteristic of the adult gastrointestinal tract. The composition and temporal patterns of development of the intestinal microbiota in a pair of fraternal twins were strikingly similar, suggesting that genetic and environmental factors shape our gut microbiota in a reproducible way.
| The adult human body typically comprises ten times more microbial cells than human cells, due largely to the extremely high density of microbes found in the human intestinal tract (typically 1011–1012 microbes/ml of luminal content). This microbial ecosystem serves numerous important functions for its human host, including protection against pathogens, nutrient processing, stimulation of angiogenesis, and regulation of host fat storage [1–7]. It is clear that this list is not yet complete; as this field of study expands, we are continually discovering new roles and relationships. Studies of gnotobiotic mice have been particularly enlightening, illustrating the essential role of the gastrointestinal (GI) microbiota in normal gut development [2,5]. In addition, numerous diseases in both adults and infants have known or suspected links to the GI microbiota, including stomach cancer [8], mucosa-associated lymphoid tissue lymphoma [9], inflammatory bowel disease [10,11], and necrotizing enterocolitis [12,13].
The composition of the adult GI microbiota has been intensely studied, using both cultivation and, more recently, culture-independent, small subunit (SSU) ribosomal DNA (rDNA) sequence-based methods [14]. The human colon ecosystem alone has been estimated to contain more than 400 bacterial species, belonging to a limited number of broad taxonomic divisions [15]. Members of the anaerobic genera Bacteroides, Eubacterium, Clostridium, Ruminococcus, and Faecalibacterium have typically been found to comprise a large majority of the human adult gut microbial community. Still, each adult's gut appears to have a unique microbial community, with a structure that remains stable on the time scale of months [3,15,16].
In contrast, the infant GI microbiota is more variable in its composition and less stable over time. In the first year of life, the infant intestinal tract progresses from sterility to extremely dense colonization, ending with a mixture of microbes that is broadly very similar to that found in the adult intestine [17]. Although the beginning and end points of this time course are well defined, the path between these points is poorly understood. There are conflicting reports in the literature regarding the composition of the neonatal GI microbiota and the factors that shape it. Several studies have reported that Bifidobacteria almost always dominate the GI microbiota of breast-fed infants by several weeks of age [17–20], while others find that they occur in only a small fraction of infants, or are not numerically dominant [21,22]. The effect of diet on the composition of the infant GI microbiota is also controversial—numerous studies have found a lower abundance of Bifidobacteria and a higher abundance of aerobic bacteria in the GI microbiota of formula-fed infants relative to breast-fed infants [20,21,23–25], yet other reports have found no such difference [26,27]. Mode of delivery has frequently been cited as one of the key factors that shape the infant microbiota [18,28,29]. The GI microbiota of infants delivered by caesarean section has been reported to differ from that of infants delivered vaginally, both in the timing of colonization and in composition [18,30–32], and in some cases, there are clearly traces of the maternal vaginal microbiota in the neonatal GI microbiota [33], yet the relative importance of mode of delivery on GI microbiota is unclear. Because of the increased incidence of GI problems in premature infants, the effect of gestational age has also been extensively studied. These studies have consistently shown that the microbiota of hospitalized, preterm infants differs from that of healthy, full-term babies [32,34–36]. Attempts to associate specific microbes with the occurrence of necrotizing enterocolitis, a condition with suspected bacterial etiology that is an important cause of morbidity and mortality in premature babies, have yielded mixed results [32,36]. Clearly, there is still much to be learned about the origins and development of the infant GI microbiota and its influence on health and disease.
We focused our study on describing the range of profiles that constitute a healthy infant GI microbiota in the hopes of discovering themes that govern its development, and in order to provide a detailed reference and a solid foundation for later studies examining the factors that influence the GI microbiota. Our study participants included 14 healthy, full-term babies, born to 13 healthy mothers (thus including one set of fraternal twins) (Table 1). Stool samples were collected according to a prescribed schedule, beginning with the first stool produced after birth: samples were collected daily at first and then with decreasing frequency over the course of 1 y, with additional sampling around key events (e.g., introduction of solid food and administration of antibiotics), yeilding an average of 26 stool samples per baby (Table 2). In addition, stool samples were collected from parents and siblings, and vaginal swabs and breast milk were collected from the mothers. We analyzed the microbiota of each of these specimens using a newly developed SSU rDNA microarray designed to give nearly comprehensive coverage of known SSU rDNA species. A subset of these samples was also analyzed by SSU rDNA clone library sequencing, for the purposes of calibrating and validating our microarray results.
To survey the composition of our sample set and to provide a basis for quantitative calibration of the microarray results, we created a reference pool by combining equal amounts of amplified SSU rDNA from each PCR-amplifiable sample (except for samples collected when the infants were ≥1 y old). We obtained 3,458 high-quality clone sequences from a library constructed from this pool, and taxonomically assigned each sequence using Ribosomal Database Project's Classifier [37]. The taxonomic distribution of these sequences is summarized in Table 3.
To assess the performance of our new microarray design relative to SSU rDNA sequencing, we sequenced SSU rDNAs amplified from each of 12 individual biological samples obtained in this study, selected for their diverse profiles by 16S rDNA microarray analysis. This study set included DNA extracted from eight baby stools, two maternal stools, one vaginal swab, and one breast milk sample. For each of these samples, we amplified SSU rDNA sequences using the same PCR primers that were used in the microarray analysis, then cloned and sequenced several hundred (mean = 342) of the amplified products for a total of 4,100 sequences.
We focused our comparison at levels 2, 3, and 4 of the prokaryotic multiple sequence alignment (prokMSA) hierarchy, which very roughly correspond to the phylum, class, and order levels in the classical taxonomic hierarchy. At these broader levels, most sequences are expected to have homology to at least one probe in our current microarray design, and rDNA sequences can generally be unambiguously classified. Microarray-based relative abundance estimates were obtained for 2,149 species and taxonomic groups by integrating data from all probes that represented any subset of the class in question, as fully described in Materials and Methods. Sequence-based estimates were obtained by taxonomically classifying each sequence by assigning the prokMSA operational taxonomic unit (OTU) code of the best BLAST match in the 2004 prokMSA database of 86,453 SSU ribosomal RNA (rRNA) gene sequences [38] (Datasets S1 and S2). Although the relative abundance of a bacterial species cannot be precisely determined from its proportional representation in a pool of amplified rDNA sequences, we expect that such estimates should be accurate within an order of magnitude and usually within a few-fold [39–41], based on previous studies that compared abundance levels estimated from sequencing SSU rDNA amplicons with counts based on in situ hybridization.
Overall, the microarray results were very similar to those obtained by sequencing, both qualitatively and quantitatively. Figure 1A shows the comparison of the community profiles of each of the 12 samples derived from our microarray analysis and by sequencing, for each taxonomic group at level 2 of the prokMSA taxonomic tree. Note that the levels (e.g., level 2) in the prokMSA taxonomy do not have a consistent correspondence with the levels (e.g., phylum) in the classical taxonomic hierarchy, and thus some of the conventional names associated with prokMSA level 2 groups can appear somewhat incongruous. Both the sequence analysis and the microarray analysis showed that the samples were dominated by a limited number of taxonomic groups—99% of the 4,100 sequences were encompassed by just three of the 22 level 2 prokMSA divisions: 2.15 (Flexibacter-Cytophaga-Bacteroides), 2.28 (Proteobacteria), and 2.30 (Gram-positive bacteria [including Firmicutes and Actinobacteria]), and the remaining 1% belonged to groups 2.10 (Prosthecobacter), 2.29 (Fusobacteria), or 2.21 (Cyanobacteria and Chloroplasts). As shown in Figure 1B and 1C, the population profiles obtained by microarray and sequencing analysis were also quantitatively similar—the Pearson correlation of the microarray- and sequencing-based estimates of relative abundance for the 12 samples was 0.97 at prokMSA taxonomic level 2 (Figure 1B), 0.89 at level 3 (Figure 1C), and 0.80 at level 4 (unpublished data).
We estimated the overall density of bacteria in each sample by a real-time quantitative PCR (qPCR) assay, using a broad-range bacterial primer and probe set (see Materials and Methods). We used the total number of rRNA gene copies (typically about five per genome [42]) per gram of stool, as estimated by this assay, to approximate the total density of bacteria. As shown in Figure 2, the total number of rRNA gene copies was relatively unstable throughout the first week of life, then persisted in most babies in the range of 109 to 1010/g of stool (wet weight). Although there was no clear effect of method of delivery on the timing of the colonization, it is noteworthy that babies 13 and 14 (the dizygotic twins), who were the only babies delivered by a planned caesarean section, and thus without rupture of the amniotic membrane and exposure to maternal birth canal microbiota during labor or delivery, had low bacterial counts (<108 rRNA gene copies/g) until the seventh day of life.
We analyzed the bacterial composition of 430 samples—363 infant stool samples, 43 adult stool samples, two sibling stool samples, 12 breast milk samples, and ten maternal vaginal swabs—by hybridization to the DNA microarray developed in this study. By combining information across multiple probes (see Materials and Methods), we obtained relative abundance estimates for 2,149 nested taxonomic groups and species in each of these samples (All probes are listed in Dataset S3; All taxa are listed in Dataset S4). As shown in Figure 3, the phylum-level diversity in the stool samples analyzed in this study was extremely limited. The vast majority of samples were dominated by just three of the 22 level 2 bacterial groups represented by our microarray: 2.15 (Flexibacter-Cytophaga-Bacteroides), 2.28 (Proteobacteria), and 2.30 (Gram-Positive Bacteria [Firmicutes and Actinobacteria]). A second major finding was the remarkable degree of interindividual variation in the colonization process. Although the taxa that populate the infant GI tract were limited at the broadest levels, each baby was distinct in the combination of microbial species that it acquired and maintained, and in the precise temporal pattern in which those species appeared and disappeared. Bacteroides, for example, dominated the early microbiota of some babies but were virtually absent at this stage in other babies. A third striking feature of this dataset was the relative stability of the microbial populations over time—even early in the course of the colonization of the infant GI tract, most taxonomic groups persisted over intervals of weeks to months.
The main dimensions of variation among the colonization profiles of different taxonomic groups were timing of colonization and temporal stability. Consistent with previous studies [28,35,43,44], the earliest colonizers were often organisms predicted to be aerobes (e.g., Staphylococcus, Streptococcus, and Enterobacteria), whereas the later colonizers tended to be strict anaerobes (Eubacteria, and Clostridia). The Bacteroides varied greatly from baby to baby in the timing of their first appearance, but were consistently present to some degree in nearly all babies by 1 y. Several other taxa, including Prevotella, Acinetobacter, Desulfovibrio, Veillonella, and Clostridium perfringens, tended to appear only transiently, sometimes appearing and disappearing repeatedly within a baby's first year of life.
We explored the similarities and differences in the composition of all of our samples by hierarchically clustering the 430 samples based on their similarity with respect to their abundance profiles for the set of 53 prokMSA level 4 taxonomic groups that had at least two samples with a relative abundance estimate greater than 1%. The clustering pattern, as reflected in the dendrogram at the top of Figure 4, highlights several critical features of the colonization program, and shows that the stool microbiota of babies 1 y of age and older is distinctly different from that at earlier ages and much more similar to that of adults. Prior to 6 mo of age, stool samples tended to cluster by baby, indicating that the differences from baby to baby are much greater than the changes over periods of weeks or months in the composition of any individual baby's microbiota. There were two notable exceptions to this baby-specific clustering. First, samples from the first few days of life often clustered away from the rest of a given baby's samples, sometimes clustering with other very early samples and sometimes with samples from other sites (e.g., baby 8 day 1 with vaginal samples). Second, samples from babies 13 and 14, who are fraternal twins, tended to intermingle. Figure 4B shows examples of several of the clustering patterns described above.
Most of the breast milk and maternal vaginal samples clustered perfectly by anatomic site of origin. As expected, all but one of the vaginal samples were overwhelmingly dominated by lactobacilli, with Staphylococci, Bacteroides, Clostridia, and Veillonella among the groups variably present as minority constituents. The vaginal sample from one of the mothers (mother of baby 11) had a distinctly different population profile, dominated instead by members of the Gamma Proteobacteria group. The microbial populations found in the milk samples were diverse, often including mixtures of enterics and species of Bacteroides, Pseudomonas, Haemophilus, Veillonella, and Streptococcus.
In order to compare the infants more systematically, we determined the nearest-neighbor sample for each sample as measured by the Pearson correlation of level 4 relative abundance estimates. Using this metric, the nearest-neighbor sample of any given baby sample was usually another sample from the same baby—the average percentage of samples from a given baby for which the most similar sample was from the same baby was 82%. Figure 5 summarizes this analysis and illustrates the interesting finding that by this measure, the most similar pair of babies by far was babies 13 and 14—fraternal twins raised in the same environment—8 of 23 (35%) of baby 13′s nearest-neighbor samples were from baby 14 (the next most similar pair was babies 11 and 14, at 17%).
The similarity of the microbial community profiles of stool samples from babies 1 y and older to each other and to those of the adult stool samples suggested that the infant GI communities converged over time toward a generalized “adult-like” microbiota. We explored this phenomenon by calculating, for each age interval, the average pairwise Pearson correlation of the population profiles of all infant samples collected at that age. As shown in Figure 6A, this analysis revealed that as time progressed, the babies' microbiota consistently converged toward a common profile. We also calculated, for each time point, the average correlation of infant samples at that time point to a generalized adult profile (centroid of 18 adult samples—nine fathers and nine mothers from this study). This analysis, shown in Figure 6B, confirmed that the profile toward which the infants' microbiota converges is similar to that of adults, and highlighted an apparent tendency for a population rearrangement to occur around 5 d after birth. Notably, the infants' GI microbiota was not significantly more similar to that of their parents than to that of other adults, as measured by the Pearson correlations of their level 4 taxonomic profiles (mean baby–parent correlation of 0.55 for within family, versus 0.62 between families for nine “triads” of contemporaneously obtained samples from baby, mother, and father obtained at 1–1.5 y of age).
To visualize the temporal patterns in the particular phylogenetic groups that populate the infant gut, we charted the relative abundance of the nine level 4 taxonomic groups that had a mean relative abundance of 1% or greater over time in each infant (Figure 7). This analysis enabled us to identify common themes and interesting differences among the colonization profiles of these babies. First, we observed that “uneven” populations (populations heavily dominated by a single taxonomic group) were common in the first several weeks but rare later in the time courses. Another notable feature in the temporal program of many of the babies was the occurrence of one or more dramatic shifts in the population structure—such shifts were frequently stabilized within one sampling interval. We were unable to identify any specific age or signal event consistently associated with such transitions, although the transition to an “adult-like” profile often followed the introduction of solid foods.
Several of the babies were treated with antibiotics either in the neonatal period (day 0–28) or in the later months (see Table 1 and Figure 2 for more details). In some cases, the treatment was associated with a striking alteration in the density or composition of the GI microbiota. For example, baby 8 received two courses of amoxicillin, one at 4 mo and one at 6 mo. In both cases, both the total density of bacteria (Figure 2) and the community composition were dramatically altered (Figures 3 and 6). Indeed, in this baby, the bacterial density in fecal samples decreased so much during the antibiotic courses that we were unable to amplify sufficient SSU rDNA for microarray analysis, so we could only compare the populations before and after the antibiotic course. However, we did not identify any consistent consequences of antibiotic treatment.
The results of both the sequence analysis of the reference pool and the microarray data analyses indicated that Bifidobacteria were only minor components of the population—a result at odds with the conventional wisdom [20,21,26]. The primers we used for broad-range PCR amplification of the reference pool (the source of the sequences) and samples for microarray analysis were potentially suboptimal for amplification of Bifidobacteria [21,26] due to three mismatches in the rDNA sequence of Bifidobacterium longum to the forward primer 8F used in this study. A survey of the 5′ sequences of full-length SSU rDNA genes showed that Bifidobacteria are outliers in their divergence from the generally conserved 8F primer sequence. We therefore carried out two independent analyses to determine whether and how the quantitative estimates of Bifidobacteria from the microarray hybridization results would need to be adjusted. First, we quantitatively evaluated the relative efficiency with which the 8F/1391R primer pair amplified SSU rDNA from two Bifidobacteria species—Bifidobacterium longum and Bifidobacterium infantis—compared to a set of three diverse common fecal bacteria—Escherichia coli, Clostridium perfringens, Bacteroides fragilis—all of which have SSU rDNA sequences with one or more mismatches to the 8F/1391R PCR primer sequences. Using a range of stoichiometric mixtures of chromosomal DNA extracted from these species, we found that after 20 cycles (the number of cycles used for our microarray analyses and for amplification of the reference pool prior to sequencing), efficiency of amplification of the two Bifidobacterial species' DNA was consistently 8-fold lower than that of the three other species, all of which amplified with nearly identical efficiencies (unpublished data). This result suggests that both the reference pool sequencing results and the microarray-based quantitation underestimated the abundance of the Bifidobacteria group by a factor of eight. Second, we used a real-time qPCR assay with a primer pair and probe optimized for detection of Bifidobacteria to obtain an independent estimate of the abundance of Bifidobacteria in each sample. The results confirmed the finding from the microarray analysis that Bifidobacteria were almost always only minor constituents of the fecal microbiota of both the infants and adults in our study population (Dataset S5 and Figure S1).
The majority of bacterial species identified in our sample set were previously reported constituents of the human GI microbiota. There were, however, a number of cases in which the microarray results indicated the presence of a bacterial species or group that was both unexpected and not represented in our sequenced reference pool. We investigated several of these cases using independent assays. For 12 of the suspect species/taxa, we used the cognate group-specific primers in a PCR assay applied to most or all of the samples in which the suspect species/taxa appeared to be present based on the microarray results, as well as a small set of samples in which the suspect species was not detected by the microarray. In one case, that of Sutterella wadsworthia, sequencing of the species-specific PCR product confirmed its presence. In seven of 12 cases, none of the array positive (or negative) samples yielded an amplified product in the PCR analysis. For four remaining cases, the ostensibly species-specific PCR assay yielded an amplified product of the expected size, but the clones sequenced from this product did not correspond to the expected species. We further investigated these four cases by sequencing a clone library obtained by amplification with the same broad-range primers that were used in preparation for microarray analysis. Although the sequencing did not confirm the presence of any of the four questionable species/taxa, it provided strong evidence for a major source of false-positive hybridization signals. Specifically, in three of the four cases, we identified a relatively abundant species whose rDNA sequence was sufficiently similar to the probe sequence that it was likely to account for the observed signal. In one case (Legionella pneumophila), which was predicted to be present at approximately 1%, we were unable to identify any candidate species that could account for the hybridization signal (i.e., none with best BLAST matches scores ≥30), among our set of 192 sequences. Since our power to detect a species present at a partial abundance of 1% was only 85%, it remains possible that this species, or another species with a similar SSU rDNA sequence, could have been present at a low abundance in the suspect samples.
Both our DNA extraction and rDNA amplification methods were optimized for bacteria and suboptimal for eukaryotes and archaea, thus we separately tested for the presence and abundance of fungi or archaea by means of qPCR assays with broad specificity for the respective taxonomic groups. Based on our qPCR analysis, fungi were intermittently detectable in stool samples at relatively low abundance (104–106 rRNA genes/g fecal wet weight), persisting for varying durations in individual babies, through the first year of life. One of the babies in this study (baby 10) was noted to have a diaper rash, as well as oral thrush, both of which are commonly caused by a fungus (Candida), and which were treated with an antifungal agent (nystatin). The qPCR analysis detected especially high levels of fungal rDNA in stool samples from this baby, particularly during the period in which these findings were described. This baby's mother also had notably high levels of fungal SSU rDNA sequences in her prenatal vaginal swab sample, but not in her “day 0” stool sample.
The prevalence of archaea was considerably lower and more variable than that of fungi or bacteria; qPCR analysis detected archaeal rRNA genes (in the range of 103–106 rRNA genes/g) in only seven babies during their first year of life, and in four of these babies, they were detected in only a single sample. In these babies, archaea appeared only transiently, and almost exclusively in the first few weeks of life; they were detected in only one infant after the fifth week of life. Limited analysis of archaeal sequences amplified from the three maternal stool samples that tested positive for archaea (mothers 4, 9, and 12) revealed a predominance of Methanobrevibacter smithii (7/8 archaeal clones identified, including at least one clone from each mother), with one additional (uncultured) archaeal phylotype. Results of qPCR analysis of fungi and archaea are included in Dataset S5 and shown graphically with bacterial qPCR results in Figure S2.
The microbial colonization of the infant GI tract is a remarkable episode in the human lifecycle. Every time a human baby is born, a rich and dynamic ecosystem develops from a sterile environment. Within days, the microbial immigrants establish a thriving community whose population soon outnumbers that of the baby's own cells. The evolutionarily ancient symbiosis between the human GI tract and its resident microbiota undoubtedly involves diverse reciprocal interactions between the microbiota and the host, with important consequences for human health and physiology. These interactions can have beneficial nutritional, immunological, and developmental effects, or pathogenic effects for the host [2,5,7,18,45].
This study began with the development of a DNA microarray with nearly comprehensive coverage of the bacterial taxa represented in the available database of SSU rRNA gene sequences. Our microarray design and experimental methods were based on lessons learned in the validation of a less comprehensive SSU rDNA microarray [46]. These previous experiments enabled us to optimize our methods for computational prediction of SSU rDNA hybridization behaviors, and to develop an experimental protocol that maximized hybridization specificity. The excellent concordance in the measurements of individual taxa determined using the new microarray design in comparison with sequencing results from corresponding SSU rDNA clone libraries (Figure 1) suggests that these design principles hold true for this platform across a diversity of taxa and give us confidence in both the comprehensiveness and accuracy of the results obtained with our new microarray probe set. It is important to note, however, that our methods of array design and analysis are imperfect and still evolving. Several of the unexpected species predicted by the microarray to be present in one or more samples could not be corroborated by sequencing. In most of these cases, sequence analysis of the sample(s) in question revealed that low-level cross hybridization of a highly abundant species was responsible for the false-positive prediction, a result that will be taken into consideration in future rounds of array design and analysis.
We used this microarray in a detailed, systematic, and quantitative study of bacterial colonization of the newborn human GI tract. We used freshly collected stool samples as surrogates for samples taken from the lumen and mucosa of the colon. Although there are undoubtedly differences in the population profiles of stool samples and corresponding mucosa, we found in a previous study that the profiles are nonetheless remarkably consistent—sufficiently so that individual stool samples can readily be matched to colonic biopsy samples from the same individual, based on the similarity in their bacterial profiles [15,46]. Thus, we believe that the results of our temporal analysis of the bacterial populations in infant stool samples provide a useful window on the resident colonic microbiota.
In view of the importance of the symbiosis between human host and gut commensals for both human host and microbial colonist, it would be easy to imagine that the program of microbial colonization of the neonatal GI tract would have evolved under strong selective pressure, acting on both the intestinal niche and its microbial colonists, to be highly deterministic and stereotyped. We might have expected that a highly restricted group of co-evolved commensals would be exceptionally well adapted to this environment and consistently dominate the colonization process in a stereotyped fashion. Indeed, the bacteria that we found in infant and adult feces, presumably reflecting the colonic microbiota, were largely restricted to only a small subset of the bacterial world—Proteobacteria, Bacteroides, Firmicutes, Actinobacteria, and Verrucomicrobia. Yet, surprisingly, we found that in the first days to months of life, the microbiota of the infant gut, and the temporal pattern in which it evolves, is remarkably variable from individual to individual. The seemingly chaotic progression of the early events in colonization, and the similarity in bacterial composition of some early infant samples to breast milk or vaginal swabs, suggests that the bacterial population that develops in the initial stages is to a significant extent determined by the specific bacteria to which a baby happens to be exposed. Notably, these maternal “signatures” did not persist indefinitely, as evidenced by our failure to find a significantly higher correlation of the overall taxonomic profiles of baby/parent pairs from the same household versus different households.
An important exception to the tale of individuality and uniqueness in the early profiles was the remarkable similarity of the temporal profiles of the fraternal twins (babies 13 and 14) (Figures 4 and 5). These twins shared both a common environment and approximately 50% genetic identity, making it impossible to determine from this study to what degree each of these commonalities is responsible for their similar colonization patterns. However, evidence from this and other studies suggests that the shared environment is a major factor. One argument in favor of this view is the lack of comparable similarity in the microbial communities of other pairs that also share 50% genetic identity, including mother:baby, father:baby, and sibling:baby (unpublished data), although this dissimilarity may be due in part to their differing stages in development. Another argument in favor of a strong environmental influence is the coincidental transient appearance of specific organisms in both twins—it is hard to imagine that the appearance of a particular microbe on a particular day could be genetically programmed. Our final argument rests on evidence from a previous study that the microbiota of genetically equivalent families from a cross of inbred mice was more similar among members of the same “household” (mother and offspring who share a cage) than between households [1].
The definition of a “healthy” intestinal microbiota encompasses a remarkable diversity of community profiles in the first 6 mo of life. Although diverse and idiosyncratic in the early months, these microbial communities became progressively more similar to one another (Figure 6A), converging toward a generic adult-like profile (Figure 6B) characterized by a preponderance of Bacteroides and Firmicutes, common occurrence of Verrucomicrobia, and very low abundance of Proteobacteria and aerobic Gram-negative bacteria in general. We hypothesize that the earliest colonization events are determined to a large extent by opportunistic colonization by bacteria to which a baby is exposed in its environment. Common environmental exposures are likely to include the maternal vaginal, fecal, or skin microbiota, as is suggested by the observed similarity of some infants' early stool microbiota to these samples, which is consistent with previous evidence of vertical transmission of microbes [33,47,48]. The diversity and variation would thus reflect the corresponding individuality of these accidental exposures. Over time, however, the fitness advantage of the taxa that typically dominate the adult colonic microbiota apparently overcomes the initial advantage of early-colonizing opportunists that are less well adapted to the intestinal environment. In addition, progressive changes in the gut environment, due to intrinsic developmental changes in the gut mucosa, transition to an “adult” diet, and the effects of the microbiota itself [44,49–51], may impose increasingly stringent selection for the most highly adapted bacteria. Thus, despite the unexpectedly chaotic early months, the establishment of the gut ecosystem in human infants turns out after all to follow a conserved, conventional program.
The transformation of the intestinal microbiota to an adult-like pattern implicitly involved replacement of species found in infants, but rarely in adults, with species characteristically found in the adult colon. One potential driving force for such a demographic change might be that the adult-like community members eventually dominate by virtue of their greater ability to establish themselves stably and irreversibly once they colonize a host. We looked for evidence of this differential “stickiness” by comparing the autocorrelations over time of the abundance of each “species” (see Materials and Methods). We found no clear evidence that the species characteristic of adult microbiota were able to establish more intrinsically stable colonization than the species characteristic of infant microbiota.
We and others have found that the individual-specific characteristics of the bacterial microbiota of adults are stable, in the sense that they remain consistently more similar within an individual over time than between individuals, for periods of a year or more, and one of the striking results of this study was the identification of relatively stable, individual-specific patterns of bacterial colonization even in the first weeks and months of life. These observations raised the interesting possibility that opportunistic colonization events in early infancy might play a significant role in defining the distinct characteristics of the same individual's microbiota into adulthood. We looked for evidence of this by comparing the intraindividual and interindividual correlations of bacterial profiles at 1 or 2 mo and 1 y, and found no significant difference (unpublished data). Thus, although these results certainly do not exclude the possibility that early colonization events play an important role in determining the adult microbiota, there does not appear to be a strong, direct correlation between the two.
Our results and conclusions differ considerably from many previous reports in several respects. One notable discrepancy between our studies and many others was the relatively low frequency and abundance of Bifidobacteria in the fecal microbiota at any age from birth to adulthood. Bifidobacteria have received disproportionate attention, in part because of their reputed beneficial effects, and many studies have reported (and reviews have repeated) that the microbiota of breast-fed infants is dominated by Bifidobacteria [17–19]. We were thus surprised by, and initially skeptical of, the apparent paucity of Bifidobacteria in nearly all of our samples, and took steps to verify that our results were accurate. Bifidobacteria-specific qPCR corroborated the conclusion from our microarray results that Bifidobacteria were rarely major constituents of the GI microbiota, at least in this study population, and that in most babies, they did not appear until several months after birth, and thereafter persisted as a minority population. Although it is conceivable that there are geographical or demographic differences in the prevalence of Bifidobacteria, we suspect that the emphasis on Bifidobacteria in studies and reviews of the infant GI microbiota may be out of proportion to its prevalence, abundance, and relevance to health.
The results presented here suggest numerous future avenues of research. An intriguing feature of the bacterial population dynamics was the occurrence of abrupt shifts punctuating intervals of relative stability. Except in one case (the antibiotic treatment of baby 8), we could not readily identify a strong candidate for the cause of the shifts we observed. Some possibilities include bacteriophage outbreaks that can selectively decimate a dominant taxonomic group [52]; stochastic, opportunistic invasion of a metabolic or anatomic niche by a fitter species; and subtle developmental or diet-induced changes in the gut environment tipping the fitness balance in the population. Other important avenues for future research will be comparing the composition and evolution of microbial communities encountered in these healthy babies to those of preterm or otherwise unhealthy babies and to investigate the effects of antibiotics, diet, and mode of delivery on the development and evolution of these communities. Even though the healthy babies in this study assumed a large range of microbial community profiles, they were similar in several respects, most notably in the major contributing phyla, in the acquisition of certain key phyla over time, and in the relative stability of their profiles over time. It may be that in other states of health or disease, we will find either species or groups that are novel to this environment, or unusual combinations of this newly defined set of “usual” species.
Importantly, although we have shown that the gut microbiota becomes increasingly stereotyped over the first year, it is clearly established that stable interindividual differences persist even in adults [15,16]. When and how do these stable “intrinsic” characteristics of the microbiota of each individual develop? How long do they persist? How do the differing stabilities of colonization by different bacteria relate to their microanatomic (e.g., crypt vs. villus vs. mucous layer) or metabolic niches? Identifying the environmental and genetic factors that determine the distinctive characteristics of each individual's microbiota, and determining whether and how these individual specific features affect the host's physiology and health, will be an important goal for future investigations, in which the microarray described in this study will be a useful tool.
The microarray contained 10,500 DNA probes (10,265 unique sequences). The probes comprised 1,379 control probes (1,144 unique sequences) and 9,121 unique taxonomically specific probes (5,938 group-level and 3183 species-level probes), consisting of 40-nucleotide (nt) sequences derived from the SSU rRNA genes, and selected for their specificity to the corresponding species or taxonomic group. The basic principles of the design are detailed in a previous report [46]. The design was based on the 2004 prokMSA SSU rDNA sequence database and phylogenetic tree [38], containing 86,453 prokaryotic SSU rDNA sequences (5,672 archaeal and 80,781 bacterial) organized into 672 archaeal and 15,765 bacterial OTUs. OTUs were defined in prokMSA as groups of sequences with identity scores (as defined in [38]) greater than either 95% or 98% (in certain medically relevant genera). We distilled this database by selecting a single high-quality sequence representative for each OTU, and trimmed the sequences to within the regions amplified by universal bacterial (Bact-8F: 5′-AGAGTTTGATCCTGGCTCAG-3′ [53] + 1391R: 5′-GACGGGCGGTGTGTRCA-3′ [54]) or archaeal (Arch344 [55]/1391R) primers. The OTUs in our pruned prokMSA database were organized into 945 (53 archaeal + 892 bacterial) taxonomic groups (nodes) containing multiple OTUs, and 92 single OTU “nodes.”
The nomenclature of the database was such that each OTU was designated by a numerical code that indicated its prokMSA taxonomy, e.g., species “1.2.3.4.5.6.007” belongs to superkingdom 1, phylum 1.2, class 1.2.3, order 1.2.3.4, family 1.2.3.4.5, genus 1.2.3.4.5.6. In this manuscript, taxonomic levels are referred to according to their depth in the OTU code, e.g., species “1.2.3.4.5.6” belongs to the more-inclusive group “1” at level 1 and less-inclusive group “1.2.3.4: at level 4.
We generated a large set of candidate probes by using BLAST [56] to predict the potential for hybridization of overlapping 40-nt sequences from each OTU in our distilled sequence database with each of the other OTUs in the database (by tiling across each sequence with window of two). A sequence was deemed a candidate probe for a specific taxonomic group if it was predicted to hybridize to at least 10% of that group's members, and not to any non-group members, using an empirically determined threshold of 28 out of 40 BLAST match–mismatches (total nucleotide matches minus mismatches to the best BLAST hit for a given rDNA sequence) as our cutoff for potential hybridization [46]. From the resulting set of candidate probes, we selected, for each node/taxonomic group, the two probes predicted to hybridize to the largest fraction of that group. We also selected probes from our candidate set such that each OTU in our distilled database was represented by probes at as many taxonomic levels as possible. Due to space limitations on the array, we were unable to print species-specific probes for every OTU in our database. Instead, we supplemented the set of group-level probes designed as described above with three additional sets of probes. First, we compiled a list of bacterial species that were either medically relevant or known human commensals. We identified the prokMSA codes for each of these species and tested the selected sequence from that OTU for species-specific 40-nt sequences, as defined by a BLAST match–mismatch score no greater than 27 to any other sequence. We also evaluated the species-specific probes from our previous array design [46] in the context of the prokMSA database, and included 467 such species-specific probes, representing 286 species. The final category of taxonomic probes were the “novel OTU” probes—316 probes designed to represent recently discovered SSU rDNA “species” that were identified in studies of the adult human colon [15] or stomach [57] (novel OTUs were defined by a 99% identity cutoff as described in the original studies). Finally, our microarray design included 1,153 control probes—positive and negative—designed for normalization and systematic examination of hybridization behaviors. The negative controls included both non-rDNA sequences and reverse complement rDNA sequences, whereas the positive controls included primer sequences and several sets of overlapping probes covering complete bacterial, archaeal, and eukaryotic SSU rRNA genes. Surface-attached oligonucleotide probes were synthesized in situ as previously described [58], with a 10-nt poly-T linker used to tether the specific 40-nt DNA probe (Agilent Technologies, http://www.chem.agilent.com). All arrays also included 307 standard Agilent control probes. All probe sequences and annotations are available in Dataset S6.
Our microarray probe set included one or more group-specific probes for 649 of the 950 taxonomic groups in prokMSA and species-specific probes for 1,590 bacterial and 39 archaeal species. Taken together, these probes ensured that 15,406 (94%) of the 16,437 species represented in the prokMSA database had at least one representative probe at some level in the tree from phylum to species, and most prokMSA species (74%) had representative probes at multiple taxonomic levels (mean of 2.4 levels per species).
Thirteen healthy pregnant women were recruited at the Stanford University Medical Center. All study participants, including 14 babies (one set of fraternal twins), nine fathers, 13 mothers, and two siblings (1–2 y old) provided informed consent or were consented for by their parents. The study design was approved by the Stanford University Administrative Panel on Human Subjects in Medical Research. At 36–40 wk of gestation, vaginal swabs (Copan Diagnostics, http://www.copanusa.com) were obtained from ten of the 13 mothers and immediately frozen at −20 °C. After birth, infant stool samples were obtained by the parents using stool collection vials (Sarstedt, http://www.sarstedt.com/php/main.php), which contained a spoon for standardized collection of approximately 300 mg of material. Infant stool samples were collected according to the prescribed schedule (Table 1) and immediately stored in home freezers. A maternal “day 0” stool sample was obtained within 0–5 days following delivery. Stool samples were transported on ice to the laboratory for processing 2 wk, 3 mo, and 6 mo after birth of the baby. Twelve of the mothers provided breast milk samples (~20 ml) 3–9 mo after delivery, and one of them also provided breast milk 6 d after delivery; these samples were collected in 50-ml tubes and frozen immediately. Nine of the study families also provided contemporaneous stool samples from mother, father, and baby 12–17 mo after the baby's birth. Upon arrival in the laboratory, all samples were immediately transferred to a −80 °C freezer, and stored there until processing. A total of 548 samples were collected, including 471 stool samples from babies, 39 stool samples from mothers, nine stool samples from fathers, two stool samples from siblings, 16 breast milk samples, and 11 vaginal swabs. Parents were instructed to keep a journal recording key events in the categories of illness, medication, dietary change, and travel. Table 2 contains selected information for each baby (e.g., gender and method of delivery).
DNA was extracted from stool samples using the QIAamp Stool DNA mini kit (Qiagen, http://www1.qiagen.com). Vaginal swabs were processed using the QIAamp DNA mini kit (Qiagen). Milk samples were first concentrated by spinning 2 ml in a microcentrifuge for 10 min at 5,000 g, removing 1,800 μl of the supernatant. The pellet was resuspended in 200 μl of the remaining supernatant, and DNA was extracted using the QIAamp DNA mini kit. Samples were processed in batches of approximately 16, and multiple extraction controls were included in each run to monitor contamination. The ratio of samples to extraction controls was 6.8 for stool, 2.8 for vaginal swabs, and 5.3 for milk.
SSU rDNA was amplified from extracted DNA using broad-range bacterial-specific primers Bact-8F (5′-AGAGTTTGATCCTGGCTCAG-3′) [53] and T7-1391R (5′-AATTCTAATACGACTCACTATAGGGAGACGGGCGGTGTGTRCA-3′) [46,54]. These primers amplify approximately 90% or more of the full-length bacterial SSU rRNA coding sequence, and provide a promoter for T7 RNA polymerase. PCR mixtures were composed of 1× PCR buffer II (Applied Biosystems, http://www.appliedbiosystems.com), 1.5 mM MgCl2, 0.05% Triton X-100, 20 mM tetramethylammonium chloride, 2% dimethyl sulfoxide, 0.1 mM concentrations of each deoxyribonucleoside triphosphate, 0.4 μM concentrations of each primer, 2.5 U of AmpliTaq DNA polymerase (Applied Biosystems), and 5 μl of extracted DNA in a final volume of 50 μl. The PCR conditions used were 5 min at 95 °C, 20 cycles of 30 s at 94 °C, 30 s at 55 °C, and 90 s at 72 °C, followed by 8 min at 72 °C. Amplification was carried out by using a GeneAmp PCR system 9700 (Applied Biosystems). In cases of extremely low yield, multiple 20-cycle reactions were pooled. PCR reactions were cleaned up in 96-well format using Invitrogen's Charge Switch PCR Purification bead-based system (Invitrogen, http://www.invitrogen.com), and stored at −20 °C.
Our common reference was an equimolar mix of SSU rDNA amplicons from each sample (infant or maternal stool, vaginal, or breast milk) collected before the subject infant was 1 y old. To create the equimolar mix, purified 20-cycle PCR products were quantitated in 96-well format using the Quant-It PicoGreen dsDNA kit (Molecular Probes, http://probes.invitrogen.com) and pooled in equal amounts. The approximate fractions of stool-, vaginal-, and milk-derived SSU rDNA in the resulting pool were 90%, 5%, and 5%, respectively. This DNA pool was used both as a template for in vitro transcription (for microarray hybridizations) and for construction of a SSU rDNA clone library.
The reference pool (an equimolar mix of SSU rDNA amplicons from most samples, described above) was cloned and sequenced as previously described [15]. We obtained 3,458 high-quality bacterial rDNA sequences of length greater than 800 nt, including both 3,163 double reads and 295 single reads. These sequences were taxonomically classified using the Ribosomal Database Project (RDP) classifier (summarized in Table 3).
We also cloned and sequenced several hundred SSU rDNA amplicons (mean = 342; range = 125–557 adequate sequences) from each of 12 diverse individual samples in the same way, yielding a total of 4,100 high-quality sequences of length greater than 800 nt. The 12 samples sequenced consisted of ten stool samples (day 11 from baby 2; day 1 from mother of baby 3; week 12 from baby 3; month 6+ from baby 8; day 1 and day 2B from baby 10; day 12 from baby 12; month 7 from baby 13; month 7 from mother of twins: babies 13 and 14; and month 7 from baby 14), one milk sample from the mother of twins 13 and 14, and one vaginal sample from the mother of twins 13 and 14.
Each sequence from these 12 individual samples was taxonomically classified according to the 2004 prokMSA taxonomy [38] using BLAST. Specifically, we used BLAST to find the sequence with the most matches in the entire prokMSA database (also trimmed to within 8F and 1391R). The top two hits were reported and compared (hits with fewer than 600 matched nucleotides were not considered), and if these two hits mapped to the same OTU, then the sequence was classified to that OTU. If the top two hits differed in their taxonomic code, then the sequence was classified only at the most-specific level shared by the top two hits. In cases in which the second-best hit was considerably worse (matches 2nd/matches 1st < 0.9), only the best hit was considered. The prokMSA OTU codes explicitly define the taxonomic classification of a sequence at all phylogenetic levels for all of the 4,100 high-quality “individual sample” sequences.
Purified SSU rDNA amplicons were used as a template for in vitro transcription-based synthesis of amino-allyl–labeled single-stranded RNA using the MEGAScript T7 In Vitro Transcription Kit (Ambion, http://www.ambion.com). Transcription reactions were cleaned up in 96-well format using Ambion's MagMax RNA Purification system and stored at −20 °C. Large batches (5–10 μg) of reference RNA (equimolar pool of all samples; see below) were labeled using Cy3 NHS ester and stored for several weeks for repeated use. On the day of hybridization, 1 μg of each sample RNA was labeled using Cy5 NHS ester as described previously [46]. We then combined 100 ng of Cy5-labeled sample and 100 ng of Cy3-labeled reference pool in a volume of 48 μl (in nuclease-free water). We then added 2 μl of Agilent's 25× fragmentation buffer, and fragmented the RNA by heating at 70 °C for 30 min before stopping the reaction by putting it on ice and adding 50 μl of Agilent's 2× hybridization buffer. Immediately before hybridization, we heated the reaction to 95 °C for 5 min, then cooled it on ice before adding 120 μl of 1× hybridization buffer to 100 μl of fragmented, labeled RNA, and loaded 200 μl of this mixture into a hybridization chamber (Agilent). Arrays were hybridized at 60 °C in a rotating oven for 14–18 h. Slides were washed in 6× SSC, 0.005% TritonX-102 for 5 min at room temperature, then in 0.1× SSC, 0.005% TritonX-102 for 5 min, and scanned immediately using an Agilent DNA Microarray Scanner. Washing and scanning were performed in a low-ozone environment [59].
Data were extracted from the scanned microarray image using the most current version of the Agilent Feature Extraction software (Versions 7.1.1–8.1.1). The raw background-subtracted Cy5 (or Cy3) fluorescence intensity values for each probe were normalized by dividing the Cy5 (or Cy3) values by the median Cy5 (or Cy3) value of the universal probe “UNIV2” (extended version of 3′ PCR primer 1391R: 5′-GTGGGGAGCGAACAGGATTAGATACCCTGGTAGTCCACGC-3′) from the corresponding array, and multiplying by 100. At this stage, values ranged from 0.01 to approximately 100; values greater than 100 were rare, occurring only when the fluorescent signal for a specific probe was brighter than that of the universal probe. The normalized Cy5 values were “decompressed” by the following correction: decompressed Cy5 equals 10 to the power of log5 of the normalized Cy5 intensity. This decompression corrects for the nonlinear relationship of the hybridization signals to the relative abundance of the target species, which we observed in a series of serial dilution experiments described in a previous report [46]. In those experiments, we found that a 10-fold change in abundance translated into approximately a 5-fold change in the corresponding fluorescence intensity. Following this transformation, the expanded range of values was 0.001 to approximately 700. We used BLAST to predict the hybridization of each of the 3,458 common reference rDNA sequences greater than 800 base pairs, using a weighting scheme described previously [46], such that a probe with a perfect match to every sequence would have an expectation of 100%, and probes with fewer or imperfect matches to the sequences in the reference pool would have correspondingly lower expected hybridization values. We then calculated the log (observed/expected) ratio for each probe in the context of our pooled sample reference mix and applied these probe-specific correction factors to the decompressed Cy5 values.
Microarray analysis of synthetic pools of defined composition allowed us to identify poorly performing probes, which were then excluded from further analyses. We evaluated probe performance using five synthetic pools (four pools of six unique rDNA PCR products, one pool of 230 unique rDNA PCR products [from [46]]) and one biological pool (the common reference described above), whose composition was characterized by deep sequencing. Probes with fluorescence intensity values greater than 0.5% of that of the universal probes despite the lack of predicted sequence homology (defined as a BLAST match–mismatch value less than 25 out of a possible 40) to any of the species in the sample were excluded from subsequent analyses. We also discarded data from probes that had an observed (decompressed) signal intensity 100-fold higher or lower than expected, in the biological reference pool analyses, (as described above in the Microarray data normalization section). The remaining set of 6,381 well-behaved probes was used in all subsequent analyses.
For each sample, we derived an estimate of the relative abundance of each taxonomic group in our phylogenetic tree using an algorithm that ensures that no species contributes more than once to the estimate of a taxonomic group abundance, and that the downstream probes (probes that represent distinct subsets of species belonging to that phylogenetic group) are incorporated into the cumulative group abundance estimate. Specifically, for each phylogenetic group in each sample, we sorted all of the downstream probes according to their microarray-based relative abundance estimates and calculated the total sum of the relative abundance estimates of all nonoverlapping probes. As a result, the specific probes added together to represent a given taxonomic group varied across samples, depending on which specific probes had the greatest hybridization signal in that sample.
Estimates of the relative abundance of each prokMSA group, at each level of the taxonomic hierarchy, in the reference pool and in the 12 individual samples analyzed by sequencing, were derived by calculating the proportion of the sequences in the corresponding rDNA library that were assigned by BLAST to that group. These sequence-based abundance estimates were directly compared to those derived from the microarray data as described in the previous section.
We used autocorrelations as a way to measure the tendency of a taxonomic group to persist once established (“stickiness”). For a given time series from time a to time b, we calculated the Pearson correlation for each baby of the vector (a + 1,…, b) and vector (a,…, b − 1) representing the log(relative abundance) estimates. The autocorrelation of each taxonomic group was then taken to be the median autocorrelation across all of the babies for which the taxonomic group in question was present at least once (defined as abundance >0.1%) in the time interval in question. In our “stickiness” analysis, we performed this analysis separately for two different sets of samples, collected at different sampling intervals, in order to avoid the confounding effects of variation in sampling intervals: (1) once weekly samples from weeks 1–12; and (2) monthly samples from months 1–6.
To screen for archaeal and fungal rDNA sequences. we first screened pools comprising all of the extracted DNA samples from each baby (i.e., one pool per baby), all parental stool samples, all vaginal samples, and all milk samples, respectively. For each pool that gave a positive result, all the component samples were then analyzed individually. To screen for archaeal rDNA sequences, we used two sets of broad-range archaeal-specific primers: A751F and U1406R [60]; and Arch333F (5′-TCCAGGCCCTACGGG-3′ [61]) and Arch958R (5′-YCCGGCGTTGAMTCCAATT-3′ [62,63]). To screen for fungi, we used the broad-range fungal-specific primers 817F [64] and 1536R-rev (5′-AATRCAATGCTCTATCCCCA-3′, adapted from [64]). PCR mixtures were similar to those used for the broad-range bacterial PCR described above, but with the following changes: for the fungal-specific PCR, the MgCl2 concentration was increased to 2 mM, and BSA was added in a final concentration of 1 mg/ml. To both fungal-specific and archaeal-specific PCR reactions, no tetramethylammonium chloride was added, and 2 μl of pooled DNA was added, with a final volume of 50 μl. The cycling program consisted of 5 min at 95 °C, 35 cycles of (30 sec at 94 °C, 30 sec at 55 °C, and 30 sec at 72 °C), followed by 8 min at 72 °C. Amplification reactions were analyzed on agarose gels.
To investigate whether mismatches to the broad-range bacterial primers could affect amplification, DNA was extracted from five bacterial reference strains: Escherichia coli (TOP10 cells, Invitrogen), Clostridium perfringens (ATCC 13124), Bacteroides fragilis (ATCC 25285), Bifidobacterium longum (ATCC 15707), and Bifidobacterium infantis (ATCC 15697), using the QIAamp DNA mini kit. DNA concentrations were measured using a UV spectrophotometer, and adjusted to correct for DNA yield, genome size, and number of SSU rRNA gene copies per genome to obtain “normalized” DNAs, each containing the same number of SSU rRNA gene copies per μl. Bacterial DNA from lysates was amplified individually or in pairs, using primers 8F and T7-1391R as described above, using either 20 or 35 cycles. For the paired reactions, normalized Bifidobacterium longum DNA was mixed with undiluted or serial dilutions of normalized DNA from Escherichia coli, Clostridium perfringens, or Bacteroides fragilis. After amplification, PCR mixtures were purified using the QIAquick PCR purification kit (Qiagen) and digested using a set of restriction enzymes selected to distinguish between the two PCR products obtained with the paired species. Digestions were analyzed on agarose gels, to quantitate the relative abundance of the PCR products representing Bifidobacterium longum and the species in the comparison group, respectively.
A separate real-time qPCR assay was used to amplify and quantify rDNA from each of four microbial groups: total bacteria, Bifidobacteria, total fungi, and total archaea. Total bacterial qPCR was performed using a 10:1 mixture of the universal forward primer 8FM (5′-AGAGTTTGATCMTGGCTCAG-3′, adapted from [53]) and corresponding Bifidobacterium longum forward primer 8FB (5′-AGGGTTCGATTCTGGCTCAG-3′, this study), with reverse primer Bact515R (5′-TTACCGCGGCKGCTGGCAC-3′, adapted from [54]) and TaqMan probe Bact338K (5′-FAM/CCAKACTCCTACGGGAGGCAGCAG/TAMRA-3′, adapted from [65]). We supplemented the universal bacterial forward primer with the Bifidobacterium longum forward primer because an analysis of SSU rDNA sequences showed that this group was an outlier in that it had three mismatches to our forward primer 8FM. Pilot studies showed that this primer mixture allowed comparable amplification of SSU rRNA genes from representative Bifidobacteria, Bacteroides, Enterobacteria, and Clostridia. Bifidobacterium genus qPCR was performed using primers Bif42F [26] and Bif164R (5′-CATCCGGCATTACCACCCGTT-3′, adapted from [66]), and probe Bif126_Taqman [26]. Total fungal qPCR was performed using primers ITS1F-F (5′-CTTGGTCATTTAGAGGAAGTAA-3′ [67]) and ITS4-R (5′-TCCTCCGCTTATTGATATGC-3′ [68], and TaqMan probe 5.8S (5′-FAM/CATTTCGCTGCGTTCTTCATCGATG/TAMRA-3′, adapted from [68]. Archaeal qPCR was performed using primers Arch333F (5′-TCCAGGCCCTACGGG-3′) [61] and Arch958R (5′-YCCGGCGTTGAMTCCAATT-3′) [63], and TaqMan probe 515F (5′-FAM/GTGCCAGCMGCCGCGGTAA/TAMRA-3′, adapted from [54]). For all qPCR assays, each 20-μl reaction contained 1× TaqMan Universal PCR master mix (Applied Biosystems), 0.9 μM of each primer (0.09 μM of primer 8FB in the total bacterial assay), 0.2 μM of the probe, 1 U of AmpliTaq Gold DNA polymerase (Applied Biosystems), and 2 μl of extracted DNA. The thermal cycling program consisted of 95 °C for 10 min, followed by either 40 cycles (bacterial and bifidobacterial assays) or 45 cycles (fungal and archaeal assays) of 95 °C for 30 s, 55 °C for 30 s, 60 °C for 45 s, 65 °C for 15 s, and 72 °C for 15 s [69]. Reactions were carried out in a Prism 7900HT Sequence Detection System (Applied Biosystems). Ten-fold serial dilutions of known quantities of rDNA from the appropriate microbial group (i.e., bacteria, Bifidobacteria, fungi, or archaea) were used to generate standard curves. Absolute rDNA abundance was calculated based on the standard curves using SDS software version 2.1 (Applied Biosystems) with the baseline set at cycles 3–15 (bacterial and bifidobacterial assays) or cycles 3–13 (fungal and archaeal assays), and the cycle threshold set within the geometric phase of the amplification curve. Sensitivity of each assay was approximately 100 rDNA molecules per PCR reaction well. Every qPCR reaction plate included two types of negative controls (reagent control and aliquot control), each in triplicate. Specificity of the bifidobacterial real-time PCR assay was tested using genomic DNA extracted from 17 bacterial reference strains: Bacillus subtilis (ATCC 6633), Bacteroides fragilis (ATCC 25285), Bacteroides thetaiotaomicron (ATCC 29148), Bifidobacterium longum (ATCC 15707), Bifidobacterium infantis (ATCC 15697), Clostridium perfringens (ATCC 13124), Clostridium putrefaciens (ATCC 25786), Enterococcus faecalis (ATCC 19433), Escherichia coli (TOP10 cells; Invitrogen), Haemophilus haemolyticus (ATCC 33390), Lactobacillus acidophilus (ATCC 4356), Lactobacillus delbrueckii (ATCC 4797), Megasphaera elsdenii (ATCC 17752), Proteus vulgaris (ATCC 13315), Pseudomonas aeruginosa (ATCC 10145), Staphylococcus aureus (ATCC 25923), and Streptococcus salivarius (ATCC 13419).
We investigated the origin of array hybridization signals representing 12 species/taxa whose presence in the samples was unexpected and uncorroborated by sequences in the reference pool. For each analysis, we used one or both of two independent assays. First, we attempted to amplify the sequences apparently detected by the array analysis, using species/taxa specific primers; when a product was obtained, it was further analyzed by sequencing. To amplify the sequences, we used a primer identical to the 40-mer probe that yielded a hybridization signal in our microarray as the 5′ primer and a 40-mer universal sequence (the reverse complement of sequence UNIV2 given above) as the 3′ primer. In some cases, we also tried to amplify the suspect sequence using a truncated (23-mer) version of the corresponding oligonucleotide probe from the microarray paired with a known group-specific PCR primer from ProbeBase [55]. All PCRs were performed under conditions identical to those used in the original amplifications of samples for microarray analysis. Positive bands of expected size were cloned and sequenced. As a second approach, four samples predicted from the microarray results to contain sequences from unexpected species were further analyzed by sequencing of SSU rDNA clone libraries (96–288 clones), generated by amplification using broad-range bacterial primers 8F and 1391R (as above), and cloning and sequencing as previously described [15]. The relative abundances predicted by microarray analysis and the numbers of clones sequenced were as follows: Vibrio (13%): 96, Deinococcus (0.1%): 192, Spirochaetes (1%): 288, and Legionella pneumophila (1%): 192.
|
10.1371/journal.pntd.0003004 | Increased Interleukin-17 in Peripheral Blood and Cerebrospinal Fluid of Neurosyphilis Patients | Treponema pallidum infection evokes vigorous immune responses, resulting in tissue damage. Several studies have demonstrated that IL-17 may be involved in the pathogenesis of syphilis. However, the role of Th17 response in neurosyphilis remains unclear.
In this study, Th17 in peripheral blood from 103 neurosyphilis patients, 69 syphilis patients without neurological involvement, and 70 healthy donors were analyzed by flow cytometry. The level of IL-17 in cerebrospinal fluid (CSF) was quantified by ELISA. One-year follow up for 44 neurosyphilis patients was further monitored to investigate the role of Th17/IL-17 in neurosyphilis. We found that the frequency of Th17 cells was significantly increased in peripheral blood of patients with neurosyphilis, in comparison to healthy donors. IL-17 in CSF were detected from 55.3% neurosyphilis patients (in average of 2.29 (0–59.83) pg/ml), especially in those with symptomatic neurosyphilis (61.9%). CSF IL-17 was predominantly derived from Th17 cells in neurosyphilis patients. Levels of IL-17 in CSF of neurosyphilis patients were positively associated with total CSF protein levels and CSF VDRL (Venereal Disease Research Laboratory) titers. Notably, neurosyphilis patients with undetectable CSF IL-17 were more likely to confer to CSF VDRL negative after treatment.
These findings indicate that Th17 response may be involved in central nervous system damage and associated with clinical symptoms in neurosyphilis patients. Th17/IL-17 may be used as an alternative surrogate marker for assessing the efficacy of clinical treatment of neurosyphilis patients.
| Syphilis, caused by the bacterium Treponema pallidum, can progress to affect the central nervous system (CNS) and cause damage in the brain and spinal cord, which is called neurosyphilis. While many neurosyphilis patients may not have any symptoms, some patients develop severe symptoms which can be life-threatening. Th17 cells are a subset of CD4+ T helper cells producing the hallmark cytokine IL-17, which are essential for effective antimicrobial host defense and are also involved in tissue damage. In this study we conduct a comparative analysis of Th17/IL-17 in the peripheral blood and cerebrospinal fluid (CSF) of syphilis patients without neurological abnormalities, and neurosyphilis patients with or without symptoms. Th17 frequency in peripheral blood was significantly increased in neurosyphilis. CSF IL-17 was increased in neurosyphilis patients, especially in those with symptomatic neurosyphilis. Levels of CSF IL-17 in neurosyphilis patients were positively associated with CNS damage. Notably, neurosyphilis patients with undetectable CSF IL-17 had better outcome upon treatment. These findings indicate that the Th17 response may be involved in central nervous system damage, clinical symptoms and prognosis of treatment of neurosyphilis patients.
| Syphilis, a sexually transmitted multi-stage disease caused by the spirochete Treponema pallidum, remains to be a global public health problem with an estimated 12 million new cases annually [1]. In recent years, China has experienced a resurgence of syphilis cases, with the national incidence rate of 32.04 per 100,000 population and with 429,677 new cases reported in 2011 [2]. T. pallidum invades the human host through genital or oral mucosa, abraded skin, enters lymphatic system and bloodstream, and then disseminates to different organs. Without treatment, this spirochetal pathogen is able to survive in the human host for several decades, causing damage in multiple organs including nervous system (neurosyphilis) [3], [4].
Neurosyphilis is a frequent and protean clinical manifestations ranging from headache and oculopathy to more serious conditions such as cerebrovascular events, paretic and tabes dorsalis [5]. The mechanisms underlying the development of neurosyphilis remain poorly understood. T. pallidum can invade the CNS at any stage of infection and provokes robust cellular immune response [6]. In the non-human primate models, strong T helper (Th) 1-type immune response can contribute to the clearance of T. pallidum in CNS [6]. The immune response elicited during infection, although aimed to eliminate organisms, may also contribute to the pathogenesis. Cytokines produced by T lymphocytes are critical for regulation of both protective and pathogenic immune responses [7].
Th17 cells, with the hallmark of producing cytokine IL-17, were identified as a subset of CD4+ T helper cells. Emerging evidence has demonstrated that Th17 cells contribute to clearance of diverse organisms (Mycobacterium tuberculosis, Pneumocystis carinii, Candida albicans and Klebsiella pneumonia et al.) [8], [9], [10], [11]. On the other hand, Th17 also mediates strong immunopathology in chronic infection. Anti-IL-17 and anti-IL-17R treatments could prevent severe Borrelia-induced destructive arthritis [12]. Hence, Th17 response in infection may be involved in both protection and progression/chronic infection.
Previous studies reported an increase of IL-17 in secondary syphilitic lesion and peripheral blood [13], [14]. Recently, Pastuszczak et al. also showed elevated CSF IL-17 levels in early asymptomatic neurosyphilis [15], suggesting that IL-17 may be involved in local immune response to T. pallidum infection. In this study, we performed a comparative analysis of Th17/IL-17 in peripheral blood and CSF in asymptomatic and symptomatic neurosyphilis patients, and evaluated the relationship between CSF IL-17 level and the clinical outcomes. Our results suggested that Th17/IL-17 is a contributing factor to the immunopathology of neurosyphilis, and may be used to monitor the prognosis of treatments of syphilis infected patients.
This study was performed at the Shanghai Skin Disease Hospital between Aug. 2010 and Dec. 2012. The study was approved by the Ethics Committee of the Shanghai Skin Disease Hospital. Written informed consents were obtained from all participants. Patients were identified and referred for enrollment by dermatologists, neurologists, psychiatrists and ophthalmologists after careful examination and evaluation. Syphilis was diagnosed at each stage of infection by a combination of compatible history, clinical features and the results of nontreponemal and treponemal tests of serum and CSF samples. The exclusion criteria include positive HIV infection; prior history of syphilis infection, or history of syphilis treatment (except for 7 serofast patients); history of systemic inflammation, autoimmune disease, other underlying acute or chronic disease, receiving anti-inflammatory medications, immunocompromised conditions, or use of antibiotics or immunosuppressive medications in the last four weeks. 70 healthy donors, who visited Shanghai Skin Disease Hospital voluntarily for a medical check-up for the purpose of STD prevention, were recruited to the study. All healthy subjects were negative for HIV and serological tests for syphilis (i.e., both serum RPR and TPPA negative), and did not have any clinical symptoms consistent with T. pallidum infection.
In this study, three groups of patients were included: 1) neurosyphilis group (including 40 subjects with asymptomatic neurosyphilis, 4 with meningovasculitis, 39 with general paresis, 8 with tabes dorsalis, and 12 with ocular neurosyphilis); 2) non-neurosyphilis group with normal CSF WBC count, CSF protein concentration and CSF-VDRL negative (including 13 subjects with primary syphilis, 30 with secondary syphilis, 19 with latent syphilis, and 7 with serofast syphilis); 3) 70 healthy donors. In this study, the serofast state must be met the following three criteria: i) syphilitic patients, despite receiving recommended standard treatment (according to Chinese National STI Treatment Guidelines), whose nontreponemal test titers (RPR) persists positive for at least two years of follow-up evaluation. ii) patients who denied high risk sexual behaviors (re-infection) following treatments; and iii) patients with RPR titers declined fourfold within 6 months after therapy. Peripheral blood from healthy donors was used for peripheral blood mononuclear cells (PBMC) isolation and for measurement of the baseline of the levels of IL-17+ cells and the frequency of Th17 cells. Since it is difficult to collect CSF from healthy donors, we used a separate control group of 29 patients who underwent orthopaedic or stone surgery (gall stone, kidney stone) but were serum RPR and TPPA negative, whose CSF samples were collected prior to spinal anaesthesia. The baseline level of IL-17 in CSF was determined using samples from the control group. All groups were well matched in the categories of gender and age. Additional information on the patient groups were presented in Table 1. CSF samples were stored at −70°C and thawed once before analyses.
First, all neurosyphilis patients have positive serum RPR and TPPA tests. The diagnosis of confirmed neurosyphilis includes reactive CSF-VDRL (Venereal Disease Research Laboratory) and CSF-TPPA tests in the absence of substantial contamination of CSF with blood. Presumptive neurosyphilis was defined as a nonreactive CSF-VDRL test but reactive CSF-TPPA with either or both of the following: (i) CSF protein concentration >45 mg/dl or CSF white blood cell (WBC) counts≥8/µl in the absence of other known causes for the abnormalities; (ii) clinical neurological or psychiatric manifestations consistent with neurosyphilis without other known causes for such abnormalities [16], [17].
Neurosyphilis is categorized as: asymptomatic, meningovascular, paretic, ocular and tabetic neurosyphilis. Asymptomatic neurosyphilis is defined by the presence of CSF abnormalities consistent with neurosyphilis and the absence of neurological/psychiatrical symptoms or signs. Meningovasculitis is defined by clinical features of meningitis and magnetic resonance image (MRI) evidence of brain lesions and/or a stroke syndrome. General paresis is characterized by personality changes, dementia and psychiatric symptoms including mania or psychosis. Patients with sensory loss, ataxia, lancinating pains, bowel and bladder dysfunction were considered as Tabes dorsalis. Ocular neurosyphilis (those who had ocular signs or symptoms but with normal CSF index were not included in this study)is defined by the presence of CSF abnormalities consistent with neurosyphilis and ocular signs or symptoms (worsening visual acuity and visual fields, floaters, papillitis, uveitis). All these forms of neurosyphilis should have no other known causes for these clinical abnormalities. A complete list of information of neurosyphilis patients were listed in Table 2.
According to Chinese National STI treatment Guidelines, syphilis patients without CNS involvement were treated with benzathine penicillin 2.4MU/qw intramuscular for 1 or 2 weeks for early syphilis, and 3 weeks for late or unknown duration syphilis. If allergic to penicillin, ceftriaxone 0.25 g/day intramuscular for 10 days were given. Neurosyphilis patients were given aqueous crystalline pencillin G, 4MU intravenously every 4 h for 14 days, or ceftriaxone intravenously with 2 g daily for 10 days if allergic to penicillin. In the 103 neurosyphilis patients, 80 patients were treated with aqueous crystalline pencillin G, 4MU intravenously every 4 h for 14 days, 23 patients were treated with ceftriaxone intravenously with 2 g daily for 10 days.
All patients were asked for follow-up after treatment. Patients were selected if the patient's written informed consent was obtained. The exclusion criteria include positive HIV infection; history of syphilis or syphilis treatment; history of systemic inflammatory, autoimmune disease, other underlying acute or chronic disease, patients receiving anti-inflammatory medications, immunocompromised, or using antibiotics or immunosuppressive medications in the last four weeks.
In this study, 44 neurosyphilis patients were enrolled and followed up at Shanghai Skin Disease Hospital. Patients returned for follow-up visits at 3, 6, 9 and 12 months after treatment. All patients underwent lumbar puncture at the 3-month visit, and lumbar punctures were repeated at 6, 9 and 12 months if the previous CSF profile was abnormal. Blood samples were collected at each follow-up visits. All patients completed their 12 months follow-up visit.
For the CSF-VDRL and the serum RPR test, a 4-fold decrease or more in titer or reversion to a nonreactive result was defined as a normal response. Stepwise Cox regression models were used to determine the influence of the following factors on the likelihood of normalization of each measure and the improvement of clinical symptoms: (1) neurosyphilis treatment regimen (intravenous ceftriaxone, vs. intravenous aqueous penicillin G); (2) syphilis stage (secondary and early latent vs. late latent and syphilis of unknown duration); (3) baseline laboratory values (greater or less than the median value for those subjects with each abnormality); (4) CSF IL-17 levels: CSF IL-17 negative (<0.5 pg/ml) and CSF IL-17 positive (≥0.5 pg/ml); and (6) clinical symptoms.
Whole blood samples (5 ml) from syphilis patients and healthy donors were used for peripheral blood mononuclear cells (PBMC) isolation. PBMC were purified from peripheral blood by centrifugation using a Ficoll-Hypaque gradient (Axis-Shield). Because resting cells do not normally produce cytokines, cells were stimulated in vitro in order for the respective cytokine genes to be activated for intracellular cytokine staining. Phorbol myristate acetate (PMA) and ionomycin are unspecific stimulator that trigger a strong production of cytokines in vitro and are widely used to evaluate intracellular cytokine production from various T lymphocyte subpopulations [18]. Monensin is used to prevent the intracellular transport of cytokines from Golgi apparatus for enhancing the sensitivity of the detection. Accordingly, PBMC were seeded into 24-well culture plates (Corning) at 2×106 cells/well and stimulated ex vivo with PMA (50 ng/ml) (Sigma) and ionomycin (1 µg/ml) (Sigma) for 4 hours. Monensin (2 uM) (eBioscience) was then added at the start of stimulation. CSF (10 ml) was centrifuged at 4°C immediately after spinal tap, and cells were stimulated as described above.
For intracellular staining, cells were first stained with ECD-labeled anti-human CD3 (Clone UCHT1, Beckman), FITC-labeled anti-human CD4 (Clone RPA-T4, Biolegend) and then fixed and permeabilized using Perm/Fix solution (Biolegend) at room temperature for 20 minutes. Cells were washed with Perm/Wash buffer (Biolegend) and stained with PE-labeled anti-human IL-17A (Clone BL168, Biolegend). Mouse IgG1 and IgG2 (BD Biosciences) were used as isotype controls. After staining, cells were analyzed with Epics XL (Beckman Coulter) and FlowJo software (Tree Star). Lymphocytes were gated according to forward and side scatter characteristics and CD4+T cells were gated based on CD3 and CD4 expression. IL-17 positive lymphocytes, CD3+, CD4+ T cells were defined by setting regions with the lower limits for cytokine positivity determined from isotype antibody.
IL-17 levels in CSF were determined using human IL-17 Quantikine ELISA kits (eBioscience) according to manufacturer's instruction. The sensitivity for detecting IL-17 is 0.5 pg/ml.
Data were presented as median and range (min, max). Differences between the groups were analyzed using the nonparametric Mann-Whitney U test. The detection rates between the groups were assessed using χ2 test or Fisher's exact test. Spearman correlation analysis was performed between the levels of IL-17 and other parameters. Stepwise Cox regression models were used to determine the influence of the factors on the likelihood of normalization of each measure. All statistical analyses were performed using SPSS 17.0 software. A value of P<0.05 was considered significant.
To investigate the potential role of Th17 in neurosyphilis, we first examined the frequency of IL-17+ among lymphocytes, CD3+, CD4+ T populations in PBMC. The baseline frequency of total IL-17+ cells, and IL-17+ CD3+cells, and IL-17+ CD4+ T cells (Th17) of PBMC in healthy individuals were 0.86% (0.19%–1.58%), 1.33% (0.48%–3.2%) and 1.7% (0.56%–2.76%), respectively (Fig. 1A, 1B & 1C). We observed a significant increase in the frequencies of IL17+, CD3+IL-17+ and Th17 cells in syphilis patients with either non-neurosyphilis or neurosyphilis compared to those in healthy individuals (Fig. 1A, 1B & 1C). However, there was no significant difference in the frequencies of IL-17+ cells, CD3+IL-17+ and Th17 cells in PBMC between syphilis patients without neurological involvement (including primary, secondary, latent and serofast syphilis patients) and neurosyphilis patients (Fig. 1A, 1B & 1C).
To further investigate whether Th17 cells in peripheral blood are different between diverse clinical presentations of neurosyphilis, we divided neurosyphilis patients into two groups, asymptomatic (n = 40) and symptomatic neurosyphilis patients (n = 63). We then compared the Th17 cell frequency in PBMC between these two groups. As shown in Fig. 1D, 1E & 1F, patients with symptomatic neurosyphilis had significant higher percentage of total IL-17+cells, CD3+IL-17+ and Th17 in PBMC than the patient group with asymptomatic neurosyphilis.
Since neurosyphilis patients had increased levels of Th17 cells in peripheral blood, we further investigated the IL-17 levels in CSF of these patients. We first compared the detection rate of IL-17 in CSF between neurosyphilis patients and non-neurosyphilis patients. We found that there was five-fold higher detection rate of IL-17 in CSF in neurosyphilis patients than that in non-neurosyphilis patients (Fig. 2A). The average levels of CSF IL-17 was also significantly higher in neurosyphilis patients (2.29 pg/ml) (range of 0–59.83 pg/ml) than that in non-neurosyphilis patients (0 pg/ml) (range of 0–2.60 pg/ml) (Fig. 2B). IL-17 was not detected in CSF of the control group (Fig. 2A & B).
We further compared the levels of CSF IL-17 between patients with asymptomatic and symptomatic neurosyphilis. The detection rates of CSF IL-17 were 47.5% and 61.9% in asymptomatic and symptomatic neurosyphilis, respectively. The level of CSF IL-17 in symptomatic neurosyphilis patients (4.91 pg/ml, range from 0 to 59.83 pg/ml) was significantly higher than that in asymptomatic neurosyphilis patients (0.715 pg/ml, range from 0 to 44.27 pg/ml). Noted that the symptomatic neurosyphilis patient group included meningovascular, paretic, ocular and tabetic neurosyphilis. Further examination showed that the level of CSF IL-17 was the highest among paretic patients (7.6 pg/ml, range from 0 to 38.07 pg/ml) (Table 3).
T. pallidum is capable of invading central nervous system and damaging local tissues. There are detectable CSF abnormalities in neurosyphilis patients, including positive CSF VDRL, pleocytosis, and/or elevated protein concentration [19]. These measurements correlate well with the disease activity [19]. Since the above data showed that neurosyphilis patients had increased CSF IL-17 levels, we further investigated a possible relationship between CSF IL-17 levels and other measurements. As shown in Fig. 3, there was a significant positive correlation between CSF IL-17 levels and CSF protein concentrations or CSF VDRL titer, but not with the CSF WBC counts.
In some neurosyphilis patients, CSF IL-17 was not detected. We thus further investigated whether there are certain factors which may contribute to this phenomenon. We found that there were no differences in age, sex, the baseline serum RPR titer, or duration of symptoms prior to diagnosis between the IL-17 positive and IL-17 negative neurosyphilis groups. However, the IL-17 positive group had higher CSF protein concentration and CSF VDRL titer and higher frequency of symptomatic neurosyphilis than that of the IL-17 negative group (Table 4).
We further analyzed IL-17-producing cells in CSF of neurosyphilis patients. Because of the limited sample sizes and lymphocyte cell numbers in CSF collected from neurosyphilis patients, CSF cells from 14 neurosyphilis patients who had high levels of CSF pleocytosis (>50 cells/ul) were collected and stimulated for intracellular staining for the purpose of this study. 14 subjects included 6 (42.9%) subjects with asymptomatic neurosyphilis, 5 (35.7%) subjects with paretic, 2 (14.3%) subjects with ocular neurosyphilis, 1 (7.14%) subjects with meningovascular neurosyphilis. The average percentage of CD4+ T cells was 58.28% (51.65%–80.1%) of total lymphocytes, and Th17 (CD3+CD4+IL-17+ cells) was 1.8% (0.25%–4.6%) (Fig. 4). However, Th17 cells accounted for 88.8% (45.1%–100%) of total IL-17+ cells (Fig. 4), indicating that Th17 is the dominant IL-17-producing cells and may play an important role in neurosyphilis.
Among 44 subjects with confirmed neurosyphilis, 22(50%) subjects were asymptomatic neurosyphilis, 15 (34.1%) were paretic neurosyphilis, 5(11.4%) were ocular neurosyphilis, 2 (4.5%) were meningovascular neurosyphilis. All enrolled patients were routinely under followed-up examination and treated with standard therapy according to the Chinese treatment Guidelines.
Factors that were included in the final regression models of normalization of each laboratory measure, the improvement of clinical symptoms and their HRs are shown in Table 5. Factors that may improve laboratory markers or clinical symptoms were analyzed using the final regression model (Table 5). The neurosyphilis treatment regimen did not influence normalization of any of the 4 laboratory markers and the improvement of clinical symptoms. Normalization of the CSF protein concentration was more likely to occur in subjects with early syphilis (p = 0.018). CSF-VDRL reactivity was less likely to become normal in patients with positive CSF IL-17 (p = 0.04) and with higher baseline CSF-VDRL titer (p = 0.019). Serum RPR reactivity was more likely to return to normal in subjects with higher baseline serum RPR titers (p = 0.008).
T. pallidum remains one of the human pathogens that cannot be cultivated in vitro to-date. A suitable animal model for studying the pathogenesis of syphilis is also lacking. These obstacles have greatly hindered the effort of elucidating the basic immunobiological traits of syphilis. As a consequence, little is known about how T. pallidum causes damage to the central nervous system.
IL-17, a potent proinflammatory cytokine, plays a key role in the induction and development of tissue injury. IL-17 results in an increased production of ICAM-1, IL-6 and IL-8, and an increased synergy of many effects of IL-1β and TNF-α, which enhances the local inflammation and leads to inflammatory destruction [20], [21]. In this study, we observed an elevated CSF IL-17 in neurosyphilis patients. A similar scenario has been observed in infectious and auto-immune CNS disorders [22], [23]. Furthermore, we found that the level of CSF IL-17 is positively associated with CSF VDRL titer and total CSF protein in neurosyphilis patients. These findings suggest that IL-17 may involve in the CNS damage in neurosyphilis patients.
Syphilis is known as a “great imitator” because it is protean, especially neurosyphilis. Based on the patient's clinical and laboratory features, neurosyphilis is divided into five diagnostic categories, including asymptomatic, meningitis, meningovascular, paretic, and tabetic neurosyphilis [5]. Meningitis involves diffuse inflammation of the meninges with signs and symptoms of meningitis including headache, photophobia, nausea, vomiting, cranial nerve palsies, and occasionally seizures. It is diagnosed within 12 months after T. pallidum infection but it is relatively rare [5]. Unfortunately, no meningitis neurosyphilis patients enrolled in this study, and thus, the involvement of IL-17 in this stage of neurosyphilis remains unknown.
The pathogenic mechanisms underlying different clinical presentations of neurosyphilis are largely unknown. In this study, we found that higher levels of IL-17 were observed in CSF of symptomatic neurosyphilis patients, especially in paretic patients, compared with asymptomatic neurosyphilis patients. Moreover, the higher CSF protein and VDRL titer were observed in symptomatic neurosyphilis patients. These results suggest that IL-17 may be associated with clinical symptoms in neurosyphilis patients. Asymptomatic neurosyphilis does occur in both early and latent stages of syphilis. It is reported that there was an elevated CSF IL-17 level in early asymptomatic neurosyphilis patients, which correlated with the extent of CSF abnormality [15]. In our study, CSF IL-17 could be detected in 66.7% (12/18) of early asymptomatic neurosyphilis patients. It is believed that asymptomatic neurosyphilis patients may develop to late neurological complications [24]. Moreover, the extent of abnormalities of CSF positively correlated with the probability of developing late neurological complications [25]. Based on these notions, we suggested that some early asymptomatic neurosyphilis patients might have persistent IL-17 inflammation response, which could damage the CNS, resulting in neurological symptoms. Regrettably, there has been no study to compare long-term outcomes between CSF IL-17 negative and positive asymptomatic neurosyphilis patients.
Pastuszczak et al. identified that there was a strong correlation between CSF IL-17 and CSF pleocytosis in early asymptomatic neurosyphilis patients [15]. But our results indicated that there was no correlated between the level of CSF IL-17 and CSF pleocytosis. The CSF pleocytosis in neurosyphilis was related to the syphilis stage besides to the CNS damage. There was a marked pleocytosis in patients with acute meningeal syphilis. In late stage, CSF WBC counts in some neurosyphilis patients were less or even normal and were inconsistent with clinical symptoms. In up to 10% of patients with tabes referred to as the “burned out” stage, the CSF cell count may be normal [5]. In our study, there were early and late stage neurosyphilis patients. There were some paresis patients the CSF WBC counts were normal, though the clinical manifestations were severe. Therefore, according to the data, the CSF WBC counts were not always correlated with the degree of CNS damage. The different stage patients were enrolled in our study, leading to be inconsistent with the previous results.
Besides CD4+ T cells, other cells are capable of secreting IL-17 [26]. It was previously shown that IL-17 is mainly secreted by CD4+ T cells (Th17) in CSF in patients with chronic inflammatory demyelinating polyradiculoneuropathy (CIDP) [27]. In this study, we observed the CD4+ T cells were accumulated in CSF in neurosyphilis patients, and they were the dominant IL-17-producing cells. This finding suggests that Th17 response is a part of the local CNS response in a sub-population of neurosyphilis patients.
Our results showed that Th17 cells were increased in CSF of neurosyphilis patients. The mechanisms underlying the increase of Th17 in CNS remain unclear. IL-17 can disrupt the tight junction molecules and activates the endothelial contractile machinery, leading to disruption of blood brain barrier (BBB) [28]. Thus, Th17 in CSF may be a consequence of passive diffusion from blood. On the other hand, microbial lipopeptides such as Helicobacter pylori HP-NAP and B. burgdorferi NapA, can induce Th17 differentiation and production of IL-17 [29], [30]. In this regard, T. pallidum TpF1 is a protein homolog of HP-NAP and NapA [31], which may be capable of promoting Th17 differentiation and expansion in CNS. Interestingly, recent data showed that TpF1 can stimulate Treg cell differentiation [32]. The mechanisms underlying the increase of Th17 in CNS in neurosyphilis need to be further elucidated.
Because Th17 response may induce the immune-mediated CNS injury, we further evaluated the relationship between IL-17 and the clinical outcome of neurosyphilis. The baseline CSF-VDRL titer, and serum RPR titer influenced the likelihood of normalization of each parameter, consistent with previous studies [33]. However, we observed that CSF IL-17 positive neurosyphilis patients were 2.43 times less likely to normalize CSF-VDRL reactivity, even after taking into account the baseline CSF-VDRL titer and the stage of syphilis. CSF VDRL titer may reflect the T. pallidum burden and the extent of CNS damage. T. pallidum invaded CNS can induce Th17 immune response and CSF IL-17 were positively correlated with CSF VDRL titer. So positive CSF IL-17 in patients may reflect higher number of T. pallidum spirochetes in CSF. Since longer time would be required to clear a higher number of T. pallidum burden, the likelihood of normalization of CSF VDRL reactivity at the end of the observation would be lower. Because the sample size is limited (the number of subjects included in the regression analyses was only ranged from 21 to 44 patients in this study), a large sample study is needed to further understanding the true immune response at different stages of neurosyphilis and its clinical significance.
In conclusion, our findings demonstrate that neurological damage in syphilis patients is associated with increased CSF Th17/IL-17 response. CSF IL-17 may be used to evaluate the clinical outcome of treatment of neurosyphilis.
|
10.1371/journal.pcbi.1005947 | Life cycle synchronization is a viral drug resistance mechanism | Viral infections are one of the major causes of death worldwide, with HIV infection alone resulting in over 1.2 million casualties per year. Antiviral drugs are now being administered for a variety of viral infections, including HIV, hepatitis B and C, and influenza. These therapies target a specific phase of the virus’s life cycle, yet their ultimate success depends on a variety of factors, such as adherence to a prescribed regimen and the emergence of viral drug resistance. The epidemiology and evolution of drug resistance have been extensively characterized, and it is generally assumed that drug resistance arises from mutations that alter the virus’s susceptibility to the direct action of the drug. In this paper, we consider the possibility that a virus population can evolve towards synchronizing its life cycle with the pattern of drug therapy. The periodicity of the drug treatment could then allow for a virus strain whose life cycle length is a multiple of the dosing interval to replicate only when the concentration of the drug is lowest. This process, referred to as “drug tolerance by synchronization”, could allow the virus population to maximize its overall fitness without having to alter drug binding or complete its life cycle in the drug’s presence. We use mathematical models and stochastic simulations to show that life cycle synchronization can indeed be a mechanism of viral drug tolerance. We show that this effect is more likely to occur when the variability in both viral life cycle and drug dose timing are low. More generally, we find that in the presence of periodic drug levels, time-averaged calculations of viral fitness do not accurately predict drug levels needed to eradicate infection, even if there is no synchronization. We derive an analytical expression for viral fitness that is sufficient to explain the drug-pattern-dependent survival of strains with any life cycle length. We discuss the implications of these findings for clinically relevant antiviral strategies.
| Viral infections such as HIV, hepatitis B, hepatitis C, and influenza may be treated with antiviral drug therapy. These drugs generally block a specific phase of the virus’s life cycle, preventing it from replicating in the body. Often, a virus population can evolve drug resistance by acquiring mutations that prevent the drug from binding and inhibiting it. Here, we propose a new mechanism of drug resistance. We use mathematical models and stochastic simulations to show that virus populations can build resistance to antiviral therapies by synchronizing their life cycle with the dosing pattern of the drug. This process, which we refer to as “drug tolerance by synchronization”, can allow the virus to increase its overall fitness in the presence of the drug without altering drug binding. We show that viral strains whose life cycle lengths are approximately integer multiples of the time between drug doses possess an advantage during drug therapy. They can outcompete unsynchronized strains and lead to therapy failure with drugs that would otherwise have been successful.
| Viral infections are a major cause of human morbidity and mortality [1]. While vaccines to prevent viral illnesses have existed for over a century, it is only in the past several decades that drugs directly targeting viral replication have been developed. Antiviral drugs now exist for pathogens including the human immunodeficiency virus (HIV) [2], hepatitis B [3] and C [4] viruses, influenza A and B viruses [5], herpes simplex viruses [6], cytomegalovirus [7], Epstein Barr virus [8], and varicella zoster virus (chickenpox virus) [9]. These drugs each target a specific phase of the virus’s life cycle, and by binding to a viral protein or otherwise interfering with a critical step in viral replication are able to reduce the virus’s growth rate. Examples of viral functions targeted by antiviral drugs include binding of the viral particle to the target cell membrane, transcription of the viral genome, integration of the virus in the host cell genome, or post-translational cleavage of viral proteins.
Antiviral treatments that are initially successful at reducing viral loads may eventually be rendered ineffective, in individual patients or in entire populations, by the emergence of drug-resistant strains [7, 10–15]. Drug resistance occurs when a viral strain gains a mutation that allows it to replicate efficiently despite the presence of the drug, and this strain subsequently outcompetes the wild-type strain to reach high levels in the viral population. Resistance can be conferred by mutations that interfere with the ability of the drug molecule to inhibit the intended viral target. For example, for antivirals that block the fusion of viral particles with the target cell membrane, the mutations which confer resistance can alter the shapes and chemical properties of viral proteins to either prevent drug binding or allow the virus to enter the cell despite the presence of the drug.
There may, however, be another mechanism by which resistance can develop in viral populations. In a 2000 paper [16], Wahl and Nowak hypothesized that a heretofore unobserved effect, which they termed cryptic resistance, may prevent a viral population from being suppressed by an antiviral drug, without it needing to evolve the ability to alter drug binding or even complete its life cycle in the presence of the drug. This insight was motivated by the realization that drug levels are not constant during the course of viral treatment, and hence the viral fitness in the presence of the drug is also time-varying. Like most medications, antiviral drugs are administered in discrete doses of constant size separated by approximately equal time intervals. Shortly after a dose, the drug is absorbed and drug levels are high, and the relevant stage of the viral life cycle is maximally inhibited. However, between doses, drug levels decay, and may eventually reach low levels where they are no longer suppressive. This pattern repeats periodically over the course of treatment. These authors suggested that the virus could avoid the effects of the drug by always completing the targeted phase of its life cycle when drug concentrations are at a minimum. If the length of the viral life cycle is a mutable trait, then in the presence of drug treatment it may evolve to become approximately equal to the duration of time between successive doses. The virus population would then become synchronized with the pattern of drug levels. In this manner, the virus could sustain itself indefinitely by “hiding” from the highest concentrations of the drug, even though it has no means of counteracting the effects of drug molecules when they are present.
Lifecycle timing, at least in some bacteriophages, is known to be a mutable trait subject to life history trade-offs [17–24]. Separately, antibiotic “tolerance” in bacteria is contrasted with traditional “resistance” in that it only implies the ability to temporarily survive high drug levels [25]. Tolerance can be heritable and traced back to particular mutations, and in certain in vitro settings, can even involve changes to the timing of particular growth cycle-dependent life cycle stages [26, 27]. However, since Wahl and Nowak’s original paper, no in vivo, in vitro, or in silico studies have examined whether cryptic resistance, perhaps more aptly called “tolerance by synchronization”, could actually evolve during antiviral treatment.
Here, we use mathematical models to show that tolerance by synchronization can plausibly arise in a viral population subjected to antiviral treatment. We start by augmenting well-established models for viral infection dynamics to account for distinct phases of the viral life cycle. This model includes a maturation rate, which can take on different values that result in different viral life cycle lengths. Fluctuating drug concentrations are incorporated as a periodic time-dependent infectivity of the virus. We evaluate the success of this tolerance strategy via two different methods. First, we examine viral fitness in a single-strain infection by determining the growth rate and long-term infection size, and we compute the life cycle durations that optimize each of these values. Next, to predict how an infection evolves, we study competition between multiple strains. Specifically, we search for strains that can outcompete others and can persist despite drug treatment. We discover that strains that are synchronized to the drug dose schedule—that is, have life cycle durations that are a near-integer multiple of the time between drug doses—can dominate the infection and cause sustained treatment failure, therefore conferring tolerance. These strains would avoid detection by most genotypic drug resistance assays, which look for mutations only in the viral protein targeted by the drug therapy, and by in vitro phenotypic susceptibility assays, which apply constant, not periodic, drug levels. Our results suggest that new experimental methods are needed to identify this potentially important new mechanism of antiviral resistance.
The standard viral dynamics model makes the simplifying assumption that infected cells produce new virus particles as soon as they are infected [28]. In reality, a virus must complete many stages of its life cycle before new virions are created. These stages may include uncoating of the viral particle, transcription and translation of the viral genome, copying of the viral genome, assembly of viral proteins, or even cell cycle-dependent events. While the exact steps involved in production vary among different viruses, all have in common a delay between infection of a cell and viral production [29]. This intracellular delay is a prerequisite for drug tolerance via synchronization.
There are two common methods for incorporating delays into a dynamic model: first, by introducing a series of maturation phases, each represented by a state variable, and second, by using delay differential equations. We explore both methods in this paper.
For the first method (Fig 1A), we posit a series of n immature phases wi, (similar to [30]):
x ˙ ( t ) = λ - β ( t ) x ( t ) y ( t ) - d x x ( t ) w ˙ 1 ( t ) = β ( t ) x ( t ) y ( t ) - ( m 1 + d w ) w 1 ( t ) w ˙ i ( t ) = m i - 1 w i - 1 ( t ) - ( m i + d w ) w i ( t ) ∀ 2 ≤ i ≤ n y ˙ ( t ) = m n w n ( t ) - d y y ( t ) (1)
The size of the population of healthy target cells is described by the state variable x; these cells are produced at rate λ and die at per capita rate dx. We consider the infected cell population as being subdivided into two subpopulations: “immature” (wi, where the subscript i indicates the particular immature phase if there are more than one) and “mature” (y) infected cells. Immature cells in phase i progress to become immature cells in phase i + 1 (or fully mature cells if i = n) at per capita rate mi, and immature cells in phase i die at per capita rate dw. Mature infected cells produce virus and lead to infection of healthy cells at rate proportional to the product of both their levels and the infectivity rate β(t), and die at per capita rate dy. We allow for the infectivity β to be time-dependent, since we assume that drug treatment acts on this parameter and that drug levels may vary over time (detailed in next section).
The maturation time τ is the time required for a newly infected cell in the first immature infected state w1 to progress and become fully mature in state y. We assume, for simplicity, that each maturation step occurs at the same rate, so mi = mn for all i. The probability distribution of maturation times is therefore a Gamma (Erlang) distribution [31] that depends on the number of maturation steps, n, and the rate constant, mn:
p ( τ ) = m n n τ n - 1 e - m n τ ( n - 1 ) ! (2)
We set the maturation rate, mn, to be a linear function of the number of maturation steps, n, such that mn = nm. With this choice of mn, the average maturation time is independent of the number of maturation steps (i.e., 〈τ〉 = n/mn = 1/m). The standard deviation in maturation times, however, is inversely proportional to both the maturation rate, m, and the square root of the number of maturation steps (i.e., σ τ = 1 / ( m n )) (Fig 1B). Therefore, for a given number of intermediate phases, n, strains with lower expected maturation times (higher maturation rates) have less deviation in their time to maturation than strains with longer expected maturation times. Moreover, for a given maturation rate, m, more maturation steps would allow for the virus to better control its maturation time; in the limit of large n, the maturation time is fixed and equal to 1/m.
The second method for modeling intracellular delay uses delay differential equations with a fixed delay, τ = 1/m, between infection and start of virion production (similar to [32–34]):
x ˙ ( t ) = λ - β ( t ) x ( t ) y ( t ) - d x x ( t ) y ˙ ( t ) = β ( t - τ ) x ( t - τ ) y ( t - τ ) e - d w τ - d y y ( t ) (3)
As for the basic viral dynamics model, we can define the basic reproductive ratio, R0, as the average number of new infections generated by a single infected cell over the course of its lifetime. If β(t) = β is constant, then, in the multi-stage model, R0 is given by
R 0 = λ β d x d y ( n m n m + d w ) n (4)
Similarly, in the fixed-delay model, R0 is given by
R 0 = λ β d x d y e - d w / m (5)
In the limit of large n, the systems described by Eqs (1) and (3) are equivalent, and thus so are the expressions for R0 given by Eqs (4) and (5). If there is the possibility for an immature cell to die before producing virus (i.e., if dw > 0), then R0 is maximized when the maturation rate, m, is maximized (so that infected cells mature as quickly as possible). If immature cells do not die (i.e., if dw = 0), then R0 is independent of the maturation rate, m. As in the basic model, R0 is sufficient to qualitatively determine the outcome of the system of equations. If R0 > 1, then the infection will persist and reach an equilibrium, and if R0 < 1, then the infection will decline and eventually go extinct [30, 35]
For both models of viral dynamics, we also implemented a multi-strain competition between strains with different average maturation times (see S1 Text for equations). When β(t) = β is constant, an R0 value can be derived for each strain. When dw = 0, all strains have the same R0 value and can co-exist at values that can be calculated analytically. When dw > 0, the strain with the shortest maturation time has the highest R0 value and eventually outcompetes all other strains.
In addition, we also created stochastic versions of each of the single- and multi-strain models, where the rate of each reaction is equal to that in the differential equation formulation, and each reaction increases or decreases a state variable by 1. The stochastic process was simulated with the Gillespie next reaction method [36]. In stochastic versions of the model, extinction can happen even when R0 > 1.
Note that throughout this paper, we use a common assumption to simplify the viral dynamics model and reduce the computational burden of simulations. For the vast majority of infections, the dynamics of free virions tend to be much faster than the dynamics of cells. Virus is produced in large quantities and is rapidly cleared in vivo. This implies that the free virus population tends to reach a “quasi-steady state” level with respect to the level of infected cells, implying that a separation of timescales can be applied to the system. Consequently, we do not explicitly track free virus but instead assume that its level is proportional to the abundance of infected cells. With this assumption, our infectivity parameter β(t) is actually a composite parameter given by β(t) = b(t)k/c, where b(t) is the infectivity of a free virion, k is the rate constant for production of free virus by mature infected cells, and c is the rate constant for clearance of free virus.
Ordinary differential equations were numerically integrated with the Scipy odeint algorithm in Python 2.7 [37], and delay differential equations were numerically integrated with the Scipy-based DDE solver ddeint [38]. For results presented in the main text, we choose parameter values that roughly correspond to HIV infection (Table 1). Throughout the paper, we often present results for the case of no death of immature infected cells (i.e., dw = 0), since in this case, the value of R0 in the presence of constant drug levels is independent of the maturation rate, m, and the number of maturation stages, n, which makes it easiest to see how R0 changes under periodic drug levels. However, we also present results for cases with other values of dw, including the most natural assumption: that immature infected cells die at the same rate as uninfected cells (i.e., dw = dx). Since immature infected cells are not yet producing new virions, they are less likely to a) experience direct cytotoxic effects of virus production and release, and b) be presenting viral epitopes and triggering cytolytic immune responses, so their death rate may resemble that of healthy uninfected cells.
We model the effect of an antiviral drug on the infection by assuming that it reduces the infectivity in a manner that depends on the drug concentration, D(t), which varies over time. We first model the periodic time-dependence of drug levels as an on-off switch. The drug is taken every T days and completely inhibits new infections for a time fT thereafter, such that β(t) = 0 for that time; we refer to this time as the on window. For the remaining time between doses, we assume the drug to be completely inactive, such that β(t) = β0 for that time; we refer to this time as the off window. We also refer to f as the efficacy of the drug; a perfect drug therapy would have efficacy equal to 1 (see Fig 1C):
D ( t ) = { 1 , if t mod T < f T 0 , if t mod T > f T β ( t ) = β 0 ( 1 - D ( t ) ) (6)
We also considered drug dynamics that follow a simple pharmacological model. When the drug is taken consistently, drug levels D(t) peak immediately at Cmax following each dose and then decay exponentially with half-life th, reaching a minimum value C min = C max 2 - T / t h, before the next dose. Infectivity is reduced in a concentration-dependent manner that is described by a Hill dose-response curve, where β0 is the infectivity in the absence of treatment, IC50 is the concentration at which 50% inhibition occurs, and M quantifies the steepness of the decay curve:
D(t)=Cmax2−t/thβ(t)=β01+(D(t)IC50)M (7)
Drug efficacy can be varied in this model by changing Cmax, th, IC50, or M. However, wherever Eq (7) are used, we choose to vary th only. For any parameter combinations used in this model, we can always find a corresponding f value for the on-off model in Eq (6) such that the time-average rate of infection is equal between the two models. Indeed, throughout the paper, we express overall drug efficacy values using the single quantity f for both models.
In addition to modeling perfect adherence to periodic drug treatment, we consider imperfect adherence. We assume there is a fixed probability of taking each dose at the scheduled time or missing it completely, and that each dose is taken independently. In the on-off model, a missed dose results in a full drug period with D(t) = 0. In the pharmacological model, drug continues to decay after a missed dose, and a subsequent dose increases the concentration by Δ = Cmax − Cmin. Example values of R0(t) under this model are shown in S1 Fig.
When viral fitness is time-varying, as is the case under fluctuating drug concentrations, the above calculations for R0 no longer hold. A time-averaged R0 value (R ¯ 0) could be calculated that takes into account the time-dependence of the infectivity, R ¯ 0 = R 0 ( 〈 β ( t ) 〉 ). However, as we will show below, this average R0 is no longer sufficient to describe the outcome of the model.
In this paper, we showed that it is possible for a virus population to synchronize its life cycle with the pattern of drug therapy. This process allows for strains to persist and cause treatment failure during anti-viral treatments where success would be expected. Although originally called “cryptic resistance” when it was first proposed by Wahl and Nowak [16] in 2000, we have opted for the updated term “tolerance by synchronization”, which reflects the current terminology in microbiology for differentiating between the ability to grow despite sustained (“drug resistant”) versus transiently (“drug tolerant”) high drug levels. Here, tolerant strains survive repeated, transiently high drug levels via heritable life cycle timing. The main condition needed for emergence of tolerance by synchronization is tight control of viral life cycle timing. This tight control (i.e. low variance) may be achieved even if each life cycle stage has random duration, as long as there are sufficiently many stages.
In order for viral synchronization to be a feasible mechanism for drug tolerance, the life cycle length of wild-type virus must be at least close in magnitude to the dose interval (or an integer multiple thereof) since it may be unrealistic to assume that a virus could dramatically change its life cycle length. To this end, we examined the viral generation time—defined as the average time from the moment one cell is newly infected until one of its offspring infects a new cell—and the recommended dosing schedules for viral infections for which targeted therapy is available (Table S1 Table). This includes HIV, hepatitis B virus, hepatitis C virus, herpes simplex virus 1 and 2, cytomegalovirus, and influenzavirus A and B. We found that in all cases, the conditions for possible synchronization were met. Although current experimental methods of characterizing resistance generally preclude identification of synchronization-based mechanisms (detailed below), there are some hints of effects in which synchronization may play a role. For example, for HIV it is known that resistance mutations to protease inhibitors can occur outside the protease gene (e.g. [44]), and for hepatitis C virus, there are many examples of clinical failure without a known resistance mutation, or multi-drug resistant strains with no known mechanism (e.g. [45] and references therein.)
The “resistance” mechanism described here is aptly “cryptic” in the sense that it would evade detection by all existing in vitro tests for drug susceptibility. Genotypic resistance tests typically look for amino acid changes in the viral sequence which codes for the protein targeted by the drug, and particularly regions important for drug binding. Since a gene influencing viral life cycle length could appear anywhere in the genome, such tests would likely miss these mutations. In order to adapt these tests, experiments would need to be done to identify genetic loci associated with life cycle length and then adapt resistance screens to look for changes at these sites. Phenotypic resistance tests suffer a similar problem: they measure in vitro viral replication against a series of drug levels, but since these levels will be constant within the culture media, synchronization cannot occur and strains conferring tolerance will appear fully susceptible. These tests would need to be conducted in a device, such as a bioreactor or microfluidic chip, that could recreate drug profiles experienced in vivo to be able to detect tolerance via synchronization.
While we have shown that tolerance via life cycle synchronization is a possible means of evading therapy in silico, to our knowledge this effect has never been evaluated experimentally. A first step would be to create an in vitro viral culture system that could deliver periodic drug concentrations, and to find a model viral infection system in which genetic determinants of life cycle control are already established. Additionally, we would like to look for evidence of this strategy emerging in patients on antiviral therapy. A first step could be to identify patients who have only partially suppressed viremia despite high adherence and who have virus that appears drug-sensitive by the genotypic and/or phenotypic resistance tests described above.
Although this work is the first we are aware of to explore the evolution of viral life cycles in response to drug treatment, previous work has explored other determinants of life cycle length evolution, in particular for bacteriophages [17–24]. A classic question has been how long a phage should wait before lysing a host cell, when there is naturally a trade-off between benefit of delaying lysis to accumulate viral progeny within the cell, and the need to rapidly spread to other potential target cells. Models of this process have characterized the determinants of lysis time in terms of, for example, host cell density and intracellular host resources [18, 19], but have not, to our knowledge, considered periodic effects. We have incorporated ideas from this work into our model, showing that drug dosing period acts in combination with trade-offs between slowing maturation (to produce more progeny) and increasing maturation (to either spread faster or avoid death) to determine the optimal life cycle length.
The emergence of tolerance by viral life cycle synchronization can only lead to therapy failure, on its own, if trough drug levels are not fully suppressive. Trough levels depend on drug dose and the kinetics of absorption and clearance (e.g. Cmax and th), and suppression at trough levels depends on the viral fitness (R0) in the absence of drug, and the parameters of the dose-response curve (Eq (7)). For drugs with a short half-life, a steep dose-response curve, and a narrow therapeutic window, it is more likely that worries about toxicity prevent reaching Cmax levels high enough to ensure that even trough concentrations prevent viral replication. Cryptic resistance is most likely to occur in this regime. Even if viral replication is suppressed for the wild-type strain throughout the dose interval, synchronization could augment partially-resistant mutations that act by standard mechanisms (e.g. altering drug binding).
Here, we have assumed that the viral infectivity is instantaneously affected by the current drug level. In reality, some drugs may need to undergo further steps, such as breakdown into active forms, active transport into cells, etc., which could delay their effect. These processes may alter the form of the periodic drug levels, but the periodic nature—and hence the potential for viral synchronization—will be preserved. However, if the drug binds irreversibly to a cellular target that does not turnover or intracellularly to a viral target that is not continually produced, or does so reversibly but with a very slow dissociation rate, then there may be no periodicity in drug effect despite a periodicity in concentrations.
We assumed that the drug acts on the infection of new target cells by free virus. However, the drug could also act on another phase of the life cycle, which could be represented as blocking the transition from one stage of immature infected cell to another. Tolerance could potentially occur in these scenarios as well, since the main requirement for its existence is for synchronized strains to possess an evolutionary advantage over the others. We saw that, in general, this is achieved through a tight control over the viral life cycle length.
For some viral infections, treatment is administered as combinations of different drugs, often with the same dosing schedules, that act on different stages of the virus life cycle. In this case it is much harder for synchronization to confer a benefit, as it may be impossible for the virus to complete the multiple targeted stages of its life cycle at the time when both drug levels are low. However, therapy failure could occur by a combination of resistance mechanisms—altered drug binding for one drug, and synchronization for the remaining drug.
In some viral infections, multiple infections of the same target cell may be common. When infection contains multiple viral strains, complex dynamics within multiply-infected cells can alter evolutionary dynamics. For example, recombination, by a variety of mechanisms, can occur and can have complex effects on selection [46]. Phenotypic mixing or multiplicity reactivation can lead to production of virions with mismatches between genotype (nucleic acid carried) and phenotype (structural and functional proteins carried), which also greatly complicates evolutionary predictions [47, 48]. Finally, within-cell competition can select for different traits than between-cell competition, for example leading to competitor colonizer trade-offs [49]. Future models, designed to more precisely capture the details of particular viral infections, could include multiply-infected cells.
Our results show that common methods for calculating R0 have limited use when considering periodically administered therapies. Even though a version of R0 can easily be constructed which takes into account the time-averaged viral fitness in the presence of fluctuating drug levels, this quantity does not discriminate between strains that can persist versus go extinct in the presence of the drug. This is because the process of synchronization does not just apply to strains that have life cycle lengths very close to integer multiples of the drug period, and hence benefit most from the effect. It also alters the long-term fitness of all strains. Strains that are most asynchronous do worse than predicted by the time-averaged R0. This failure of existing methods for R0 is due to the combined presence of both time-dependent effects and a stage-structured model (e.g. immature versus mature cells). We have derived a method for calculating a version of R0 that does account for synchronization, and although there appears to be no simple expression for this quantity, it can be calculated numerically and can predict which strains will persist under periodic drug levels.
The system we analyze in this paper has parallels to population-level epidemic models in which there may be periodic fluctuations in disease parameters. While the most common source of periodicity is seasonality [50], which likely occurs on a timescale much longer than the generation time of infection, other periods, such as weekly changes in contact rates due to work/school days versus weekends, may occur on timescales for which interactions with the generation time are more relevant. In fact, mathematicians studying such models have independently suggested constructs for the basic reproductive ratio under periodic model coefficients [51], proved that their definitions represent persistence thresholds [52], suggested algorithms to actually compute these quantities [53], and determined the scenarios under which the simpler time-averaged approach is correct [53]. They have even observed a phenomenon they call “resonance” [54], in which the early growth rate of such models is enhanced if the period of environmental change is close to some natural timescale of the infection, which is analogous to the effect we observe and call synchronization.
While we have discussed tolerance via synchronization in the context of viral infections, this is by no means the only possible case. Other microbial causes of infection, such as bacteria and protozoans, could also use this mechanism to avoid life-cycle-stage-specific drug targeting, assuming their life cycle length is greater than or equal to the drug period. Similarly, cancer chemotherapy may be administered with dose intervals in the right range for cell cycle synchronization to occur. These cellular organisms, as opposed to viruses, may not actually have to evolve synchrony via genetic changes, but may have the cellular machinery to make them capable of regulating cell-cycle length phenotypically.
In fact, in particular laboratory protocols used to grow and evolve bacterial cultures in the presence of antibiotics, an effect called “tolerance by lag” has been observed which bears some similarities to the viral life cycle synchronization that we describe here. In the context of antibiotic resistance, “tolerance” is defined as the ability to temporarily survive high levels of bacteriocidal antibiotic, and can occur by either genetic or non-genetic mechanisms. When bacteria are moved from a stressful environment where growth has been suppressed (for example, crowded culture media), to a resource-rich environment (e.g., diluted into new media), there is a well-documented “lag” phase before cell growth and division resumes. If cultured bacteria are repeatedly allowed to grow until stasis-via-overcrowding before being diluted and transferred, and if the transfer always involves a transient period of (bacteriocidal) antibiotic-treated media, then the population will evolve an altered lag time which matches with the length of antibiotic exposure [25–27]. While this work suggests bacteria too could use synchrony as a resistance mechanism, it remains unclear whether the in vitro growth protocol designed to force a lag phase at the time of treatment is relevant to any process that naturally occurs during infection.
|
10.1371/journal.ppat.1007080 | Neisseria gonorrhoeae employs two protein inhibitors to evade killing by human lysozyme | The bacterial pathogen Neisseria gonorrhoeae (Gc) infects mucosal sites rich in antimicrobial proteins, including the bacterial cell wall-degrading enzyme lysozyme. Certain Gram-negative bacteria produce protein inhibitors that bind to and inhibit lysozyme. Here, we identify Ng_1063 as a new inhibitor of lysozyme in Gc, and we define its functions in light of a second, recently identified lysozyme inhibitor, Ng_1981. In silico analyses indicated that Ng_1063 bears sequence and structural homology to MliC-type inhibitors of lysozyme. Recombinant Ng_1063 inhibited lysozyme-mediated killing of a susceptible mutant of Gc and the lysozyme-sensitive bacterium Micrococcus luteus. This inhibitory activity was dependent on serine 83 and lysine 103 of Ng_1063, which are predicted to interact with lysozyme’s active site residues. Lysozyme co-immunoprecipitated with Ng_1063 and Ng_1981 from intact Gc. Ng_1063 and Ng_1981 protein levels were also increased in Gc exposed to lysozyme. Gc lacking both ng1063 and ng1981 was significantly more sensitive to killing by lysozyme than wild-type or single mutant bacteria. When exposed to human tears or saliva, in which lysozyme is abundant, survival of Δ1981Δ1063 Gc was significantly reduced compared to wild-type, and survival was restored upon addition of recombinant Ng_1981. Δ1981Δ1063 mutant Gc survival was additionally reduced in the presence of human neutrophils, which produce lysozyme. We found that while Ng_1063 was exposed on the surface of Gc, Ng_1981 was both in an intracellular pool and extracellularly released from the bacteria, suggesting that Gc employs these two proteins at multiple spatial barriers to fully neutralize lysozyme activity. Together, these findings identify Ng_1063 and Ng_1981 as critical components for Gc defense against lysozyme. These proteins may be attractive targets for antimicrobial therapy aimed to render Gc susceptible to host defenses and/or for vaccine development, both of which are urgently needed against drug-resistant gonorrhea.
| The mucosal pathogen Neisseria gonorrhoeae has acquired resistance to almost all recommended antibiotics, and no gonorrhea vaccine currently exists. Attractive targets for therapeutic discovery include bacterial factors that, when inactivated, enhance bacterial susceptibility to host-derived antimicrobial components. The bacterial cell wall-degrading enzyme lysozyme is abundant in mucosal secretions and innate immune cells. To resist killing by lysozyme, some bacteria produce proteins that bind to and directly inhibit the activity of lysozyme. Here, we demonstrate lysozyme inhibitory activity in the N. gonorrhoeae protein Ng_1063. We found that both Ng_1063 and a second, recently described lysozyme inhibitor, Ng_1981, contribute to full resistance of N. gonorrhoeae to lysozyme, including resistance to lysozyme-rich mucosal secretions and human neutrophils. Although Ng_1063 and Ng_1981 are both inhibitors of lysozyme, they are distinct in their sequences, biological activities, and cellular localizations. Because both Ng_1063 and Ng_1981 are extracellular, we propose they can be targeted for vaccines and drugs that sensitize Gc to human antimicrobial defenses.
| Neisseria gonorrhoeae (Gc) is a Gram-negative diplococcus and the causative agent of the sexually transmitted infection gonorrhea. The World Health Organization (WHO) estimates 78 million cases of gonorrhea occur each year, with over 800,000 cases reported annually in the United States [1–3]. The lack of a protective vaccine, widespread prevalence of antibiotic-resistant Gc, and treatment failures with last line therapeutics have prompted the United States Centers for Disease Control to label antibiotic-resistant Gc as an urgent threat to public health [1, 4–6]. Likewise, the WHO lists the control and elimination of Gc infection as a high priority [7]. Dissecting Gc pathogenesis and virulence is critical for the development of novel therapeutics and vaccines.
Gc colonizes mucosal sites, including the cervix, urethra, pharynx, conjunctiva, and rectum. Colonization initiates an inflammatory response, culminating in the robust recruitment of neutrophils, an innate immune cell with antimicrobial killing activities [8]. Thus, Gc survival during human infection requires defenses against the antimicrobial molecules made both by the mucosal epithelium and neutrophils. For instance, Gc evades killing by cationic antimicrobial proteins by modifying surface lipooligosaccharide and producing multidrug efflux pumps, and these activities are important for Gc survival from human neutrophils, survival in the mouse model of Gc cervicovaginal colonization, and/or survival in a human male urethral-infection model [9–11].
The antimicrobial protein lysozyme is ubiquitous at the mucosal sites colonized by Gc, reaching concentrations as high as 2 mg/mL in the conjunctiva and 1 mg/mL in the cervix [12, 13]. Lysozyme is also abundant in phagocytes like neutrophils [14, 15]. Lysozyme hydrolyzes the glycan backbone of bacterial cell wall peptidoglycan, causing lysis and death. Because lysozyme is highly cationic, it can also kill bacteria through an enzymatic-independent mechanism, purportedly via pore formation on bacterial membranes [16–19]. Bacteria have evolved numerous, non-redundant mechanisms to thwart the killing activities of lysozyme, which in many cases contribute to enhanced survival and virulence in vivo (reviewed in [20]). These observations highlight lysozyme as a critical player in host defense, and, in turn, underscore the importance for lysozyme resistance to a pathogen’s success.
Lysozyme resistance in Gc is mediated by envelope integrity, peptidoglycan modifications, and protein inhibitors of lysozyme. In Gram-negative bacteria like Gc, the outer membrane (OM) restricts periplasmic access of lysozyme [21]. Consequently, maintenance of envelope integrity is vital for Gram-negative resistance to lysozyme. We recently reported that two cell wall turnover proteins, the lytic transglycosylases LtgA and LtgD, contribute to envelope integrity in Gc, and ΔltgAΔltgD mutant Gc is more sensitive to killing by lysozyme and human neutrophils [22]. Another mechanism of lysozyme resistance in Gc involves O-acetylation of peptidoglycan, which can sterically hinder lysozyme’s hydrolytic activity [23, 24]. O-acetylation, however, only contributes to lysozyme resistance in Gc if the envelope is also compromised, as when LtgA and LtgD are lost [22, 24].
Recently, we identified a third mechanism of lysozyme resistance in Neisseria, the production of a protein inhibitor of lysozyme [25]. Protein inhibitors have only been identified in Gram-negative bacteria, and their expression contributes to bacterial survival within mucosal secretions, survival from phagocytes, and/or survival in vivo [26–30]. Inhibitors against c-type lysozymes, like human lysozyme, are classified as Ivy (inhibitor of vertebrate lysozyme), PliC (periplasmic lysozyme inhibitor of c-type lysozyme), or MliC (membrane-bound lysozyme inhibitor of c-type lysozyme). PliC- and MliC-type inhibitors share structural similarity and conserved sequence motifs that are distinct from Ivy-type inhibitors [31]. All function by insertion of one or more protein loops into the active site of lysozyme to interfere with its peptidoglycan-hydrolyzing activity [31]. We found that purified recombinant Ng_1981 (also known as the Ng-Adhesin Complex Protein, Ng-ACP) from Gc inhibits the enzymatic activity of lysozyme in vitro and contributes to gonococcal tolerance to lysozyme [25]. The homolog of Ng_1981 in Neisseria meningitidis, NMB_2095 (Nm-ACP), is 94% identical to Ng_1981 and binds directly to lysozyme with micromolar affinity [25]. Although the structure of NMB_2095 exhibits overall similarity with PliC/MliC-type inhibitors, NMB_2095 and Ng_1981 lack the conserved PliC/MliC sequence features and therefore were classified as a novel type of lysozyme inhibitor [25, 31].
Gram-negative bacteria often produce multiple, non-redundant inhibitors of lysozyme [26–28, 32–35], prompting us to investigate whether Gc also produces more than one inhibitor. In this work, we report that the open reading frame ng1063 (KEGG GENOME, Gc strain FA1090) encodes for a protein that shares sequence and structural homology with MliC-type inhibitors. We tested the hypothesis that Ng_1063 functions as a lysozyme inhibitor in Gc and further defined its biological activity in the context of Ng_1981. Our findings suggest that Ng_1063 functions as a bona fide inhibitor of lysozyme and that Gc employs both Ng_1063 and Ng_1981, which exhibit distinct properties and localizations, for optimal defense against lysozyme.
Ng_1063 shares overall amino acid sequence similarity with MliC from P. aeruginosa (23% identity, 40% similarity) and E. coli (16% identity, 32% similarity) (Fig 1A). S89 and K103 of P. aeruginosa MliC interact with the active site residues of lysozyme and are required for MliC inhibitory function [36]; the corresponding residues, S83 and K103, are conserved in Ng_1063 (Fig 1A). Like other MliC inhibitors, Ng_1063 is predicted to be a lipoprotein as determined by signal sequence analysis using lipoP 1.0 (http://www.cbs.dtu.dk/services/LipoP/). We characterized the potential for Ng_1063 as an MliC-type inhibitor of lysozyme using PHYRE2 (www.sbg.bio.ic.ac.uk/phyre2) to predict the Ng_1063 three-dimensional structure. PHYRE2 predicted Ng_1063 to have an MliC-type fold with 99.6% confidence. Alignment of the Ng_1063 predicted structure with P. aeruginosa MliC in PyMOL shows overlap of S83 and K103 in Ng_1063 with the corresponding residues of P. aeruginosa MliC (Fig 1B). Threading of Ng_1063 through the P. aeruginosa MliC-lysozyme co-crystal structure (PDB 3f6z) positions the S83 and K103 residues of Ng_1063 in close proximity to the active site residues of lysozyme, E53 and D70 (Fig 1C). Based on these in silico analyses, we hypothesized that Ng_1063 is a protein inhibitor of lysozyme.
To test this hypothesis, we first examined if Ng_1063 protein could rescue the lysozyme-mediated lysis of Micrococcus luteus, a Gram-positive bacterium that is intrinsically sensitive to lysozyme. We generated purified, recombinant Ng_1063 (r1063) protein (see Materials and methods) and found it prevented the lytic activity of lysozyme on M. luteus in a concentration-dependent manner (Fig 2A). As Gc is the relevant context for Ng_1063 activity, we next investigated the effect of r1063 on survival of Gc after exposure to lysozyme. In contrast to M. luteus, WT Gc is relatively resistant to lysozyme. Thus, we used an ΔltgAΔltgD Gc mutant, which we previously demonstrated has increased sensitivity to lysozyme, as a tool to assess r1063 inhibitory activity with Gc [22]. As expected, ΔltgAΔltgD mutant Gc was markedly reduced in survival in the presence of lysozyme, compared to WT (strain MS11) (Fig 2B). Preincubation of lysozyme with r1063, but not an unrelated Neisserial OM protein (macrophage infectivity potentiator (MIP) [25]), restored survival of the ΔltgAΔltgD mutant to WT levels (Fig 2B). Overexpression of ng1063 in ΔltgAΔltgD mutant Gc was also sufficient to rescue bacterial survival to WT levels (Fig 2C). Although the ΔltgAΔltgD mutant has reduced envelope integrity that correlates with lysozyme sensitivity [22], overexpression of ng1063 in this mutant did not affect bacterial susceptibility to vancomycin, an indicator of cellular permeability (S1 Table). These observations strongly suggest that Ng_1063 protects Gc by inhibiting lysozyme, rather than enhancing envelope integrity.
Together, these results identify Ng_1063 as a new lysozyme inhibitor made by Gc.
Besides Ng_1063, Gc produces another inhibitor of lysozyme, Ng_1981, which we recently characterized [25]. Ng_1981 shares limited sequence similarity with Ng_1063 (19% identity, 35% similarity) (S1A Fig) yet is predicted to share a similar overall three-dimensional structure [25]. Recombinant Ng_1981 (r1981) inhibits the lytic activity of lysozyme against M. luteus [25], and we found that pretreatment of lysozyme with r1981 or overexpression of ng1981 restored ΔltgAΔltgD mutant survival after exposure to lysozyme to WT levels (S1B and S1C Fig). Overexpression of ng1981 did not affect the sensitivity of ΔltgAΔltgD mutant Gc to vancomycin (S1 Table).
We next tested whether both Ng_1063 and Ng_1981 interact with lysozyme in the physiological context of the bacterial cell. We generated a Δ1981Δ1063 mutant, where ng1063 was replaced with a null allele and ng1981 was disrupted by insertion-deletion. Δ1981Δ1063 mutant Gc was then complemented with C-terminal FLAG-tagged versions of either ng1063 or ng1981 at an ectopic chromosomal location under IPTG regulation (see Materials and methods). As a negative control, we used Gc expressing C-terminal FLAG-tagged ltgA or ltgD, also under IPTG regulation (see Materials and methods) [37]. These strains were induced for FLAG-protein expression and incubated with lysozyme or vehicle control. FLAG-tagged proteins and their interacting partners were then immunoprecipitated from bacterial cell lysates. Lysozyme co-precipitated with Ng_1063-FLAG and Ng_1981-FLAG, but not with LtgA-FLAG, LtgD-FLAG, or anti-FLAG resin alone, showing the lysozyme inhibitors both interact with lysozyme in intact bacteria (Fig 3A).
In some bacteria, exposure to lysozyme induces a signaling cascade that results in upregulation of lysozyme resistance factors [38–41]. Therefore, we next examined whether treating Gc with lysozyme under sublethal conditions (see Materials and methods and Fig 4A) affected Ng_1063 and Ng_1981 protein abundance. To do so, we engineered a Gc strain carrying 1063(WT)-FLAG at its native locus to probe for Ng_1063 using the anti-FLAG antibody. There was a significant increase in Ng_1063 protein from Gc treated with lysozyme for 3 hrs, compared with vehicle (Fig 4B and 4C). Similarly, using anti-r1981 antisera, Ng_1981 protein was significantly increased upon lysozyme treatment, compared to vehicle (Fig 4D and 4E). On average, Ng_1063 and Ng_1981 protein were increased 2-fold and 3.5-fold, respectively, in Gc treated with 1 mg/mL lysozyme (Fig 4C and 4E).
Taken together, these findings indicate that both Ng_1063 and Ng_1981 interact with lysozyme in intact Gc, and Gc alters expression of Ng_1063 and Ng_1981 proteins after exposure to lysozyme.
We next took a genetic approach to define how Ng_1063 and Ng_1981, singly and in combination, affect Gc resistance to killing by lysozyme. To our surprise, survival of Δ1063 mutant Gc was equivalent to WT after exposure to lysozyme (Fig 5A). One possible explanation for this phenotype is that the outer membrane barrier of Gc limits the access of lysozyme to the periplasm where Ng_1063 could be functioning; however, loss of ng1063 had no effect on bacterial survival compared with parent Gc under the following envelope-compromising conditions: subinhibitory concentrations of the pore-forming cationic antimicrobial peptide LL-37 (S2A and S2B Fig); membrane-destabilizing treatment with EDTA (S2C Fig); and genetically-induced loss of envelope integrity using the ΔltgAΔltgD mutant (S3 Fig). In contrast, at the same concentrations of lysozyme used for the Δ1063 mutant, Δ1981 mutant Gc was significantly compromised for survival compared to WT bacteria, reaching a near 5-fold decrease in survival when exposed to 1 mg/mL lysozyme (Fig 5A). Complementation with ng1981 restored Δ1981 mutant survival to WT levels (Fig 5A). Loss of ng1981 also reduced survival of the ΔltgAΔltgD mutant after exposure to lysozyme, an observation that highlights how multiple, mechanistically distinct factors contribute to lysozyme resistance in Gc (S3 Fig). Thus, Ng_1981 is confirmed as an important factor for Gc defense against lysozyme.
Importantly, deletion of both ng1063 and ng1981 markedly reduced Gc survival in the presence of lysozyme in a concentration-dependent and time-dependent manner (Fig 5A and 5B). The sensitivity to lysozyme in Δ1981Δ1063 mutant Gc was greater than either single mutant alone, and this effect was particularly apparent at higher concentrations of lysozyme (Fig 5A and 5B). As seen in Fig 5B, recoverable CFU from Δ1981Δ1063 mutant Gc was reduced 100-fold compared to Δ1981 single mutant Gc after 4 hrs exposure to lysozyme. Similarly, loss of both ng1063 and ng1981 significantly impaired the survival of ΔltgAΔltgD mutant Gc after lysozyme exposure, in comparison to loss of ng1981 alone (S3 Fig). Complementation with ng1063 in Δ1981Δ1063 mutant Gc restored survival to the level of the single Δ1981 mutant (Fig 5A). In the absence of lysozyme, single and double mutants grew similarly to WT Gc (Fig 5B). These findings indicate that in addition to Ng_1981, Ng_1063 is also important for Gc defense against lysozyme. Moreover, these findings shed light on the biological activities of Ng_1063 and Ng_1981 in Gc, where Ng_1981 can fully compensate for loss of ng1063, yet Ng_1063 cannot fully compensate for loss of ng1981.
Lysozyme can kill bacteria through its hydrolase activity and through an enzyme-independent mechanism that relies on its cationicity [16, 17]. We investigated if Ng_1063 and Ng_1981, like other proteinaceous inhibitors of lysozyme, defend Gc from the enzymatic activity of lysozyme. To this end, we exposed Δ1981Δ1063 mutant Gc to lysozyme that had been boiled, eliminating its enzymatic activity [22, 42]. Survival of the Δ1981Δ1063 mutant was unaltered by boiled lysozyme compared to WT Gc (Fig 5C), confirming that Ng_1063 and Ng_1981 inhibit the enzymatic activity of lysozyme.
We next tested the ability of Ng_1063 and Ng_1981 to defend Gc against other peptidoglycan-degrading enzymes (i.e., peptidoglycan muramidase activity) (S4A Fig). Δ1981Δ1063 mutant Gc was significantly reduced over 6-fold in survival after incubation with the chicken egg white c-type lysozyme, compared to WT, and survival was rescued by overexpression of ng1063 (S4B Fig). At the concentrations tested, the single Δ1981 mutant was unaffected by chicken egg white lysozyme (S4B Fig). In contrast, in the presence of mutanolysin, a distinct bacteria-derived muramidase (S4A Fig), survival of Δ1981 and Δ1981Δ1063 mutant Gc was equivalent to WT (S4C Fig). Mutanolysin was active in this setting, as evidenced by the reduced survival of ΔltgAΔltgD mutant Gc (S4C Fig) [22]. These findings indicate that Ng_1063 and Ng_1981 can inhibit some but not all muramidases, possibly owing to co-evolution of Gc with humans, which possess one c-type lysozyme.
Residues S89 and K103 of P. aeruginosa MliC interact with the active site of lysozyme and are required for lysozyme inhibition (see Fig 1) [36]. Thus, we evaluated if the corresponding residues in Ng_1063, S83 and K103, similarly contributed to lysozyme inhibition by complementing Δ1981Δ1063 Gc with C-terminal 3X-FLAG-tagged Ng_1063, in which each of these residues was replaced with an alanine. Immunoblotting with a FLAG antibody showed equivalent expression between Ng_1063 WT, S83A, and K103A variants in the Δ1981Δ1063 mutant upon IPTG induction (S5 Fig). In the presence of lysozyme, Δ1981Δ1063 mutant Gc expressing WT Ng_1063-FLAG exhibited a significant, greater than 60-fold increase in survival, which was not reproduced with expression of either S83A or K103A Ng_1063-FLAG (Fig 5D). This finding suggests that these residues are important for the lysozyme inhibitory activity of Ng_1063.
A MUSCLE alignment between Ng_1063 and Ng_1981 shows possible conservation of the MliC inhibitory serine (S76 in Ng_1981) and lysine residues (K99 in Ng_1981) (S1A Fig). We previously found that the N. meningitidis homolog to Ng_1981, NMB_2095, exhibits structural homology to MliC/PliC inhibitors, yet computational docking models between NMB_2095 and lysozyme predicts a mode of binding that differs from MliC/PliC inhibitors [25]. Because mutation of residues in NMB_2095 that were predicted to interact with lysozyme (i.e., Asn79, Tyr84, and Gly95) failed to alter its inhibitory activity [25], we investigated the role of S76 and K99 in Ng_1981 inhibition of lysozyme. We complemented Δ1981 mutant Gc with C-terminal 3X-FLAG-tagged Ng_1981, where S76 and K99 were replaced with an alanine residue. We observed equivalent expression of Ng_1981(WT)-FLAG, Ng_1981(S76A)-FLAG, and Ng_1981(K99A)-FLAG by immunoblot upon IPTG induction (S1D Fig). Complementation with either WT, S76A, or K99A versions of ng1981 was sufficient to significantly increase Δ1981 mutant survival in the presence of lysozyme (S1E Fig). This finding supports our previous conclusions that Ng_1981 and its homolog NMB_2095 behave differently from MliC/PliC inhibitors and are thus novel inhibitors of lysozyme.
Mucosal sites that are colonized by Gc are bathed in fluids with high concentrations of lysozyme, including tears (2 mg/mL) and saliva (0.12 mg/mL) [13]. We therefore tested the possibility that Ng_1063 and Ng_1981 contribute to Gc survival when exposed to pooled human tears or pooled human saliva. Compared to WT Gc, Δ1981Δ1063 mutant Gc was significantly reduced in survival in the presence of tears as well as saliva (Fig 6A and 6B). Survival of the single Δ1981 mutant was equivalent to WT at the dilutions tested (Fig 6A and 6B). We confirmed the secretions contained active lysozyme by their ability to lyse M. luteus (S6A and S6B Fig). Pretreatment of the secretions with r1981 was sufficient to inhibit the lytic activity of secretions against M. luteus (S6A and S6B Fig), and pretreatment with r1981 restored Δ1981Δ1063 mutant survival to WT levels (Fig 6A and 6B).
Neutrophils are another abundant source of lysozyme and are heavily recruited to sites of Gc infection [14, 15]. Thus, we tested if Ng_1063 and Ng_1981 are important for Gc survival from adherent, interleukin 8-treated primary human neutrophils, a model to recapitulate the physiological state of tissue-migrated neutrophils during gonorrheal disease [43]. Δ1981 mutant Gc exhibited a modest but statistically significant reduction in survival in the presence of human neutrophils, compared to WT and ng1981 complemented Gc (Fig 6C). In contrast, survival of Δ1063 Gc was equivalent to WT after exposure to neutrophils, and Δ1981Δ1063 mutant Gc was equally sensitive to neutrophils as the single Δ1981 mutant (Fig 6C). Under the conditions tested, Gc resides in a phagosome that exhibits limited fusion with neutrophil primary granules, which contain lysozyme [14, 15, 44, 45]. In comparison with Δ1981Δ1063 mutant Gc, ΔltgAΔltgD mutant Gc is markedly sensitive to even small amounts of lysozyme. Because we previously linked lysozyme sensitivity of the ΔltgAΔltgD mutant with increased killing by human neutrophils [22], we next tested whether overexpression of ng1063 or ng1981 enhanced ΔltgAΔltgD survival in the presence of human neutrophils, as was the case in vitro with purified human lysozyme (see Fig 2C and S1C Fig). However, neither overexpression of ng1063 nor ng1981 was sufficient to rescue the survival defect of ΔltgAΔltgD mutant Gc in the presence of human neutrophils, pointing to factors in addition to lysozyme for the sensitivity of ΔltgAΔltgD mutant to neutrophils (S7 Fig).
Together, these findings suggest that Ng_1063 and Ng_1981 help defend Gc from physiologically relevant sources of lysozyme that would be encountered in its obligate human host.
Although Ng_1063 and Ng_1981 both interact with lysozyme and display a similar ability to inhibit the enzymatic activity of lysozyme, their biological activities in Gc are distinct. To gain insight into the mechanism underlying these differences, we examined the localization of Ng_1063 and Ng_1981 in Gc. Ng_1063 is a putative lipoprotein predicted to be extracellularly exposed on the surface of Gc [46], whereas Ng_1981 is predicted to be a soluble, periplasmic protein, according to LipoP 1.0 and CELLO bioinformatic analyses. To test these predictions, we assessed the surface exposure of each protein, using Δ1063 and Δ1981 mutant Gc complemented with IPTG-inducible WT copies of C-terminal FLAG-tagged Ng_1063 and Ng_1981, respectively. FLAG-complemented Gc were incubated with anti-FLAG antibody, and fluorescence intensity per individual bacterium was quantified by imaging flow cytometry. We detected strong surface labeling of FLAG protein from ng1063-FLAG complemented Gc (Fig 7A–7C). In contrast, ng1981-FLAG complemented Gc displayed negligible surface expression of FLAG protein, which was no different from the negative control (Fig 7A–7C).
As a complementary approach, we visualized FLAG-tagged protein expression using immunofluorescence microscopy. We observed staining with the FLAG antibody in ng1063-FLAG complemented Gc, but not in ng1981-FLAG complemented Gc under non-permeabilizing conditions, in agreement with the imaging flow cytometry data (Fig 7D). Surface expression of Ng_1981-FLAG was also not detectable using polyclonal anti-r1981 antisera (S8 Fig). When Gc was permeabilized with methanol and Triton-X-100 prior to incubation with the FLAG antibody, both ng1063-FLAG and ng1981-FLAG complemented Gc exhibited peripheral staining, suggestive of localization to the bacterial envelope (Fig 7D). Together, these findings indicate that Ng_1063, but not Ng_1981, is exposed extracellularly on the surface of Gc.
The lack of Ng_1981 surface staining was unexpected because its homolog in N. meningitidis, NMB_2095, was found to be surface-exposed, where it contributes to bacterial adhesion to epithelial cells [47]. A clue to resolving this discrepancy came from the results in Fig 3, where we observed less FLAG-tagged Ng_1981 protein than FLAG-tagged Ng_1063 protein in lysates prepared from pelleted bacteria, despite using the same overexpression system. We thus considered the possibility that a fraction of Ng_1981 is released from Gc. To directly test this possibility, we evaluated the presence of Ng_1981 and Ng_1063 protein in bacterial whole cell lysates and supernatants. We used Gc carrying 1063(WT)-FLAG at its native locus to simultaneously detect Ng_1063-FLAG using the anti-FLAG antibody and Ng_1981 using anti-r1981 antisera. Conditioned supernatants from this strain contained Ng_1981, but not Ng_1063, while both proteins were present in the whole cell lysates (Fig 7E). In addition, we did not detect the cytoplasmic protein ZWF in conditioned supernatants (Fig 7E), providing further evidence that Ng_1981 is not released as the result of autolysis.
To test whether Ng_1981 in Gc supernatants contains functional inhibitory activity against lysozyme, we co-cultured Δ1981Δ1063 mutant Gc with Ng_1981-complemented Gc in the Δ1981Δ1063 background. Under these conditions, the presence of the ng1981 complement significantly increased the survival of the non-complemented Δ1981Δ1063 mutant after exposure to lysozyme (Fig 7F). This observation suggests that extracellularly released Ng_1981 can protect Gc from lysozyme, in keeping with our initial findings with ectopically added recombinant protein (S1 Fig).
Taken together, these results indicate that Gc produces two distinct inhibitors of lysozyme: Ng_1981 is both released extracellularly and is in the bacterial envelope, whereas Ng_1063 is OM-anchored and surface-exposed.
The success of Gc as a human pathogen requires that it defends itself against antimicrobial components found at mucosal surfaces. In this work, we identified Ng_1063 as a new functional homolog of the MliC-type lysozyme inhibitors in Gc. We then interrogated its functions alongside another, recently characterized lysozyme inhibitor, Ng_1981 [25]. We found that Ng_1063 and Ng_1981 interact with lysozyme in the physiological context of the bacterium, and both proteins’ expression was increased upon exposure to lysozyme. Gc lacking both ng1063 and ng1981 was markedly reduced in survival after exposure to human lysozyme and lysozyme-containing human secretions, and this effect was greater than either single mutant alone. Ng_1063 was exposed on the surface of Gc, while Ng_1981 was not; however, a fraction of Ng_1981 protein was released into the extracellular milieu. Based on these results, we conclude that Gc produces two distinct inhibitors of lysozyme that together confer full resistance to this abundant antimicrobial defense protein.
MliC-type inhibitors of lysozyme are periplasmic-facing lipoproteins that form an eight-stranded antiparallel β-barrel structure [31]. The structure of P. aeruginosa MliC in complex with lysozyme revealed the protrusion of two loops from the inhibitor into the active site of lysozyme, each loop contributing one key, conserved inhibitory residue (S89 and K103) [31, 36]. Our in silico analyses predicted that Ng_1063 is a lipoprotein with sequence and structural similarity to MliC-type inhibitors. We obtained biological support for this prediction by showing that the corresponding inhibitory residues in Ng_1063, S83 and K103, were required for its inhibitory activity. In contrast, S83 and K103 were dispensable for the interaction of Ng_1063 with lysozyme in vivo (Fig 3B). This finding implicates a second binding site in Ng_1063 for lysozyme, which has been shown for a shallow binding pocket consisting of Y92 and T98 in P. aeruginosa MliC [36]. Intriguingly, we found that C-terminal FLAG-tagged Ng_1063 is on the surface of Gc. Since Ng_1063 has no predicted membrane-spanning regions, the entire Ng_1063 protein is likely exposed extracellularly, where it could bind and inhibit lysozyme. This observation contrasts with most MliC-type inhibitors, which are predicted to reside within the inner leaflet of the OM, facing the periplasm [31]. Since most studies have not experimentally verified localization of MliC proteins, we anticipate other microbes produce MliC-type proteins that are also surface-exposed. It remains unclear how Ng_1063 becomes surface exposed, but a likely candidate involves the recently identified surface lipoprotein assembly modulator (Slam) [48].
Several bacterial species, including E. coli, P. aeruginosa, Yersinia pestis, and Edwardsiella tarda, produce more than one inhibitor of lysozyme, and often these proteins are non-redundant [26–28, 32–35]. Here, we found that Gc produces two proteins, Ng_1063 and Ng_1981, which both bind to and inhibit lysozyme. Each protein’s expression is increased upon exposure to lysozyme. While other bacteria regulate lysozyme resistance factors via alternative sigma factors or two component systems [38–41, 49], such regulatory networks are limited in Gc; thus how this regulation is occurring in Gc is not yet clear. Despite their shared ability to inhibit lysozyme, Ng_1063 and Ng_1981 have several distinguishable properties. First, even though Ng_1063 and Ng_1981 share overall structural similarity with MliC/PliC-type inhibitors [25], these proteins share little sequence similarity (19% identity; 35% similarity). In fact, Ng_1981 lacks conserved MliC/PliC sequence motifs [25, 31]. Further, we found that mutation of the Ng_1981 residues S76 and K99, which align with the key inhibitory residues of Ng_1063, failed to alter Ng_1981 inhibitory activity against lysozyme. Ng_1981 protein is also predicted to be soluble whereas Ng_1063 is a predicted lipoprotein. In contrast to Ng_1063 that is surface-exposed, we found evidence that Ng_1981 is present both in the Gc envelope and in the extracellular milieu. It remains unclear whether extracellular Ng_1981 is present as a soluble protein or resides within or associates with outer membrane vesicles; however, we found evidence that ng1981 expressing Gc could rescue survival of lysozyme-susceptible Gc in trans, suggesting that extracellular Ng_1981 retains inhibitory activity. Future studies will investigate how Ng_1981 is released by Gc.
Gc employs both Ng_1063 and Ng_1981 for optimal defense against lysozyme, but in a non-redundant manner. Loss of ng1981 significantly reduced Gc survival when exposed to lysozyme or human neutrophils. This finding extends upon our previous work, which showed sensitivity of the Δ1981 mutant in a different genetic background (strain FA1090) to lysozyme [25], and suggests that the presence of Ng_1063 is insufficient to fully compensate for loss of ng1981. In contrast, survival of Gc lacking ng1063 was equivalent to WT after exposure to lysozyme. Even though OM-permeabilization has been used to reveal a protective role for lysozyme inhibitors in other bacteria [34, 35], survival of Δ1063 mutant Gc remained unaltered from WT when the Gc envelope was compromised. Although our reductionist conditions did not reveal a phenotype for the single Δ1063 mutant, it is possible that this protein contributes to Gc survival in vivo, warranting future investigation. This finding implies that Ng_1981 sufficiently compensates for loss of ng1063 under our in vitro conditions. The differential ability of Ng_1981 and Ng_1063 to protect Gc from lysozyme may be explained by their differential localization, as well as differences in their expression or affinity for lysozyme, all of which are subjects for future investigation. Nevertheless, the role of Ng_1063 as a barrier to lysozyme in Gc was revealed in the absence of ng1981, where Δ1981Δ1063 mutant Gc exhibited increased sensitivity to lysozyme over either single mutant. Importantly, ng1063 and ng1981 are expressed during human infection in female patients, implying that both proteins contribute to colonization and/or pathogenesis in vivo [50]. Moreover, pathogenic and non-pathogenic Neisseria carry homologs for both Ng_1063 (S2 Table) and Ng_1981 [25], and in the Ng_1063 homologs, the S83 and K103 equivalent residues are 100% conserved (S3 Table). Since commensal Neisseria, like the pathogens, colonize mucosal sites, conservation of Ng_1063 and Ng_1981 amongst Neisseria may be important for residence at these lysozyme-rich surfaces. Together, these findings support a role for both Ng_1063 and Ng_1981 in Gc lysozyme resistance and implicate these inhibitors as important virulence determinants in Gc.
In addition to Ng_1063 and Ng_1981, Gc O-acetylates its peptidoglycan to directly block lysozyme-mediated hydrolysis [22, 24]. Gc also expresses two cell wall turnover proteins, LtgA and LtgD, that maintain envelope integrity, a physical barrier to lysozyme [22]. Thus, lysozyme resistance in Gc is multifactorial. Given the diverse mechanisms employed by Gc to resist lysozyme, if and how these factors are connected at the level of gene regulation or protein-protein interactions remains an open question. However, it is noteworthy that each factor contributes unequally to lysozyme resistance. For example, O-acetylation is dispensable for Gc survival in the presence of lysozyme, as long as envelope integrity is also maintained [22, 24]. In contrast, both an ΔltgAΔltgD mutant and an Δ1981Δ1063 mutant exhibit sensitivity to low concentrations of lysozyme, where the ΔltgAΔltgD mutant is the most sensitive [22]. Together, these mechanistically distinct approaches underscore that lysozyme resistance in Gc is vital to the infectivity of this pathogen. By extension, these observations open the possibility of targeting such factors to enhance bacterial susceptibility to lysozyme in the human host. For instance, chemicals that interfere with Ng_1063 and Ng_1981 could help to treat infections with strains exhibiting increased resistance to traditional antibiotics. Ng_1063 and Ng_1981 are also potential vaccine candidates, based on their extracellular localization, expression during human infection, and relative conservation among Gc strains (S2 Table) [25, 50, 51]. A vaccine against Ng_1063 and Ng_1981 is attractive because antibodies raised against them could promote Gc killing by multiple mechanisms. For instance, antibody deposition on surface Ng_1063 could promote bacterial killing via opsonophagocytosis and serum bactericidal activity, while function-neutralizing antibodies against both extracellular Ng_1063 and Ng_1981 could sensitize Gc to lysozyme in mucosal secretions and in neutrophils [51]. Thus, we propose that Ng_1063 and Ng_1981 are worthy targets for consideration in the context of revitalized global efforts to develop new antibiotics and vaccines against the urgent threat of drug-resistant gonorrhea.
Human Subjects: Venous blood was collected from adult healthy human subjects with their signed informed consent. All materials collected were in accordance with a protocol (#13909) approved by the University of Virginia Institutional Review Board for Health Sciences Research.
Animals: Healthy, specific pathogen-free rabbits were immunized with r1981 by Davids Biotechnologie GmbH (Regensburg, Germany). Davids Biotechnologie GmbH has a permit from the Veterinäramt Regensburg for housing specific-pathogen free, healthy rabbits according to §11 TierSchG (Az31.4.4/ScP1). The company is registered for immunization of animals under the Aketenzeichen: AZ 2532.44/14 at the approving authority Umweltamt Regensburg/Veterinärwesen. All immunizations were done in accordance with National Institute of Health standards for animal welfare (NIH animal welfare number A5646-01).
All Gc used in this study are in the MS11 background and are listed in S4 Table. WT Gc is a RecA+ MS11 strain nonvariably expressing the VD300 pilin [22], and ΔltgAΔltgD mutant Gc was used from previous work [52].
A null allele of ng1063 was made by introducing a stop codon in-frame into the coding region of the gene. A 5’ flank was PCR amplified using the forward primer 5’AAAAATTTACATTCCTCCGGGCGGGC3’ and the mutagenesis reverse primer 5’GCACAGGCCTACAATCTAGAAACCGATACG3’ (in-frame stop codon; XbaI site). A 3’ flank was amplified by PCR using the mutagenesis forward primer 5’TTTCTAGATTGTAGGCCTGTGCCGTG3’ (XbaI site; in-frame stop codon) and the reverse primer 5’AGCAGGTTTAAAGTTGGCATTGAGCCG3’. The 5’ and 3’ flanks were combined via overlap extension PCR, and the resulting product was introduced into Gc by spot transformation [53]. Bacteria from the transformation were screened by PCR followed by digestion with XbaI, and positive transformants were confirmed by sequencing the ng1063 allele. The ng1981 mutant was made by transforming Gc with pGEM-Δ1981, where 1981 is disrupted with a kanamycin cassette [25]. Transformants were selected on 50 μg/mL kanamycin and positive transformants confirmed by PCR.
To make the ng1981 complementation construct, ng1981 was PCR amplified using the forward primer 5’CCCCCGGGCGCCTTTTTACAAAC3’ (SmaI site) and reverse primer 5’GGAACCGCGGAAAAACAGCGTTTTCAG3’ (SacII site). The resulting product was digested with SmaI and SacII and ligated into the isopropyl β-D-1-thiogalactopyranoside (IPTG)-inducible complementation plasmid pKH35 [54]. The resulting plasmid (pKH35_1981) was used to spot transform Gc, where ng1981 was integrated between the lctP and aspC chromosomal loci. Bacteria were selected with 8 μg/mL chloramphenicol and confirmed by DNA sequencing. To make the ng1063 complementation construct, ng1063 was PCR amplified using the forward primer 5’ACATAGCCGCGGGCTTTAATGTG3’ (SacII site) and the reverse primer 5’ GACAGGAGATATCCAGAACGAAACG3’ (EcoRV site). The resulting product was digested with SacII and EcoRV and ligated into the anhydrotetracycline (AT)-inducible complementation plasmid pMR68 [55]. The resulting plasmid (pMR68_1063) was used to spot transform Gc, where ng1063 was integrated between the iga and trpB chromosomal loci. Transformants were selected for using 10 μg/mL erythromycin and confirmed by DNA sequencing.
To make the ng1063(S83A) point mutant, a 5’ flank was PCR amplified with the forward primer 5’ACATAGCCGCGGGCTTTAATGTG3’ (SacII site) and the reverse primer 5’GTTCGCCCGCTGCGGCAACGTC3’ (codon for alanine). A 3’ flank was PCR amplified with the forward primer 5’GACGTTGCCGCAGCGGGCGAAC3’ (codon for alanine) and the reverse primer 5’GACAGGAGATATCCAGAACGAAACG3’ (EcoRV site). The 5’ and 3’ flanks were combined via overlap extension PCR. The resulting product was digested with SacII and EcoRV, and ligated into pMR68 to make pMR68_1063(S83A). To make the ng1063(K103A) point mutant, a 5’ flank was PCR amplified with the forward primer 5’ACATAGCCGCGGGCTTTAATGTG3’ (SacII site) and the reverse primer 5’ CTTCGCCGCCCGCCTGGTGCCAC3’ (codon for alanine). A 3’ flank was PCR amplified using the forward primer 5’GTGGCACCAGGCGGGCGGCGAAG3’ (codon for alanine) and the reverse primer 5’ GACAGGAGATATCCAGAACGAAACG3’ (EcoRV site). The 5’ and 3’ flanks were combined via overlap extension PCR. The resulting product was digested with SacII and EcoRV, and ligated into pMR68 to make pMR68_1063(K103A).
To make the ng1981-3XFLAG construct, ng1981 was PCR amplified using the forward primer 5’CCCCCGGGCGCCTTTTTACAAAC3’ (SmaI site) and the reverse primer 5’TTGAATTCACGTGGGGAACAGTCTTTG3’ (EcoRI site). The resulting product was digested with SmaI and EcoRI, and then ligated into pMR100, a C-terminal 3XFLAG vector [56], to make pMR100_1981(WT). For additional 3’ homology, the 3’ region for ng1981 was PCR amplified using the forward primer 5’ GGCAAGCTTAAACAGCGTTTTCATTTCTG3’ (HindIII site) and the reverse primer 5’ GGCTCGAGGCCGCGGTCATTAAAAAAGAC3’ (XhoI site and SacII site). The resulting product was digested with HindIII and XhoI, and then ligated into pMR100_1981(WT). The resulting plasmid, pMR100_1981(WT)_3’homology, was digested with SmaI and SacII to release the 1981(WT)-3XFLAG-3’homology fragment, which was subsequently ligated into pKH35 to make pKH35_1981(WT)_3’homology. pKH35_1981(WT)_3’homology was used to spot transform Gc. Transformants were selected with chloramphenicol and were confirmed by DNA sequencing.
To make the ng1063(WT)-3XFLAG, ng1063(S83A)-3XFLAG, and the ng1063(K103A)-3XFLAG constructs, ng1063 was PCR amplified from MS11 genomic DNA, pMR68_1063(S83A) plasmid DNA, and pMR68_1063(K103A) plasmid, respectively, with the forward primer 5’ACATAGCCCGGGGCTTTAATGTG3’ (SmaI site) and the reverse primer 5’TTGAATTCACGGGCGCGGCAGGAAGTTTC3’ (EcoRI site). The resulting products were digested with SmaI and EcoRI and each ligated into pMR100 to make pMR100_1063(WT), pMR100_1063(S83A), and pMR100_1063(K103A). For additional 3’ homology, the 3’ region of ng1063 was PCR amplified using the forward primer 5’AAAAAGCTTAGCCTGTTTGAACCGCCG3’ (HindIII site) and the reverse primer 5’ GACTCGAGCCCGCGGACTTTAGGC3’ (XhoI site and SacII site). The resulting product was digested with HindIII and XhoI, and then ligated into pMR100_1063(WT), pMR100_1063(S83A), and pMR100_1063(K103A). The resulting plasmids, pMR100_1063(WT)_3’homology, pMR100_1063(S83A)_3’homology, and pMR100_1063(K103A)_3’homology, were digested with SmaI and SacII to release the 1063-3XFLAG-3’homology fragments, which were subsequently ligated into pKH35 to make pKH35_1063(WT)_3’homology, pKH35_1063(S83A)_3’homology, and pKH35_1063(K103A)_3’homology. Gc were spot transformed with these plasmids. Transformants were selected with chloramphenicol and were confirmed by DNA sequencing.
To make the ng1981(S76A)-3XFLAG construct, a 5’ flank of ng1981 was PCR amplified from MS11 genomic DNA using the forward primer 5’CCCCCGGGCGCCTTTTTACAAAC3’ (SmaI site) and the reverse primer 5’ GTCCATATTGTCCGCTTTATCCAAATTG3’ (codon for alanine). A 3’ flank of ng1981 was PCR amplified using the forward primer 5’CAATTTGGATAAAGCGGACAATATGGAC3’ (codon for alanine) and the reverse primer 5’TTGAATTCACGTGGGGAACAGTCTTTG3’ (EcoRI site). The 5’ and 3’ flanks were combined using overlap extension PCR. The resulting amplicon was digested with SmaI and EcoRI for replacement of WT ng1981 in the pMR100_1981(WT)_3’homology backbone to make pMR100_1981(S76A)_3’homology. pMR100_1981(S76A)_3’homology was subsequently digested with SmaI and SacII as above for insertion into pKH35 to make pKH35_1981(S76A)_3’homology. To make the ng1981(K99A)-3XFLAG construct, a 5’ flank of ng1981 was PCR amplified from MS11 genomic DNA using the forward primer 5’CCCCCGGGCGCCTTTTTACAAAC3’ (SmaI site) and the reverse primer 5’GTTTGCGGTAGGACGCGCTGTCCATTG3’ (codon for alanine). A 3’ flank of ng1981 was PCR amplified using the forward primer 5’CAATGGACAGCGCGTCCTACCGCAAAC3’ (codon for alanine) and the reverse primer 5’TTGAATTCACGTGGGGAACAGTCTTTG3’ (EcoRI site). Overlap extension PCR was used to combine the 5’ and 3’ flanks, and the resulting amplicon was digested with SmaI and EcoRI to make pMR100_1981(K99A)_3’homology as above. pMR100_1981(K99A)_3’homology was digested with SmaI and SacII and inserted into pKH35 to make pKH35_1981(K99A)_3’homology. Gc were transformed with the pKH35 constructs via spot transformation, and transformants were selected for with chloramphenicol and were confirmed by DNA sequencing.
To replace the native ng1063 gene with ng1063 with a C-terminal 3X-FLAG tag, the 5’ flank of ng1063 was PCR amplified from MS11 genomic DNA with the forward primer 5’AAAAATTTACATTCCTCCGGGCGGGC3’ and the reverse primer 5’CGGTCAGCGCGAAAAACCTGGTATT3’. The 3’ flank was PCR amplified from the pKH35_1063(WT)_3’homology plasmid with the forward primer 5’AATACCAGGTTTTTCGCGCTGACCG3’ and the reverse primer 5’TCTTGCAAGCGTTGGCAAACAGC3’. The 5’ and 3’ flanks were combined via overlap extension PCR, and the resulting product was spot transformed into Gc. Transformants were screened by PCR using the forward primer 5’AATACCAGGTTTTTCGCGCTGACCG 3’ and the reverse primer 5’AGCAGGTTTAAAGTTGGCATTGAGCCG3’, and positive transformants confirmed by DNA sequencing.
Piliated, opa-negative Gc were grown on Gonococcal Medium Base (GCB, Difco) plus Kellogg’s supplements [57] at 37°C with 5% CO2 (v/v). Gc was inoculated into liquid medium (GCBL with Kellogg’s supplements and NaHCO3) and repeatedly diluted until Gc reached mid-logarithmic stage, as described [58]. For experiments using human neutrophils, the absence of Opa expression in Gc was confirmed by Western blot with the 4B12 pan-Opa antibody. For IPTG- and AT-inducible constructs, 1 mM IPTG and 10 ng/mL AT, respectively, were added to Gc growing in liquid culture for at least 5 hr.
To compare protein sequences, the MUltiple Sequence Comparison by Log-Expectation (MUSCLE, http://www.ebi.ac.uk/Tools/msa/muscle/) tool was used. For alignment of ng1063 alleles, the Clustal omega tool was used. To determine percent identity and percent similarity between proteins, the Sequence Manipulation Suite (Ident and Sim; http://www.bioinformatics.org/sms2/ident_sim.html) feature was used on the MUSCLE alignment between two proteins in question, with signal sequences included. Protein signal sequences were predicted using the LipoP 1.0 Server (http://www.cbs.dtu.dk/services/LipoP/), and envelope localization was predicted using CELLO (subCELlular LOcalization predictor; http://cello.life.nctu.edu.tw/). The protein sequence of Ng_1063 (MS11), excluding the predicted signal sequence, was used for structure prediction via the PHYRE2 server (www.sbg.bio.ic.ac.uk/phyre2). The predicted Ng_1063 structure was aligned with the known structure of MliC from P. aeruginosa in complex with hen egg white lysozyme (PDB 3f6z, [36]) using PyMOL Molecular Graphics System. For allelic sequence comparisons across Neisseria, the NEIS1425 (ng1063) allele was analyzed on December 1, 2017 using the PubMLST database (http://pubmlst.org/perl/bigsdb/bigsdb.pl?db=pubmlst_neisseria_isolates).
The ng1063 gene sequence, optimized for E. coli expression and encoding the entire coding sequence for ng1063 (NEIS1425, http://pubmlst.org/neisseria/, 381 bp), was synthesized in vitro (GeneArt, Invitrogen). The ng1063 gene was cloned into the pET22b(+) system (Novagen) and inserted between the NdeI and XhoI restriction sites fused to a C-terminal hexa-histidine tag. The resulting recombinant plasmid, pET22b::1063(WT), was transformed into E.coli DH5α cells for plasmid amplification, and subsequently transformed into competent E. coli BL21 (DE3) pLysS (NEB) cells for protein expression. The transformants were cultured in LB broth at 37°C to mid-logarithmic phase, and expression of recombinant protein was induced by addition of IPTG to a final concentration of 1 mM. After growth at 37°C for an additional 4 h, the cells were harvested and insoluble recombinant Ng_1063 (r1063) protein was purified by nickel iminodiacetic acid (Ni-IDA) affinity chromatography under denaturing conditions. Bound protein was eluted using 100 mM NaH2PO4, 10 mM Tris-HCl, 6 M GuHCl and 250 mM imidazole buffer, pH 8.0, precipitated with 5% v/v Trichloroacetic acid (TCA), and subsequently resuspended in phosphate buffered saline (PBS), with 0.5% w/v SDS for solubilization. Protein concentration was determined using the BCA™ Protein Assay (Pierce). The molecular mass of mature r1063-His-tag without the leader peptide sequence (predicted by SignalP 4.1 Server) is 12.5 kDa, as confirmed by SDS-PAGE. Recombinant MIP (rMIP) and recombinant Ng_1981 (r1981) were prepared as described in [25, 59].
Rabbits (n = 2) were hyper-immunized subcutaneously with r1981 using the services of David Biotechnologie GmbH, Regensburg, Germany. Rabbits were immunized with r1981 (100 μg per dose per rabbit) emulsified in Freund’s Complete Adjuvant for the primary injection (day 0) and Freund’s Incomplete Adjuvant for a subsequent four injections at ~14 day intervals, with terminal bleeding on day 63. All sera were stored at -20°C until needed.
Lysis kinetics of freeze-dried Micrococcus luteus cells (ATCC 4698) were performed as previously described [25]. M. luteus cells were exposed to 2 μg/ml human lysozyme (Sigma) in the absence or presence of increasing concentrations of PBS-diluted r1063. Bacterial lysis was measured by the change in optical density (OD595) at 25 °C over time, using a spectrophotometer (microplate reader). Pooled human tears (filter sterilized, LEE Biosolutions) were diluted to 0.1X in H2O and pretreated with 200 μg/mL recombinant 1981 (r1981) [25], or vehicle (H2O), for 20 min at 37°C. Pooled human saliva (Lee Biosolutions) was similarly diluted to 0.5X with 125 μg/mL r1981, or vehicle (H2O), for 20 min at 37°C. M. luteus, which had been prepared as previously described [22], was exposed to pretreated secretions for a final concentration of 0.01X tears (20 μg/mL r1981) and 0.05X saliva (12.5 μg/mL r1981). Bacterial lysis was measured over time at OD450 at 37°C using a Victor3 multilabel plate reader (Perkin-Elmer).
Mid-logarithmic phase Gc was suspended in 0.5X GCBL (diluted with H2O; Kellogg’s supplements and NaHCO3 were also at 0.5X), with IPTG and AT, if necessary, prior to exposure to the antimicrobial protein at 37°C with 5% CO2 (v/v). The final concentration of Gc with lysozyme was 5x105 CFU/mL, unless otherwise indicated. Gc survival is expressed as the percent of Gc surviving after exposure to the antimicrobial divided by the percent of Gc surviving in vehicle, and normalized to the vehicle control (= 100%).
Human Lysozyme: Human lysozyme was prepared as in [22]. Gc was incubated with lysozyme for 3 hr unless otherwise indicated. Lysozyme with inactivated hydrolase activity was prepared by boiling as in [22]. Ethylenediaminetetraacetic acid (EDTA) treatment with lysozyme was performed as previously described [22] where the final concentration of Gc with EDTA and lysozyme was 5x107 CFU/mL.
For lysozyme pretreatment with recombinant proteins, vehicle (PBS), lysozyme alone (3 μg/mL), lysozyme + r1063 (3 μg/mL lysozyme + 1.5 μg/mL r1063), lysozyme + rMIP [25, 59] (3 μg/mL lysozyme + 1.5 μg/mL rMIP), and lysozyme + r1981 (3 μg/mL lysozyme + 1.9 μg/mL r1981) were pretreated for 20 min at 37°C. Because r1063 was solubilized in a PBS buffer with SDS, SDS was also added for an equivalent final concentration for all conditions. Gc were exposed to pretreated samples for 5 hr at the following final concentrations: vehicle (PBS), lysozyme + r1063 (1.5 μg/mL lysozyme + 0.75 μg/mL r1063), lysozyme + rMIP (1.5 μg/mL lysozyme + 0.75 μg/mL rMIP), and lysozyme + r1981 (3 μg/mL lysozyme + 0.97 μg/mL r1981).
For mixed Gc experiments, Δ1981Δ1063 mutant Gc was mixed with Δ1981Δ1063::1981(WT)-FLAG complemented Gc at a ratio of 1:1 for a total of 5 x 105 CFU/mL final. Gc were incubated together or separately for 1 hr prior to exposure to lysozyme for 3 hr. In mixed infection, Δ1981Δ1063 mutant Gc was differentiated from Δ1981Δ1063::1981(WT)-FLAG based on chloramphenicol resistance.
LL-37: For the LL-37 antimicrobial assay, Gc was incubated with LL-37 (from Dr. William Shafer, Emory University) as described [22]. For LL-37 pretreatment, 5x105 CFU/mL Gc was pretreated with 0.4 μg/mL LL-37 for 25 min at 37°C. The bacteria were centrifuged, supernatant removed, and the bacterial pellet resuspended in equivalent volume of 0.5X GCBL, prior to exposure to human lysozyme for 3 hr.
Chicken Egg White lysozyme: Chicken Egg White lysozyme (Sigma) was reconstituted in 10 mM Tris-HCl, pH 8.0, and incubated with Gc for 3 hr at 1,000 μg/mL.
Mutanolysin: Gc was incubated with mutanolysin (Sigma) as previously described [22].
Human Secretions: Pooled human tears and pooled human saliva were pretreated with vehicle or r1981 exactly as in the M. luteus assay above. Gc was incubated with untreated or treated secretions for 3 hr at the indicated final concentration.
Gc was exposed to vancomycin Etests (bioMérieux) as described [22] except that Gc was first grown to mid-logarithmic phase in liquid culture, and, if needed, induced with IPTG or AT for 5 hr prior to testing.
Mid-logarithmic phase Gc (7.5x108 CFU of Δ1981Δ1063, ΔltgA::ltgA-FLAG, ΔltgD::ltgD-FLAG, Δ1981Δ1063::1981(WT)-FLAG, Δ1981Δ1063::1063(WT)-FLAG, and Δ1981Δ1063::1063(S83A)-FLAG, Δ1981Δ1063::1981(K103A)-FLAG) in 0.5X GCBL with IPTG were exposed to vehicle or 1,000 μg/mL human lysozyme for 3 hr at 37°C with 5% CO2. Gc was subsequently centrifuged, washed once with PBS, and resuspended in 1 mL ice cold lysis buffer (1% v/v Triton-X-100, 150mM NaCl, 2mM EDTA, 20mM Tris-HCl pH 7.5, 1X v/v protease inhibitors (Calbiochem Set V)) for 30 min with end-over-end rotation at 4°C. Insoluble debris and unlysed Gc were pelleted at 10,000xg for 15 min at 4°C. The supernatant (“whole cell lysate”) was removed and incubated with lysis buffer-equilibrated M2 FLAG Affinity Gel (Sigma) for 2 hr with end-over-end rotation at 4°C. The resin was washed six times with ice cold wash buffer (1% v/v Triton-X-100, 150mM NaCl, 2mM EDTA, 20mM Tris-HCl pH 7.5), and subsequently resuspended in 1X sample buffer. Western blots were probed with anti-lysozyme (Abcam 108508), stripped (200mM glycine, 0.1% w/v SDS, 1% v/v Tween-20), and reprobed with anti-FLAG (M2, Sigma).
For sublethal conditions of lysozyme, 7.5x107 CFU/mL (as opposed to 5x105 CFU/mL used for antimicrobial assays) of Gc, which had been grown to mid-logarithmic phase, was exposed to vehicle or increasing concentrations of human lysozyme in 0.5X GCBL at 37°C with 5% CO2 (v/v) for 3 hr. Gc survival was assessed at 0 hr and 3 hr under these conditions, as described above. Otherwise, Gc exposed to vehicle or lysozyme were centrifuged, supernatant removed, and pellet resuspended in sample buffer (12 mM Tris-HCl pH 6.8, 0.4% w/v SDS, 5% v/v Glycerol) without β-mercaptoethanol (BME) or bromophenol blue. Protein concentration was determined by BCA™ Protein Assay, followed by addition of BME and bromophenol blue. 5 μg of total protein was separated by SDS-PAGE. Rabbit anti-r1981 antisera and mouse anti-FLAG antibody (M2, Sigma) were used to visualize native Ng_1981 (from WT Gc) and Ng_1063-FLAG (from Gc with 1063(WT)-FLAG at its native locus) protein, respectively. Blots were stripped and re-probed with rabbit anti-Zwf antibody to confirm equivalent loading. ImageJ was used to measure the percent area for each protein band. The relative band density at each concentration of lysozyme was determined by dividing the vehicle-normalized band density of Ng_1981/Ng_1063-FLAG protein by the vehicle-normalized band density of Zwf protein. Vehicle-normalized band densities were determined by dividing the band density of Ng_1981/Ng_1063-FLAG/Zwf at one concentration of lysozyme by the band density for the corresponding vehicle-treated sample.
Neutrophils were isolated from the venous blood of healthy human subjects as described [60], and used within 2 hr of isolation. Neutrophils were adhered to plastic coverslips in the presence of 10 nM human Interleukin 8 (R&D) in Roswell Park Memorial Institute 1640 medium (RPMI) with 10% v/v FBS at 37°C with 5% CO2 (v/v) for at least 30 min prior to infection. Mid-logarithmic phase Gc at a multiplicity of infection of 1 was exposed to neutrophils in a synchronous manner, as previously described [60].
Gc (WT, Δ1063::1063(WT)-FLAG, and Δ1981::1981(WT)-FLAG) was grown to mid-logarithmic phase in the presence of IPTG to induce protein expression. Gc (7.5x107 CFU) was centrifuged, resuspended in 30 μg/mL 5-(and-6)-carboxylfluorescein diacetate, succinimidyl ester (CFSE) in Dulbecco’s phosphate-buffered saline (DPBS) + 5 mM MgSO4, and incubated for 15 min at 37°C. Bacteria were washed once with DPBS + 5 mM MgSO4 and subsequently resuspended in RPMI (no phenol red) with 10% v/v FBS for 10 min at 37°C. Bacteria were centrifuged, resuspended in 8 μg/mL anti-FLAG (M2, sigma) in 100 μL RPMI (no phenol red) with 10% v/v FBS, and incubated for 30 min at 37°C. Bacteria were washed twice with RPMI with 10% v/v FBS, resuspended in 100 μL of 1:400 goat anti-mouse coupled to Alexa Fluor 647 (Thermo) in RPMI (no phenol red) with 10% v/v FBS, and incubated for 30 min at 37°C. Bacteria were washed twice with RPMI with 10% v/v FBS and resuspeded in 2% w/v paraformaldehyde. Data were acquired with an ImageStreamX Mark II cytometer operated by INSPIRE software (Amnis) and analyzed using IDEAS Application v6.2 software (Amnis). For analysis, focused cells were gated as determined by a high gradient root mean square for image sharpness. From the focused cells, single cells were gated as defined by a high intensity of CFSE and low side scatter. Focused, single cells were used for analysis of intensity of anti-FLAG/AlexaFluor647.
1 mL of induced, mid-logarithmic phase Gc (Δ1063:1063(WT)-FLAG and Δ1981:1981(WT)-FLAG) was centrifuged and processed as described [37]. For cells that were not permeabilized, PBS with 5% v/v normal goat serum was used as a blocking agent. For FLAG staining, the anti-FLAG antibody was used at 27 μg/mL in PBS with 5% v/v normal goat serum, and goat anti-mouse coupled to Alexa Fluor 647 (Thermo) was used at 1:400 in PBS with 5% v/v normal goat serum. For 1981 staining with antisera, rabbit anti-r1981 antisera was diluted 1:100 in PBS with 5% v/v normal goat serum, and goat anti-rabbit coupled to Alexa Fluor 647 (Thermo) was used at 1:400 in PBS with 5% v/v normal goat serum. Bacterial DNA was stained with 18 μM DAPI (Sigma). Cells were visualized using a Nikon E800 with Hamamatsu Orca-ER camera using Nikon Elements software and processed using Adobe Photoshop CS3.
Using 1063(WT)-FLAG native Gc, 7.5x107 CFU of mid-logarithmic phase Gc was centrifuged and resuspended in 1 mL Hanks’ balanced salt solution (HBSS) with 10 mM 4-(2-hydroxyethyl)-1-piperazineethanesulfonic acid and 5 mM sodium bicarbonate. Bacteria were incubated at 37°C with 5% CO2 (v/v) for 3 hr and subsequently centrifuged. Supernatants were collected from the bacterial pellet, passed through a 0.2 μm filter, concentrated to a ~60 μL volume using a 3 kDa centrifugal filter unit (Amicon), and brought up to a final volume of 100 μL with sample buffer. Bacterial pellets were washed once with 1X PBS and subsequently lysed (“whole cell lysates”) with 100 μL sample buffer. Equivalent volumes were loaded in an SDS-PAGE gel, and gel transferred for Western blot analysis. Blots were probed with rabbit anti-r1981 antisera to detect native Ng_1981 protein, and then stripped and reprobed with mouse anti-FLAG (M2, sigma) to detect native Ng_1063 protein. Blots were also probed with rabbit anti-Zwf (Aleksandra Sikora, Oregon State University [61]) as a cell lysate loading control.
Experimental values presented display the mean ± the standard error of the mean (SEM) of at least three independent replicates. For experiments using neutrophils, at least 3 independent donors were used. Unless otherwise indicated, a two tailed student’s t-test was performed, and significance was determined as a p-value less than 0.05.
|
10.1371/journal.pntd.0002181 | Characterization of Rift Valley Fever Virus MP-12 Strain Encoding NSs of Punta Toro Virus or Sandfly Fever Sicilian Virus | Rift Valley fever virus (RVFV; genus Phlebovirus, family Bunyaviridae) is a mosquito-borne zoonotic pathogen which can cause hemorrhagic fever, neurological disorders or blindness in humans, and a high rate of abortion in ruminants. MP-12 strain, a live-attenuated candidate vaccine, is attenuated in the M- and L-segments, but the S-segment retains the virulent phenotype. MP-12 was manufactured as an Investigational New Drug vaccine by using MRC-5 cells and encodes a functional NSs gene, the major virulence factor of RVFV which 1) induces a shutoff of the host transcription, 2) inhibits interferon (IFN)-β promoter activation, and 3) promotes the degradation of dsRNA-dependent protein kinase (PKR). MP-12 lacks a marker for differentiation of infected from vaccinated animals (DIVA). Although MP-12 lacking NSs works for DIVA, it does not replicate efficiently in type-I IFN-competent MRC-5 cells, while the use of type-I IFN-incompetent cells may negatively affect its genetic stability. To generate modified MP-12 vaccine candidates encoding a DIVA marker, while still replicating efficiently in MRC-5 cells, we generated recombinant MP-12 encoding Punta Toro virus Adames strain NSs (rMP12-PTNSs) or Sandfly fever Sicilian virus NSs (rMP12-SFSNSs) in place of MP-12 NSs. We have demonstrated that those recombinant MP-12 viruses inhibit IFN-β mRNA synthesis, yet do not promote the degradation of PKR. The rMP12-PTNSs, but not rMP12-SFSNSs, replicated more efficiently than recombinant MP-12 lacking NSs in MRC-5 cells. Mice vaccinated with rMP12-PTNSs or rMP12-SFSNSs induced neutralizing antibodies at a level equivalent to those vaccinated with MP-12, and were efficiently protected from wild-type RVFV challenge. The rMP12-PTNSs and rMP12-SFSNSs did not induce antibodies cross-reactive to anti-RVFV NSs antibody and are therefore applicable to DIVA. Thus, rMP12-PTNSs is highly efficacious, replicates efficiently in MRC-5 cells, and encodes a DIVA marker, all of which are important for vaccine development for Rift Valley fever.
| Upon outbreak of zoonotic viral diseases in herds of animals, early detection of naturally infected animals and prevention of further viral spread are important for minimizing the impact of outbreak in the society. Vaccination may compromise the identification of infected animals since both natural infection and vaccination induce antibodies specific to the pathogen. Therefore, new generation vaccines should have a marker to differentiate infected from vaccinated animals (DIVA). Rift Valley fever virus (RVFV) is a mosquito-borne zoonotic pathogen which can cause hemorrhagic fever, neurological disorders or blindness in humans and a high-rate abortion in ruminants. MP-12 strain, a live-attenuated candidate vaccine, is safe and immunogenic, but lacks a DIVA marker. In this study, we developed and characterized improved MP-12 viruses which encode a DIVA marker by replacing the virulence gene with that of serologically distinct viruses belonging to the same genera. The novel MP-12 variant with such DIVA marker was highly efficacious and replicated efficiently in human diploid cells for vaccine production, and will become alternative candidate vaccines of MP-12 for veterinary applications.
| Rift Valley fever virus (RVFV), which belongs to the family Bunyaviridae, genus Phlebovirus, is a zoonotic pathogen transmitted by mosquitoes, and the causative agent for Rift Valley fever (RVF). RVF is characterized by a high rate of abortion and fetal malformation in pregnant ruminants, febrile illness in adult ruminants, and lethal acute hepatitis in newborn lambs [1]. In humans, patients suffer an acute febrile illness with occasional complications including partial or complete blindness, hemorrhagic fever, or neurological disorders [2], [3], [4], [5]. RVFV can be transmitted through the drought-resistant eggs of infected floodwater Aedes mosquitoes which thus play a role in maintaining RVFV in endemic areas. Other mosquito species are also involved in RVFV transmission if RVFV-infected mosquito population increases subsequent to heavy rain fall or increase in mosquito habitats [6], [7], [8], [9]. RVF has been endemic to sub-Saharan Africa, and has spread into Madagascar, Comoro, Egypt, Saudi Arabia and Yemen [10], [11], [12], [13], [14], [15], [16], [17]. The development of an effective vaccine against RVF is important for non-endemic countries to prevent further spread of RVFV. RVFV is classified as an NIAID Category A Priority Pathogen and an overlap select agent by the U.S. Department of Health and Human Services (HHS) and Agriculture (USDA) [18]. RVFV is transmitted via aerosol, and the handling of virus should be done in biosafety level (BSL) 3+ or 4 laboratories.
RVFV has a tripartite negative-stranded RNA genome composed of Small (S)-, Medium (M)- and Large (L)-segments. The S-segment encodes N and NSs genes in an ambi-sense manner, the M-segment contains a single open reading frame (ORF) which encodes NSm, 78-kD protein, NSm-Gn, Gn, and Gc proteins from different AUGs and co-translational cleavage; and the L-segment encodes RNA-dependent RNA polymerase [19], [20], [21], [22]. The nonstructural protein, NSs is a major virulence factor, and it inhibits host general transcription by inhibiting host basal transcription factor (TF) IIH; TFIIH is one of six general transcription factors (TFIIA, TFIIB, TFIID, TFIIE, TFIIF and TFIIH) [23], composed of 10 different proteins; i.e., XPD, XPB, p8, p34, p44, p52, p62, MAT1, cyclin H and cdk7 [24], [25], and essential for RNA synthesis by cellular RNA polymerase I and II [26], [27]. NSs binds to and sequester TFIIH p44 [28] and also promotes the degradation of TFIIH p62 [29]. On the other hand, NSs inhibits the activation of interferon (IFN)-β promoter by interacting with Sin3A-associated protein (SAP30) and recruiting repressor complex containing histone deacetylase-3 (HDAC-3) [30], [31], while NSs promotes the degradation of dsRNA-dependent protein kinase, PKR [32], [33], [34].
Smithburn vaccine generated by mouse brain passages of RVFV Entebbe strain has been sold as a veterinary vaccine in South Africa, Kenya, Zimbabwe, Namibia, Egypt and Israel [11]. This vaccine itself caused abortion in pregnant ruminants and also reassorted with natural wild-type (wt) RVFV due to the use during an outbreaks [11]. MP-12 vaccine was generated by 12 serial plaque-passages in human diploid MRC-5 cells in the presence of the chemical mutagen 5-fluorouracil [35], [36]. MP-12 is highly immunogenic in ruminants, and can also induce sufficient immune response in humans [37], [38], [39], [40], [41]. MP-12 is excluded from the select agent rule in the U.S., and can be handled in BSL-2 laboratories. MP-12 vaccine was manufactured by using certified MRC-5 cells as an Investigational New Drug for human clinical trials [36]. It is known that RVFV causes spontaneous truncation of NSs gene during passages in mammalian Vero or BHK-21 cells which lack functional type-I IFN system [42] or in Aedes aegypti larvae (Aag2) cells [43], [44]. We recently characterized the genetic subpopulations of MP-12 vaccine Lot 7-2-88 and found that MP-12 vaccine retains highly stable attenuation mutations in the M- and L-segments during its cultivation in MRC-5 cells [36]. Since MP-12 is attenuated by only point mutations at M- and L-segments, a potential of reversion to virulence cannot be excluded, and MP-12 requires further improvement for veterinary use. Another concern is a lack of DIVA (Differentiating Infected from Vaccinated Animals) markers. In previous study, we found 27% of mice vaccinated with MP-12 induce detectable anti-NSs antibody [45]. Though the immunogenicity of MP-12 NSs is poor, the presence of anti-NSs antibody in vaccinated group will compromise DIVA strategy. Without DIVA markers, it is impossible to monitor infected animals in herds of vaccinated ruminants during RVF outbreaks.
In this study, we aimed to develop MP-12 variants which encode a DIVA marker and replicate efficiently in MRC-5 cells. Although recombinant MP-12 lacking NSs gene in the S-segment encodes a negative DIVA marker [45]; i.e., a lack of anti-NSs antibody response, it does not replicate efficiently in type-I IFN-competent MRC-5 cells. The Phlebovirus genus consists of the sandfly fever group including serologically distinct Punta Toro serocomplex, Naples serocomplex, Icoaraci serocomplex, Frijoles serocomplex, Sicilian serocomplex, RVFV, and the Uukuniemi group [46], and some of the different phlebovirus NSs are reported to be able to interfere with the host type-I IFN system [34], [47], [48]. In this study, we developed recombinant MP-12 encoding NSs of Punta Toro virus Adames strain (PTV) (rMP12-PTNSs) or Sandfly fever Sicilian virus (SFSV) (rMP12-SFSNSs) in place of MP-12 NSs [46]. It should be noted that a lack of MP-12 NSs serves as a negative DIVA marker to identify animals exposed to wt RVFV, and we did not attempt to identify vaccinated animals by detecting antibody specific to anti-PTV NSs or anti-SFSV NSs. We characterized the functions of those NSs, and determined the immunogenicity and efficacy of rMP12-PTNSs and rMP12-SFSNSs in the outbred CD1 mouse model. Our results suggested that rMP12-PTNSs, but not rMP12-SFSNSs, replicates efficiently in MRC-5 cells, while both are as efficacious as parental MP-12, and also did not induce antibodies cross-reactive to RVFV NSs. Thus, rMP12-PTNSs is an alternative candidate vaccine which can be amplified in MRC-5 cells and encodes a negative DIVA marker.
VeroE6 cells (ATCC CRL-1586), 293 cells (ATCC CRL-1573), MRC-5 cells (ATCC CCL-171) and MEF cells [49] were maintained in Dulbecco's modified minimum essential medium (DMEM) containing 10% fetal calf serum (FCS). BHK/T7-9 cells that stably express T7 RNA polymerase [50] were maintained in MEM-alpha containing 10% FCS with 600 µg/ml of hygromycin. Penicillin (100 U/ml) and streptomycin (100 µg/ml) were added to the culture media. MP-12 vaccine Lot 7-2-88 (kindly provided from Dr. J.C. Morrill at the University of Texas Medical Branch: UTMB) was amplified twice in MRC-5 cells for experiments. rMP12-PTNSs, rMP12-PTNSs-Flag, rMP12-SFSNSs and rMP12-SFSNSs-Flag were rescued from plasmid DNAs in BHK/T7-9 cells as described previously [51], and passaged once in VeroE6 cells. rMP12-NSsR173A and rMP12-NSs-Flag were reported previously [32], [52]. RVFV ZH501 strain stock was generated after one VeroE6 cell passage of an original ZH501 reference collection vial (Serial #JM1137) at UTMB [53]. Sendai virus Cantell strain was purchased from Charles River (North Franklin, CT).
The plasmid encoding anti-viral-sense of MP-12 S-segment at the downstream of the T7 promoter, pProT7-S(+), was described previously [51]. VeroE6 cells were infected with PTV Adames strain or SFSV Sabin strain (provided by Dr. R.B. Tesh, UTMB), and the total RNA was extracted at 3 dpi. First stranded cDNA was synthesized with Superscript II Reverse Transcriptase (Invitrogen), and the NSs ORF was amplified with Phusion DNA polymerase (New England Biolabs) by using specific primers with uniquely incorporated HpaI and SpeI restriction sites for cloning. The PCR fragments of PTV Adames or SFSV NSs were cloned into pProT7-S(+) [51] in place of MP-12 NSs, designated as pProT7-S(+)PTNSs or pProT7-S(+)SFSNSs, respectively. Similarly, Flag-tag was added at the C-terminus of those NSs, and the resulting plasmids were designated as pProT7-S(+)PTNSs-Flag or pProT7-S(+)SFSNSs-Flag, respectively. The pcDNA3.1mycHisA plasmids encoding CAT [32] or NSs of PTV Adames or SFSV NSs (without tag) were generated, and designated as pcDNA3.1mycHisA-PTNSs or pcDNA3.1mycHisA-SFSVNSs. For reporter assay, IFNb-pGL3 plasmid was kindly provided by Dr. R. Lin at McGill Univ. [54], 4×IRF3-luc plasmid was kindly provided by Dr. S. Ludwig at ZMBE, Westfälische-Wilhelms-University [55], and pPRDII-luc plasmid was kindly provided by Dr. M. Gale Jr. at Univ. of Washington [56]. The pRL-SV40 plasmid was purchased from Promega (Madison, WI).
The recombinant MP-12 encoding NSs truncation or mutation were recovered by using a plasmid combination of pProT7-M(+), pProT7-L(+), pT7-IRES-vN, pT7-IRES-vL, pCAGGS-vG and either of pProT7-S(+)PTNSs or pProT7-S(+)SFSNSs. BHK/T7-9 cells were transfected with those plasmids as described previously [51]. Recombinant MP-12 encoding NSs of PTV Adames or SFSV were designated as rMP12-PTNSs or rMP12-SFSNSs, respectively. Recovered recombinant MP-12 were amplified once in VeroE6 cells, titrated by plaque assay, and used for experiments. In addition, we also generated recombinant MP-12 encoding C-terminus Flag-tagged NSs of PTV Adames or SFSV, and those mutants were designated as rMP12-PTNSs-Flag or rMP12-SFSNSs-Flag, respectively.
Total RNA was extracted from infected or mock-infected cells using TRIzol reagent. Denatured RNA was separated on 1% denaturing agarose-formaldehyde gels and transferred onto a nylon membrane (Roche Applied Science, Indianapolis, IN). Northern blot analysis was performed as described previously with strand-specific RNA probes to detect RVFV anti-sense S-segment/N mRNA, mouse IFN-β mRNA, or mouse ISG56 mRNA [57].
Western blot analysis was performed as described previously [32]. The membranes were incubated with anti-human PKR monoclonal antibody (BD Biosciences), anti-mouse PKR monoclonal antibody (B-10, Santa Cruz, CA), anti-RVFV mouse polyclonal antibody (a kind gift from Dr. R.B.Tesh, UTMB), anti-Flag-tag M2 monoclonal antibody (Sigma), or anti-β-actin goat polyclonal antibody (I-19; Santa Cruz, CA.) overnight at 4°C and with secondary antibodies (Santa Cruz, CA) for 1 hr at room temperature.
MRC-5 cells were infected with MP-12, rMP12-C13type, rMP12-PTNSs or rMP12-SFSNSs at a m.o.i of 0.01 at 37°C for 1 h, washed cells twice with media, and incubated at 37°C. Culture supernatants were collected at 0 (after removal of viral inocula), 24, 48, 72 and 96 hpi, and used for plaque assay [51], [58]. Viral titers in culture supernatants from VeroE6 cells or MEF cells infected with those viruses at a m.o.i of 0.01, which were collected at 72 hpi, were also titrated.
The pcDNA3.1mycHisA plasmids encoding CAT [32] or pcDNA3.1mycHisA-PTNSs or pcDNA3.1mycHisA-SFSVNSs were linearized, and in vitro transcribed by using mMESSAGE mMACHINE T7 Ultra kit (Ambion, Grand Island, NY) according to the manufacturer's instructions. The linearized CAT DNA contained myc-His tag at the 3′end.
Transfection of plasmid DNA or in vitro synthesized RNA was performed by using TransIT-LT1 or TransIT-mRNA Transfection Kit (Mirus, Madison, WI) according to manufacturer's instructions, respectively.
293 cells in 12-well plate were transfected with 1 µg of IFNb-pGL3 plasmids, 4×IRF3-luc plasmid, or pPRDII-luc plasmids in addition to 0.1 µg of pRL-SV40 plasmid. At 24 hours post transfection, cells were mock-infected or infected with 300 µl of SeV (100 U/ml), and mock-transfected or immediately transfected with 1 µg of in vitro synthesized RNA encoding CAT (control) or NSs of MP-12, PTV Adames strain or SFSV. Cells were collected at 16 hpi, and the relative luciferase activity was measured by Dual-Luciferase Reporter Assay System (Promega, Madison, WI) according to manufacturer's instructions.
The analysis of host general transcription suppression was described previously [29]. Briefly, 293 cells were mock-infected or infected with either MP-12, rMP12-PTNSs or rMP12-SFSNSs at a m.o.i of 3. Cells were treated with 0.5 mM 5-ethynyluridine (EU) from 12 to 13 hpi before harvesting at 13 hpi. As a control for transcriptional suppression, cells were treated with 5 µg/ml of ActD concurrently with the EU treatment. Incorporated EU was detected with an AlexaFluor 647-coupled azide (Invitrogen), and viral antigens were stained with anti-RVFV antibodies followed by an AlexaFluor 488-coupled secondary antibody. Cells were analyzed by flow cytometry on an LSRII Fortessa instrument (BD Biosciences).
For testing the efficacy of MP-12 NSs mutants, 5-week-old female CD1 outbred mice (Charles River, North Franklin, CT) were inoculated subcutaneously with PBS (mock) (n = 10) or 1×105 pfu of MP-12 (n = 20), rMP12-NSR173A (n = 10), rMP12-PTNSs (n = 9) or rMP12-SFSNSs (n = 10). Those mice were challenged with 1×103 pfu of wt RVFV ZH501 strain (i.p) at 45 days post vaccination. The challenge experiment was performed at an animal biosafety level 4 facility at the UTMB Shope laboratory. Mice were observed for 21 days after challenge, and the body weight was monitored daily. Sera were collected at 1, 2, 3, and 42 days post vaccination (retro-orbital bleeding), and at 21 days post wt RVFV challenge (cardiac puncture). Survival curves of mice (Kaplan-Meyer method) were analyzed by Graphpad Prism 5.03 program (Graphpad Software Inc, La Jolla, CA.).
The PRNT80 was determined as described previously [45]. Briefly, Each 20-µl of mouse sera serially diluted 4-fold was transferred into flat-bottom 96-well plates containing 5 µl of MP-12 virus (50 pfu/well) (final dilutions of sera: 1∶10, 1∶40, 1∶160∼). After incubation at 37°C for 1 hour, 150 µl of DMEM with 10% FBS was added to the well. The 150 µl of the mixture was transferred into a 24-well plate with confluent VeroE6 cells, and the plate was incubated at 37°C for 1 hour. After removal of inocula, virus titer was determined by plaque assay. Eighty % of the average number of plaques in 6 different wells that had mock-immunized mice sera was used as the cut-off number (typically 8∼9). The highest dilution of sera that produced the number of plaques below the cut-off number was designated as the PRNT80 neutralizing antibody titer.
His-tagged RVFV N proteins [45], which were expressed and purified using recombinant baculovirus, or purified GST-tagged RVFV NSs C-terminal 48 amino acids [45], which were expressed and purified using E,coli, were coated onto 96-well ELISA plates overnight at 4°C at a concentration of 100 ng/well. After washing 3 times with PBS containing 0.1% Tween 20 (PBS-T), the wells were blocked with PBS-T containing 5% skim milk at 37°C for 2 hours. Then, wells were incubated with serum samples (2-fold dilutions for anti-N IgG, and 1∶100 for anti-NSs IgG) at 37°C for 1 hour. Wells were washed for 3 times with PBS-T and reacted with HRP-conjugated anti-mouse IgG (Santa Cruz, CA) at 37°C for 1 hour. After washing with PBS-T for 3 times, ABTS was added to wells. The plate was incubated at room temperature for 30 min, and the optical density (OD) at 405 nm was recorded. The cut-off value, 0.176, was defined as mean+2×standard deviation of 24 normal mouse serum samples (1∶400) for anti-N IgG, while the cut-off value of 0.204 was defined as mean+3×standard deviation of 24 normal mouse serum samples. The highest dilution of sera that made a OD value larger than the cut-off was designated as the anti-N antibody titer. Because the anti-NSs antibody level was low, an OD value of a 1∶100 dilution was used for demonstrating the presence of anti-NSs antibody.
Statistical analyses were performed by using the Graphpad Prism 5.03 program (Graphpad Software Inc, La Jolla, CA). Unpaired t-test or Mann-Whitney U-test was used for the comparison of two groups. Survival curves of mice were analyzed by log-rank (Mantel-Cox) test.
Mouse studies were performed in facilities accredited by the Association for Assessment and Accreditation of Laboratory Animal Care (AAALAC) in accordance to the Animal Welfare Act, NIH guidelines, and US federal law. The animal protocol was approved by the University of Texas Medical Branch (UTMB) Institutional Animal Care and Use Committee (IACUC) (protocol #1007038). All the recombinant DNA and RVFV were created upon the approval of the Notification of Use by the Institutional Biosafety Committee at UTMB. The wt RVFV ZH501 strain was used at the Robert E. Shope BSL4 laboratory at the UTMB in accordance with NIH guidelines and US federal law.
In this study, we aimed to develop a modified MP-12 vaccine encoding NSs derived from serologically distinct phleboviruses. We attempted to select the phlebovirus NSs which can inhibit type-I IFN induction. Using reverse genetics for RVFV MP-12 strain, we recovered recombinant MP-12 encoding NSs of PTV Adames strain (rMP12-PTNSs) or SFSV (rMP12-SFSNSs) (Fig. 1A). In VeroE6 cells, rMP12-PTNSs formed clear plaques similar to those of MP-12, while rMP12-SFSNSs formed turbid plaques similar to those of rMP12-C13type [51], [58] (Fig. 1B). PTV Adames strain NSs [48] and SFSV NSs inhibit IFN-β gene [34], while SFSV NSs does not promote the degradation of PKR [34].
We first tested the replication capability of those viruses in type-I IFN incompetent VeroE6 cells and mouse embryonic fibroblast (MEF) cells (Fig. 2A and B). Cells were infected with the indicated virus at a multiplicity of infection (m.o.i) of 0.01, and culture supernatants were collected at 72 hpi for viral titration. In VeroE6 cells, MP-12 and rMP12-PTNSs replicated to a similar level, while rMP12-C13type and rMP12-SFSNSs replicated slightly more efficiently than MP-12 (Fig. 2A). In MEF cells at 72 hpi, MP-12 replicated 1.51, 0.99 or 0.81 log more than rMP12-C13type, rMP12-PTNSs or rMP12-SFSNSs, respectively (Fig. 2B). Subsequently, we determined viral growth kinetics in MRC-5 cells, because MRC-5 cells were used for the cultivation of MP-12 vaccine [36]. In MRC-5 cells, rMP12-SFSNSs did not replicate efficiently, and the replication kinetics was similar to that of rMP12-C13type, while rMP12-PTNSs replicated efficiently at the level nearly similar to that of MP-12 (Fig. 2C). At 72 hpi, MP-12 replicated 2.27, 0.54 or 2.14 log more than rMP12-C13type, rMP12-PTNSs or rMP12-SFSNSs, respectively (Fig. 2C).
RVFV NSs promotes the degradation of PKR thus inhibiting the phosphorylation of eIF2α, which promotes viral protein synthesis [32], [52]. To understand the functional difference between RVFV NSs and other phlebovirus NSs, we first tested whether they degraded PKR. To avoid the up-regulation of PKR by type-I IFN, we used type-I IFN-incompetent VeroE6 cells [59], [60]. VeroE6 cells were mock-infected or infected with MP-12, rMP12-C13type, rMP12-PTNSs or rMP12-SFSNSs at a m.o.i of 3 and cells were collected at 16 hpi for Western blot analysis (Fig. 3A). As expected, cells infected with MP-12 promoted the degradation of PKR, while cells infected with MP-12 encoding NSs of PTV or SFSV expressed abundant PKR at 16 hpi. The NSs accumulation of rMP12-PTNSs or rMP12-SFSNSs could not be detected by mouse polyclonal antibodies against PTV or SFSV probably due to the sensitivity of the antibody to detect NSs (data not shown). To evaluate the level of each NSs accumulation, NSs of PTV or SFSV were fused to Flag-tag at the C-terminus, and the resulting viruses were designated as rMP12-PTNSs-Flag or rMP12-SFSNSs-Flag, respectively. As a control for NSs-Flag expression, we also used rMP12-NSs-Flag, which encodes an MP-12 NSs with a C-terminus Flag-tag [32]. VeroE6 cells were mock-infected or infected with rMP12-NSs-Flag, rMP12-C13type, rMP12-PTNSs-Flag or rMP12-SFSNSs-Flag at a m.o.i of 3, collected at 16 hpi, and subjected to Western blot analysis using anti-Flag antibody. As shown in Fig. 3B, we confirmed that NSs of PTV Adames and SFSV were expressed at 16 hpi. Collectively, we concluded that the rMP12-PTNSs and rMP12-SFSNSs do not promote the degradation of PKR.
We found that both the rMP12-PTNSs and rMP12-SFSNSs do not promote the degradation of PKR. To clarify whether the rMP12-PTNSs or rMP12-SFSNSs can induce host general transcription suppression, 293 cells were mock-infected or infected with MP-12, rMP12-PTNSs or rMP12-SFSNSs at a m.o.i of 3, and nascent RNA was labeled with 5-ethynyluridine (EU) [61], a uridine analog, from 12 to 13 hpi. As a control for transcriptional suppression, mock-infected cells were incubated with actinomycin D (ActD) (5 µg/ml) concurrently with the EU treatment. The incorporated EU was covalently linked to azide conjugated with AlexaFluor 647, and cells were further stained with anti-RVFV antibody to detect expression of viral proteins [29]. Then, the level of EU-incorporation and expression of viral proteins were analyzed by flow cytometry. Fig. 4A depicts the acquired data as a dot plot with the expression of viral proteins (anti-RVFV) on the x-axis and the incorporation of EU (RNA) on the y-axis. Quadrant gates were set to that the majority of mock infected cells (79.5%) were in the upper left quadrant, the majority of ActD treated cells (92.6%) in the lower left quadrant and the majority of anti-RVFV positive cells in either the right upper or lower quadrant. When cells were infected with MP-12, 95.2% of total cells (or 98.1% of anti-RVFV positive cells) showed reduced EU incorporation when compared to mock infected cells. Similarly, 98.99% of anti-RVFV positive cells showed reduced EU incorporation when cells were infected with rMP12-PTNSs. In contrast, when cells were infected with rMP12-SFSNSs, 78.84% of anti-RVFV positive cells still incorporated EU at the same level as mock infected cells. Fig. 4B depicts the level of EU incorporation of the anti-RVFV negative population (mock and ActD) and anti-RVFV positive population (MP-12, rMP12-PTNSs and rMP12-SFSNSs infected cells) as a histogram where the RNA fluorescence intensity (EU incorporation) is plotted on the x-axis and the cell count is plotted on the y-axis. These data suggest that MP-12 and rMP12-PTNSs are able to suppress host transcription, whereas rMP12-SFSNSs has no negative effect on host RNA synthesis.
We found that rMP12-PTNSs but not rMP12-SFSNSs induces host general transcription suppression. To know whether rMP12-PTNSs or rMP12-SFSNSs can inhibit IFN-β gene up-regulation, type-I IFN-competent MEF cells were mock-infected or infected with MP-12, rMP12-C13type, rMP12-PTNSs or rMP12-SFSNSs at a m.o.i of 3, and total RNA was extracted at 7 hpi. Accumulation of mouse IFN-β, ISG56 mRNA, RVFV antiviral-sense S-segment RNA and N mRNA was analyzed by Northern blot as described previously [51], [62]. Cells infected with rMP12-C13type induced IFN-β and ISG56 mRNA, while those infected with MP-12, rMP12-PTNSs or rMP12-SFSNSs did not induce IFN-β and ISG56 mRNA synthesis (Fig. 5). These results suggest that both PTV and SFSV NSs inhibit the accumulation of IFN-β mRNA, consistent with previous studies [34], [48]. Taken together, the results suggest that rMP12-PTNSs inhibits both host general transcription and IFN-β mRNA synthesis and that rMP12-SFSNSs inhibits IFN-β mRNA synthesis but not host general transcription.
To further study these observations, we used dual luciferase reporter assays to analyze the activation of IFN-β promoter and two critical transcription factors for IFN-β gene upregulation; IFN regulatory factor-3 (IRF-3) and Nuclear Factor-Kappa B (NF-κB), in the presence of MP-12, PTV or SFSV NSs [63], [64]. 293 cells were transfected with (1) IFNb-pGL3 plasmids encoding firefly luciferase (FFluc) under the human IFN-β promoter [54], (2) 4×IRF3-luc plasmid encoding FFluc under 4 copies of the IFN-β promoter positive regulatory domain (PRD)I/III motif (IRF-3 binding element) [55], or (3) pPRDII-luc plasmids encoding FFluc under the IFN-β promoter PRDII motif (NF-κB-binding element) [56]. As a transfection control, pRL-SV40 plasmid encoding Renilla luciferase (rLuc) under the constitutively-active SV40 promoter was co-transfected with above plasmids. At 24 hours post transfection, cells were mock-infected or infected with 100 U/ml of Sendai virus (SeV), and immediately mock-transfected or transfected with in vitro synthesized RNA encoding either chloramphenicol acetyl transferase (CAT) (control) or NSs of MP-12, PTV Adames strain or SFSV. Cells were collected at 16 hpi, and the relative luciferase activity was measured. When we defined the FFluc activities obtained from SeV-infected cells transfected with IFNb-pGL3, 4×IRF3-luc or pPRDII-luc as 100%, FFluc activities of mock-infected cells transfected with IFNb-pGL3, 4×IRF3-luc or pPRDII-luc showed 5.5%, 2.8% or 22% of the SeV-infected cells, respectively (Fig. 6 A-C, left panels, gray bars). On the other hand, rLuc activities of mock-infected cells transfected with pRL-SV40 (control plasmid) were similar or increased compared to those of SeV-infected cells transfected with pRL-SV40 (Fig. 6 A-C, right panels, gray bars). The results suggest that SeV infection specifically induces the activation of IFN-β gene, IRF-3 and NF-κB. Compared to CAT RNA control (Fig. 6 A-C, left panels, blue bars), cells transfected with MP-12 NSs RNA reduced the FFluc expression level derived from the IFN-β promoter (Fig. 6 A, left panels, red bars), IRF-3 (Fig. 6 B, left panels, red bars) and NF-κB (Fig. 6 C, left panels, red bars). Our results were consistent with the findings that RVFV NSs inhibits the general transcription factor TFIIH and induces general transcription suppression regardless of IRF-3 or NF-κB activation [28], [29]. Similarly, cells transfected with NSs RNA of PTV Adames reduced the FFluc expression level derived from the IFN-β promoter (Fig. 6 A, left panels, yellow bars), IRF-3 (Fig. 6 B, left panels, yellow bars) and NF-κB (Fig. 6 C, left panels, yellow bars). Interestingly, cells transfected with SFSV NSs RNA consistently increased the rLuc activity derived from the constitutively-active SV40 promoter (Fig. 6 A–C, right panels, pink bars). SFSV NSs reduced FFluc expression level derived from the IFN-β promoter (Fig. 6 A, left panels, pink bars) and IRF-3 (Fig. 6 B, left panels, pink bars), but not that from NF-κB (Fig. 6 C, left panels, pink bars). These results suggest that MP-12 NSs and PTV Adames NSs inhibit the reporter activities derived from the IFN-β promoter, IRF-3 and NF-κB, while SFSV NSs inhibits reporter activities derived from the IFN-β promoter and IRF-3 but not those from NF-κB. In addition, SFSV NSs seems to increase the expression of constitutively active genes through a currently unknown mechanism.
Based on the experiments described above, we confirmed that rMP12-PTNSs induces host general transcription suppression, inhibits the up-regulation of IFN-β gene, and does not promote PKR degradation, while rMP12-SFSNSs inhibits the up-regulation of IFN-β gene, but does not induce host general transcription suppression or PKR degradation. Next, we tested the efficacy of MP-12, rMP12-PTNSs or rMP12-SFSNSs against wt RVFV challenge in outbred mice. We also tested the previously reported rMP12-NSsR173A, which encodes mutant MP-12 NSs R173A [52]. Like rMP12-PTNSs, rMP12-NSsR173A inhibits host general transcription, suppresses activation of IFN-β gene but does not promote the degradation of PKR [52]. Five-week-old outbred CD1 mice were mock-vaccinated with PBS (n = 10), or subcutaneously vaccinated with 1×105 pfu of MP-12 (n = 20), rMP12-NSR173A (n = 10), rMP12-PTNSs (n = 9) or rMP12-SFSNSs (n = 10). We used outbred mice to evaluate the immunogenicity and efficacy of mice with different genetic background. Mice were monitored daily, and intraperitoneally challenged with 1×103 pfu of RVFV ZH501 strain at 45 days post vaccination. We performed mouse IFN-α ELISA using mouse sera at 1 day post vaccination as described previously [45]. At 1 day post vaccination, mice vaccinated with MP-12 or rMP12-PTNSs did not increase the level of IFN-α in sera, while serum IFN-α was detected in some mice vaccinated with rMP12-NSsR173A (33%) or rMP12-SFSNSs (10%) (data not shown). Low levels of viremia (100 pfu/ml) were detected in some mice vaccinated with MP-12 (20% at 3 days post vaccination), rMP12-R173A (10% at 2 days post vaccination) or rMP12-PTNSs (10% at 3 days post vaccination) (data not shown). One mouse vaccinated with MP-12 that suffered viremia was dead at 13 days post vaccination, and 1 mouse vaccinated with rMP12-SFSNSs became moribund, and was euthanized at 14 days post vaccination. We analyzed the euthanized mouse vaccinated with rMP12-SFSNSs histopathologically, and found no lesions in liver and spleen, while the encephalitis characterized by mild perivascular cuffing with mononuclear cells and neuronal necrosis with infiltration of microglia were observed. Viral N antigens were diffusely detected in the parenchymal area, and neurons in hippocampus, cortex and medulla contained viral antigens (Fig. S1). None of mice vaccinated with rMP12-NSsR173A or rMP12-PTNSs died before wt RVFV challenge. The result suggests that rMP12-SFSNSs retains neuroinvasiveness and neurovirulence similar to those of parental MP-12.
In the challenge study, 90% of mock-vaccinated mice died within 8 days post infection, while 63%, 50%, 78% or 89% of mice vaccinated with MP-12, rMP12-NSR173A, rMP12-PTNSs or rMP12-SFSNSs were protected from wt RVFV challenge, respectively (Fig. 7A). The survival curve was statistically analyzed with Log rank test, and the difference in the survival curves among mice immunized with MP-12, rMP12-NSsR173A, rMP12-PTNSs or rMP12-SFSNSs were not statistically significant. A mock-vaccinated mouse that survived after wt RVFV challenge did not show reduction of body weight by more than 5% (Fig. S2B), and developed 1∶2,560 of neutralizing antibody at 21 days post wt RVFV challenge (data not shown), while 7 to 20% drop in body weight was observed before euthanasia in 80% of mock-vaccinated moribund mice after wt RVFV challenge (Fig. S2A).
Among the mice vaccinated with MP-12, 9 (47%) mice developed both neutralizing antibodies; plaque reduction neutralizing test (PRNT80) titer 1∶640 to 1∶2,560, and anti-N IgG titer 1∶800 to 1∶204,800 at 42 days post vaccination, and those mice all survived wt RVFV challenge (Fig. 7B and C, Table 1). The 9 surviving mice showed body weight at 90 to 110% range during the observation period (Fig. S2C). On the other hand, 10 (53%) mice did not develop neutralizing antibodies; 3 survived (Fig. S2D) and 7 died (Fig. S2E and F) after wt RVFV challenge. Among the 3 survivors, 2 developed anti-N IgG (1∶204,800) and the other (#14-3) did not raise anti-N IgG. The mouse #14-3 showed temporal 16% body weight reduction at 7 days post challenge (Fig. S2D), and also developed neutralizing antibodies (1∶2,560) at 21 days post challenge. On the other hand, rMP12-NSsR173A showed poor efficacy (50% survival) (Table 1). Three surviving mice (30%) vaccinated with rMP12-NSsR173A developed both neutralizing antibody (1∶10 to 1∶160) and anti-N IgG (1∶400 to 1∶6,400) (Fig. S2G), while 2 surviving mice (20%) had developed anti-N IgG (1∶800 to 1∶6,400) without neutralizing antibody (Fig. S2H). On the other hand, the remaining 5 mice (50%) died without the presence of neutralizing antibodies (3 mice developed 1∶400 to 1∶3,200 of anti-N IgG) (Fig. S2I and J). Together with previous observation that rMP12-NSsR173A does not efficiently accumulate viral proteins due to PKR-mediated eIF2α phosphorylation [52], this result suggests that MP-12 encoding a mutant NSs, which inhibits host general transcription including IFN-β gene, but does not promote PKR degradation, is not immunogenic, and poorly induces neutralizing antibodies.
Compared to vaccination with MP-12 or rMP12-NSsR173A, all mice vaccinated with rMP12-PTNSs or rMP12-SFSNSs developed anti-N IgG (Fig. 7C, Table 1). Furthermore, mice vaccinated with rMP12-PTNSs or rMP12-SFSNSs showed significantly higher titers of neutralizing antibodies and anti-N IgG than those vaccinated with rMP12-NSsR173A (Fig. 7B and C). None of the survived mice vaccinated with rMP12-PTNSs or rMP12-SFSNSs showed a decrease in body weight below 90% (Fig. S3). The results suggest that rMP12-PTNSs and rMP12-SFSNSs have slightly higher efficacy than MP-12 and induce neutralizing antibodies at equivalent level to those induced by MP-12 in spite of a lack of PKR degradation function.
DIVA is important for vaccination of ruminants. Inclusion of negative DIVA marker helps to detect animals exposed to wt RVFV during outbreak. Unfortunately, MP-12 vaccine does not have a DIVA marker, and further improvement of MP-12 is required for an efficient detection of naturally infected animals during RVF outbreak. We determined whether mice vaccinated with rMP12-PTNSs or rMP12-SFSNSs can induce IgG reactive to RVFV NSs or not. For the purpose, we used IgG ELISA using C-terminus RVFV NSs [45]. We used the C-terminus NSs because the antigen is soluble, and detect anti-NSs antibody at higher sensitivity than IgG ELISA using purified whole RVFV NSs (Fig. S4). As shown in Fig. 7D, none of the mice vaccinated with rMP12-PTNSs or rMP12-SFSNSs induced detectable levels of IgG cross-reactive to the C-terminus of RVFV NSs, while 26% of mice vaccinated with MP-12 had detectable anti-NSs IgG in our IgG ELISA. The presence of anti-RVFV NSs antibody in vaccinated animals will compromise the DIVA strategy to detect animals exposed to wt RVFV. On the other hand, the rMP12-PTNSs and rMP12-SFSNSs did not induce anti-RVFV NSs antibody detectable in this IgG ELISA, and are applicable to DIVA. We also noted that all mice vaccinated with MP-12, rMP12-PTNSs or rMP12-SFSNSs raised anti-NSs IgG after wt RVFV challenge, suggesting the vaccinations do not confer sterile immunity, and wt RVFV replicated in vaccinated mice.
Live-attenuated Smithburn vaccine, generated by serial passage of wt RVFV Entebbe strain in mouse brain, has been used in endemic areas as a veterinary vaccine since 1950s, and its genetic subpopulations have never been reported. Grobbelaar et al. suggested that the use of Smithburn strain in endemic countries could result in the spread of RVFV strain by multiple use of automatic syringes for both viremic and uninfected ruminants during outbreaks, and may have resulted in reassortment of Smithburn strain with circulating wt RVFV during an outbreak [11]. Thus, a live-attenuated RVF vaccine that contains virulent subpopulations or non-attenuated segments, should not be used in viremic animals.
MP-12 vaccine strain was generated by serial 12-time plaque passages in MRC-5 cells in the presence of 5-fluorouracil [35]. The safety and immunogenicity of MP-12 in ruminants were reported [38], [39], [40], [41], and MP-12 is excluded from HHS/USDA select agent rule in the U.S. The MP-12 vaccine generated from the master seed currently has the Investigational New Drug (IND) status in the U.S., and was manufactured by using certified MRC-5 cells for human clinical trials [36]. We recently characterized the genetic subpopulations of an MP-12 vaccine lot, and found that the major population of MP-12 is highly stable and no reversions to the parental ZH548 strain were detected [36]. Different from viral passages in MRC-5 cells, RVFV induces NSs gene truncation during passages in cells lacking an intact type-I IFN system [43], [44]. The genetic stability and consistency of immunogenicity profiles are important factors in vaccine development, and the use of MRC-5 cells may be an important factor to maintain the original populations of MP-12 during manufacturing. One study also suggested that MP-12 strain with unknown passage history caused abortion and teratogenic effect in lambs when it was used for pregnant ewes at 35 to 56 days of pregnancy [65]. In this study, we aimed to develop MP-12 encoding functional NSs gene derived from serologically distinct phleboviruses to encode a DIVA marker in MP-12 while retaining the original efficacy and ability to efficiently replicate in MRC-5 cells. To encode a DIVA marker without affecting the ability to replicate in type-I IFN-competent MRC-5 cells, we designed MP-12 encoding serologically distinct phlebovirus NSs (PTV NSs or SFSV NSs), which is known to inhibit host IFN-β [34], [48]. The rMP12-PTNSs and rMP12-SFSNSs efficiently inhibited host IFN-β gene up-regulation induced by the MP-12 replication. Interestingly, rMP12-PTNSs but not rMP12-SFSNSs replicated efficiently in MRC-5 cells, indicating that rMP12-PTNSs can be considered as an alternative candidate vaccine of MP-12 with DIVA marker, which can be produced by using MRC-5 cells. Any MP-12 variants including rMP12-SFSNSs, which does not replicate in MRC-5 cells, will require genetic stability test in type-I IFN-incompetent cells to optimize the vaccine production.
PTV Adames strain [66] inhibits the human IFN-β gene [48]. SFSV NSs also inhibits the IFN-β gene expression if it is expressed from wt RVFV in place of RVFV NSs, while it does not promote the degradation of PKR [34], [47]. Consistent with wt RVFV encoding SFSV NSs, rMP12-SFSNSs inhibited the up-regulation of IFN-β gene and did not promote PKR degradation. We found that PTV NSs also does not promote PKR degradation (Fig. 1). Furthermore, PTV NSs inhibited host general transcription and the IFN-β promoter, while SFSV NSs inhibited the IFN-β promoter but not host general transcription (Fig. 4–6). Interestingly, rMP12-PTNSs formed clear plaques indistinguishable to those of parental MP-12, while rMP12-SFSNSs formed turbid plaques similar to those of rMP12-C13type (Fig. 1B) [51], [58].
RVFV NSs promotes the degradation of PKR [32], [34], and we previously found that the expression of dominant-negative PKR in place of MP-12 NSs increases the accumulation of dendritic cells infected with the MP-12 mutant [45]. On the other hand, the immunogenicity and efficacy of MP-12 encoding NSs mutant, which inhibits host general transcription but not PKR, have not been studied. The rMP12-NSsR173A inhibits host general transcription including IFN-β gene without promoting the degradation of PKR [52]. In this study, we found that rMP12-PTNSs inhibits host general transcription including IFN-β gene, and does not promote the degradation of PKR. Mice vaccinated with rMP12-PTNSs but not rMP12-NSsR173A induced high level of neutralizing antibodies. The result suggested that the host transcription suppression induced by RVFV NSs negatively affects the vaccine efficacy if PKR is not inhibited. On the other hand, MP-12 encoding PTV NSs was highly efficacious even though it induces host general transcription suppression without inducing PKR degradation. It might be possible that PTV NSs has other unknown functions to support RVFV replication in the presence of host general transcription suppression. We noted that SFSV NSs possesses an unknown function to increase host gene expression, as indicated by up-regulation of a constitutively expressed SV40 reporter gene (Fig. 6). The gene up-regulation was induced independently of PKR degradation, and may contribute to consistently high level of anti-N IgG in rMP12-SFSNSs vaccinated mice. Further studies are currently being conducted in our laboratory to elucidate the detailed mechanism of PTV NSs and SFSV NSs in host gene expression. Both SFSV and PTV cause self-limiting febrile illness in humans, and no significant diseases in animals. We observed that one mouse vaccinated with rMP12-SFSNSs was dead at 9 days post vaccination with viral encephalitis (Fig. S1). Vaccine-related viral encephalitis was also observed in mice vaccinated with parental MP-12 (data not shown), and we did not observe any significant increase of mouse death related to rMP12-SFSNSs vaccination compared to MP-12 vaccination. Considering that MP-12 vaccine encodes fully functional NSs of RVFV, rMP12-PTNSs and rMP12-SFSNSs are similar with, or most probably more attenuated than parental MP-12 due to a lack of function to promote PKR degradation. Since mouse is the most susceptible species to RVFV infection, the vaccine safety should be test in ruminants and nonhuman primates before further consideration as a vaccine candidate.
In summary, MP-12 mutants encoding NSs of PTV or SFSV are highly efficacious in mice and encode a DIVA marker, while rMP12-PTNSs also replicates efficiently in MRC-5 cells, which is useful for the vaccine manufacturing process. The immunogenicity and safety profile of rMP12-PTVNSs in ruminants and nonhuman primates will need to be tested to develop this virus as an alternative of MP-12 vaccine for veterinary and human use, respectively.
|
10.1371/journal.ppat.1001305 | Critical Role of the Virus-Encoded MicroRNA-155 Ortholog in the Induction of Marek's Disease Lymphomas | Notwithstanding the well-characterised roles of a number of oncogenes in neoplastic transformation, microRNAs (miRNAs) are increasingly implicated in several human cancers. Discovery of miRNAs in several oncogenic herpesviruses such as KSHV has further highlighted the potential of virus-encoded miRNAs to contribute to their oncogenic capabilities. Nevertheless, despite the identification of several possible cancer-related genes as their targets, the direct in vivo role of virus-encoded miRNAs in neoplastic diseases such as those induced by KSHV is difficult to demonstrate in the absence of suitable models. However, excellent natural disease models of rapid-onset Marek's disease (MD) lymphomas in chickens allow examination of the oncogenic potential of virus-encoded miRNAs. Using viruses modified by reverse genetics of the infectious BAC clone of the oncogenic RB-1B strain of MDV, we show that the deletion of the six-miRNA cluster 1 from the viral genome abolished the oncogenicity of the virus. This loss of oncogenicity appeared to be primarily due to the single miRNA within the cluster, miR-M4, the ortholog of cellular miR-155, since its deletion or a 2-nucleotide mutation within its seed region was sufficient to inhibit the induction of lymphomas. The definitive role of this miR-155 ortholog in oncogenicity was further confirmed by the rescue of oncogenic phenotype by revertant viruses that expressed either the miR-M4 or the cellular homolog gga-miR-155. This is the first demonstration of the direct in vivo role of a virus-encoded miRNA in inducing tumors in a natural infection model. Furthermore, the use of viruses deleted in miRNAs as effective vaccines against virulent MDV challenge, enables the prospects of generating genetically defined attenuated vaccines.
| MicroRNAs (miRNAs), encoded in the genomes of a number of organisms including several viruses, belong to a class of small RNA molecules that can function as key regulators of gene expression influencing various biological processes and diseases including cancer. Among all the miRNAs, miR-155 has been well documented for its direct role of oncogenesis in a number of species including chickens. Remarkably, miR-K12-11 and miR-M4, the miRNAs encoded by the oncogenic Kaposi's sarcoma-associated herpesvirus (KSHV) and Marek's disease virus (MDV) respectively, have been shown to be functional orthologs of miR-155. There are no animal models of KSHV-induced tumors to examine the oncogenic potential of miR-K12-11. However, using recombinant mutant viruses in excellent models of MDV-induced lymphomas in their natural chicken hosts, we demonstrate that miR-M4 is critical for the induction of tumors. This is the first study that clearly demonstrates a direct role for a single miRNA in inducing cancer in an in vivo animal model. The ability of gga-miR-155 to rescue the oncogenic potential of miR-M4-deleted virus demonstrated the conservation of oncogenic functions the two miRNAs. Moreover, we show that virus attenuated by deleting the miRNAs can function as vaccines against virulent virus infection, enabling the prospects of generating novel molecularly-defined vaccines.
| Oncogenic viruses account for nearly a fifth of all human cancers [1]. In addition to their devastating effects on human health, virus-induced neoplastic diseases are valuable models in understanding the molecular pathways and dynamics of cancer. Although most of the past studies on the molecular mechanisms of cancer induced by oncogenic viruses have primarily centred on classic oncogenes [2], increasing demonstration of the widespread role of micro(mi)RNAs in cancer has added new dimensions to the molecular mechanisms of neoplastic transformation [3], [4]. Moreover, the identification of miRNAs encoded by human oncogenic viruses such as Kaposi's sarcoma-associated herpesvirus (KSHV) and Epstein-Barr virus (EBV) suggested that virus-encoded miRNAs may also contribute towards the oncogenicity of the virus. The potential of some of these miRNAs to target a number of host genes associated with the oncogenic pathways has further strengthened the case for their role in inducing malignancies [5], [6]. However, despite the identification of several potential targets [7], direct in vivo role of virus-encoded miRNAs in the induction of malignancies is yet to be demonstrated, due at least in part, to the absence of suitable animal disease models.
Marek's disease (MD), a naturally occurring neoplastic disease of poultry [8], is an excellent model for herpesvirus-induced lymphomas [9], [10], [11], [12]. All birds exposed to the causative Marek's disease virus (MDV), unless protected by vaccination or have genetic resistance to the disease, develop rapid-onset T-cell lymphomas usually from 3–4 weeks after infection. Among the virus-encoded genes, Meq protein encoded from the MDV EcoRI Q fragment is considered to be primarily associated with the oncogenicity of the virus, since its deletion [13] or inhibition of its interaction with host proteins such as c-Jun, c-Fos and C-terminal binding protein [11], [14], [15] can abolish the oncogenic potential of the virus. Although these studies clearly demonstrate the contribution of Meq in oncogenesis, other viral genes could also be important in the induction of MD lymphomas. We and others have shown that MDV also encodes 14 miRNAs located within three clusters from each of the long (IRL/TRL) and short (IRS/TRS) repeat regions of the viral genome [16], [17]. Most of these miRNAs are expressed at high levels in MD lymphomas and transformed T-cell lines [18] suggesting their direct role in transformation. The six miRNAs in cluster 1, thought to be processed from a single primary transcript upstream of the Meq locus [16] included miR-M4, the functional ortholog of miR-155 [19] and KSHV-miR-K12-11 [20], [21]. MDV miR-M4 has recently been shown to inhibit the translation of viral proteins involved in the cleavage and packaging of herpesvirus DNA [22]. However, based on the direct association of miR-155 in several cancers [23], [24], including the recent findings on its roles in TGF-β pathway and lymphomagenesis [25] and modulation of mismatch repair and genomic instability [26], we hypothesised that miR-M4 has a major role in MDV oncogenicity. This was also supported by recent observations of high levels of its expression in tumors correlated with the pathogenicity of viral strains [17].
In the present study, we explored the functional role of the miRNAs encoded within the cluster 1 of the MDV genome, including that of the miR-155 functional ortholog miR-M4, in inducing T-cell lymphomas using the well-established models of MD in the natural chicken hosts of the virus. Using a series of mutant viruses generated by reverse genetics techniques [27], [28] on the full-length infectious bacterial artificial chromosome (BAC) clone of the highly oncogenic RB-1B virus [29], [30], we were able to directly examine the in vivo functions of MDV-encoded miRNAs in the induction of MD lymphomas. Our studies demonstrate the critical role of the cluster 1 miRNAs, especially of the miR-M4, in the induction of lymphomas, thus providing the direct evidence on the in vivo function of miRNAs in virus-induced cancer.
Members of the Family Herpesviridae account for most of the currently identified virus-encoded miRNAs, and these are thought to play important roles in the distinct biological and pathogenic features of these viruses. MDV encoded-miRNAs are expressed as 3 clusters from the repeat regions, making each miRNA present as two identical copies in the viral genome (Fig. 1). We have previously demonstrated high levels of expression of these miRNAs both in vitro and in vivo [16], [31]. However, these studies did not show whether any of these miRNAs are essential for viral replication. In order to address this, we examined the role of cluster 1 miRNAs using a series of mutant viruses generated by reverse genetics techniques on the pRB-1B5 BAC clone [29]. In the mutant virus construct miR-00, the region encoding the miRNA cluster and the overlapping R-LORF8 in the TRL or IRL regions were deleted by recombination using galK-kan and thyA-spec antibiotic selection cassettes. While galK-kan cassette replaced the entire region encoding all the six miRNAs from one of the repeat regions, thyA-spec cassette that targeted a shorter region removed all the miRNAs except miR-M9 and miR-M4 from the other repeat region.
For generation of viruses with mutations in all of the miRNAs within the cluster, we synthesised a gene designed to introduce mutations into the stem loop structures of each of the six miRNAs in the cluster 1 without affecting the R-LORF8 frame (Fig. 2). In the miR-S0 and the miR-SS constructs respectively, one or both copies of the miRNA locus were replaced with this synthetic gene sequence, preventing them expressing any of cluster 1 miRNAs. The revertant miR-W0 and miR** viruses both expressed all the miRNAs, but in the latter, the introduction of a stop codon is expected to abolish the translation of R-LORF8. These two constructs were included to delineate between the functions of miRNAs and the overlapping R-LORF8 gene. The details of mutations in each of the individual recombinant viruses are summarised in Table 1. Reconstitution of the viruses in primary chicken embryo fibroblast (CEF) transfected with the BAC DNA demonstrated the infectivity of each of the above constructs. Demonstration of viral proteins Meq, pp38 and pp14 in CEF infected with the parent pRB-1B5 and different mutant viruses by western blotting showed that the mutations did not affect the expression of these proteins (Fig. 3A). This was further confirmed by specific staining of Meq and pp38 in the nucleus and cytoplasm respectively of CEF infected with a selection of the mutant viruses (Fig. 3B). Moreover, analysis of the in vitro growth kinetics showed very similar replication kinetics for both the parent and mutant viruses (Fig. 4A) indicating that these miRNAs are not essential for replication of MDV in vitro. For further analysis of the effect of mutations in the miRNA cluster on MDV gene expression, we compared the relative levels of ICP4 and Meq transcripts by quantitative RT-PCR in CEF infected with the wild type RB-1B, parent pRB-1B5 and the mutant viruses. MDV-transformed lymphoblastoid cell line 265L and CEF infected with the Meq-deletion mutant (RB-1BΔMeq) virus were used as controls. These studies showed that the transcription of ICP4 and Meq were not affected by mutations in the miRNA locus (Fig. 5A–B), although the levels of ICP4 transcripts were low with variations among different viruses in comparison to the Meq transcripts.
In order to examine whether these miRNAs are essential for MDV replication in vivo, we asked whether viruses with mutations in the miRNA locus could be re-isolated from birds in later stages of infection. Successful isolation of MDV by co-culturing of the peripheral blood leukocytes (PBL) of birds infected with different viruses 7–8 weeks after infection confirmed that the mutations in the miRNA locus did not affect the viral replication and their ability to establish latent infection in vivo. Moreover, quantitative RT-PCR data of Meq, LAT and ICP4 transcripts in CEF co-cultured with the PBL from infected birds (Fig. 5C) indicated that the modifications in the miRNA locus did not affect the expression of these transcripts in viruses isolated from these birds. All these results confirmed that the miRNAs encoded from this locus did not affect MDV replication either in vitro or in vivo.
Having shown that the mutations in the miRNA locus did not affect the replication and viral gene expression in vitro or the establishment of latency in vivo, we investigated whether the miRNAs encoded within this cluster are associated with the oncogenicity of the virus using the established infection models in the natural chicken hosts [29]. Compared to the 100 percent incidence of MD in birds infected with the parent pRB-1B5 virus (Fig. 6A), none of the birds infected with the miR-00, miR-S0 and miR-SS viruses developed any gross or microscopic lesions of MD, showing that the loss of expression of the cluster 1 miRNAs directly affected the oncogenic potential of MDV. Rescue of the oncogenic phenotype by the revertant miR-W0 virus to the levels close to that of the parent virus, even by restoring only one copy of the miRNA cluster, provided the strong evidence on the distinct role of these miRNAs in the induction of lymphomas. Moreover, the induction of MD in 90% of the birds by the miR** virus (Fig. 6A) demonstrated that miRNAs, but not R-LORF8, are involved in the induction of lymphomas.
Having demonstrated that the miRNAs encoded in the cluster 1 are essential for the oncogenic potential of the virus, we wanted to refine our investigation to identify whether any single miRNA within this cluster plays a more important role in the oncogenic potential of the virus. Since MDV miR-M4 in this cluster is a functional ortholog of the oncogenic miR-155 [19] implicated in a number of neoplastic disorders [23], we decided to focus on this particular miRNA. For this, we used the modified pRB1B5 clone [30] to generate two additional mutant viruses where only the miR-M4 was made non-functional. These included the constructs miR-4-00, where both copies of the pre-miR-M4 sequences were deleted, and the miR-4-0-mu2, where one of the deleted copies was replaced with a mutant miR-M4 carrying a 2-nucleotide mutation in the seed sequence (Fig. 7). We have previously shown that this mutation in the seed region abolished miR-M4 function [19]. We also determined the complete sequence of the ∼178-kb genome of the miR-4-0-mu2 construct by deep sequencing and ruled out unintended mutations or rearrangements generated during the BAC manipulation. Revertant construct miR-4-0-R was also generated by restoring one copy of the wild type miR-M4 sequence. A summary of the miRNA expression profiles of these constructs are in Table 1. The mutant viruses showed similar growth kinetics in CEF further demonstrating that miR-M4 deletion did not affect replication in vitro (Fig. 4B).
We also examined the expression levels of MDV-encoded miRNAs miR-M4, miR-M5 and miR-M6 by qRT-PCR on RNA extracted from CEF infected with wild type RB-1B, the parent pRB-1B5 and the different mutant viruses (Fig. 8A–C). MDV-transformed cell line 265L and RB-1BΔMeq virus were used as controls. The absence of miR-M4 and miR-M5 transcripts in the cells infected by miR-S0 and miR-SS viruses (the low level of miR-M4 but not miR-M5 in miR-00 virus is due to the presence of one copy of miR-M4), confirmed the deletion of these miRNAs in these constructs (Fig. 8A,B). As expected, the expression of miR-M6 located within the LAT region was not affected by the mutations within the cluster 1 miRNA locus (Fig. 8C).
When the oncogenicity of these viruses was evaluated in MD lymphoma model in the natural chicken hosts, the parent pRB-1B5 virus induced 100 percent disease (Fig. 6B). Compared to this, none of the birds infected with miR-4-0mu2 virus showed any clinical evidence, gross tumors or microscopic lesions in any of the organs. There was 90% reduction of MD incidence in birds infected with miR-4-00 virus (Fig. 6B), clearly demonstrating the important role of miR-M4 in inducing lymphomas. The rescue of the oncogenic phenotype in 70% of birds by the revertant miR-4-0R virus even by restoring one copy of the wild type of miR-M4, further confirmed the significant role of miR-M4 in the induction of lymphomas.
Since miR-M4 is a functional ortholog of the host-encoded gga-miR-155 [19], we wanted to examine whether gga-miR-155 can replace the oncogenic function of miR-M4 in MDV. For this, we generated an additional MDV construct miR-4-0-155 (Table 1), in which gga-miR-155 pre-miRNA including the loop sequence of the chicken BIC transcript [32] was introduced in the position of the viral miR-M4 pre-miRNA. After demonstrating that the in vitro growth kinetics on CEF was similar (Fig. 4B), we examined the oncogenic potential of miR-4-0-155 virus in the MD lymphoma models in chickens. Demonstration of the incidence of MD at similar levels (70%) in birds infected with miR-4-0-155 virus and the miR-4-0R revertant virus (Fig. 6B) indicated that gga-miR-155 can function as an oncogenic miRNA in the context of the MDV cluster 1 miRNAs. Although the levels of MD incidence were similar, the onset of tumors induced by miR-4-0-155 virus was slow, with most of the tumors detected much later than those induced by the parent pRB-1B5 or the miR-4-0R revertant viruses (Fig. 6B). Tumours induced by the mutant viruses, including the chimeric miR-4-0-155 virus, were typical MD lymphomas with neoplastic lymphoid lesions in multiple organs (Fig. 9A). In order to demonstrate the expression of gga-miR-155 or miR-M4 in the tumors induced by miR-4-0-155 and miR-4-0-R viruses respectively, we carried out quantitative RT-PCR on RNA samples extracted from the lymphomas collected from infected birds at post mortem. Demonstration of expression of miR-M4, but not gga-miR-155, in lymphoid tumors harvested from birds infected with miR-4-0-R virus (bird numbers #2166S, 2171S, 2172S and 2178K) indicated the direct correlation between the expression of miR-M4 and the induction of lymphomas (Fig. 9B–C). Despite the low levels of expression, detection of gga-miR-155 but not miR-M4 in tumors of birds infected with miR-4-0-155 virus (bird numbers #2153L, 2153S, 2163L and 2183S) by quantitative RT-PCR and Northern blotting (Fig. 9D) suggested a direct role for gga-miR-155 in induction of these tumors.
Demonstration of the total loss of oncogenicity persuaded us to investigate whether MDV made non-oncogenic by inactivation of miRNAs can function as effective vaccines against virulent MDV infection. For this, groups of birds were vaccinated with a recombinant (pCVI988-10) clone [33] of the widely used CVI988 vaccine strain or miR-SS virus that has mutations in all the miRNAs encoded from both copies of the cluster 1. Birds from all the groups were challenged with the very virulent plus MDV (vv+MDV) strain 675A along with unvaccinated controls. Comparison of the incidence of MD in the 3 groups during the 68-day experimental period showed that miR-SS virus provided the same level of protection as the widely used Rispens vaccine against infection by the vv+MDV strain 675A that induced 100 per cent disease in unvaccinated chickens (Fig. 10).
Despite the increasing evidence of the potential role of herpesvirus-encoded miRNAs functioning as determinants of oncogenicity [34], direct role of any of these miRNAs in the induction of tumors in vivo has not yet been demonstrated. Using lymphoma models in the natural chicken hosts infected with modified recombinant viruses generated from infectious BAC clones of the oncogenic RB-1B strain of MDV, our study provides the first direct evidence of the role of miRNAs in the induction of tumors. The total abolition of tumors in birds infected by viruses that do not express miRNAs in cluster 1, and subsequent rescue of the oncogenic phenotype by the revertant miR-W0 virus demonstrated the direct role of these miRNAs in the induction of MD lymphomas. Moreover, the inability of miR-S0 and miR-SS viruses to induce lymphomas and restoration of the oncogenic phenotype by miR** virus revealed that the miRNAs, and not the overlapping R-LORF8 gene, are the major determinants of oncogenicity. Important role of these miRNAs in oncogenesis is supported by the high levels of their expression in MD tumors and transformed cell lines [16], [18].
Although other miRNAs encoded within this cluster might be contributing to the induction of MD lymphomas [35], we focussed mainly on examining the role of miR-M4 for various reasons. Firstly, MDV miR-M4 is expressed at very high levels in lymphomas, in some cases accounting for even up to 72% of all MDV-encoded miRNAs [16], [17], [18], supporting its potential as a major determinant of MDV oncogenicity. Secondly, based on the seed sequence homology and the potential for regulating common targets such as Pu.1, BACH1, CEBPβ, HIVEP2, BCL2L13 and PDCD6, we have previously shown that miR-M4 is a functional ortholog of gga-miR-155 [19], a miRNA known to be directly associated with several cancers [23], [36] and molecular mechanisms of cancer pathogenesis [25], [26]. Finally, overexpression of miR-155 has been shown to be associated with lymphocyte transformation by viruses such as EBV [37] and reticuloendotheliosis virus strain T [38]. Significant reduction in the incidence of MD in birds infected with miR-4-00 virus, and the total loss of oncogenicity of miR-4-0-mu2 virus clearly indicated that miR-M4 plays a critical role in the induction of MD lymphomas. Although we have not carried out similar investigations on other miRNAs in this cluster, the dramatic suppression of oncogenicity subsequent to the loss of miR-M4 function to the levels similar to those shown by viruses with mutations in the whole cluster 1 miRNAs, suggested that miR-M4 is very important for the oncogenicity of MDV. However the expression of miR-M4 alone appeared to be not sufficient since miR-00 virus expressing low levels of miR-M4 (Fig. 8A) remained non-oncogenic, suggesting the contribution of other miRNAs in the cluster to the oncogenicity. This is also supported by our recent observation that viruses deleted in miR-M5, another miRNA within this cluster, retained oncogenicity although not to the same extent as the wild type viruses (unpublished data). The exact mechanisms of the loss of oncogenic phenotype by the miR-M4-deleted/mutated viruses remain to be investigated. However, it is not due to the lack of expression of viral genes such as Meq, ICP4 and LAT as we were able to demonstrate the expression of the proteins or transcripts in infected CEF or PBL (Fig. 3 and 5). It is also not due to the inability to express miRNAs such as miR-6 outside the cluster 1 (Fig. 8C). MDV miR-M4 is known to influence directly the expression of a number of viral and cellular targets [19], [22]. The regulation of gene expression by these miRNAs can also be due to indirect effects, and may include genes such as Meq. Although the expression levels of Meq in CEF infected with miR-M4-deleted/mutated viruses were not affected (Fig. 5B), the relative levels of Meq in the PBL of infected birds were lower than those from birds infected with viruses expressing the miRNAs (Fig. 5C). Meq is well-known for its key role in MDV oncogenesis [11]–[15]. Although Meq by itself is weak in its oncogenic potential, modulation of the levels of Meq expression by these miRNAs could be important in MDV oncogenicity. Our recent studies have shown that Meq may have a regulatory role in the expression of cluster 1 miRNAs by binding to its promoter region (unpublished) suggesting that Meq-miRNA regulatory loop could be important in the induction of tumors. Thus viral oncogenesis should be seen as a very complex process involving the interaction of multiple factors and regulatory processes including those by the virus-encoded miRNAs. Disruption of any one of these key elements could have a significant effect on the oncogenic pathways. Although the precise understanding of all the molecular processes would need further studies, the present findings demonstrate that miR-M4, most likely acting through the miR-155 pathway, functions as a key factor contributing to MD oncogenesis.
It is remarkable to show that a two-nucleotide mutation in the miR-M4 seed sequence in the context of the whole viral genome was sufficient for the total inhibition of oncogenicity of the miR-4-0-mu2 virus (Fig. 6B). One of the birds (#2168L) infected with miR-4-00 virus did develop MD lymphoma even in the absence of miR-M4 (Fig. 9B, C). This is unlikely to be due to any spurious mutations or recombination events in the virus, as we were able to confirm the absence of miR-M4 or miR-155 expression in the tumor tissues of this bird (Fig. 9D). We are yet to identify the exact mechanisms underlying the development of lymphomas in this bird in the absence of miR-M4. However, occurrence of such tumors underlines the complex and multifactorial nature of oncogenic processes; moreover the individual differences between birds in immunocompetence and genetic susceptibility could also contribute to the onset of such tumors. Nevertheless, the near total elimination of oncogenicity in the absence of miR-M4, and the rescue of oncogenicity in up to 70 per cent of the birds when the miR-M4 expression was restored in the revertant miR-4-0-R virus, provided the first clear evidence on the in vivo role of a virus-encoded miRNA in the induction of tumors.
The conservation of the seed sequence and the ability to regulate common sets of targets by miR-M4 and gga-miR-155 [19] prompted us to examine whether the miR-M4 functions of MDV can be replaced by gga-miR-155. The demonstration of the ability of the chimeric miR-4-0-155 virus to restore the oncogenic potential to the same levels as the miR-4-0-R (Fig. 6B), further demonstrated the significance of the miR-M4/miR-155 pathway in MD lymphomagenesis. Although the types of tumors induced by recombinant viruses expressing miR-M4 or gga-miR-155 were not distinguishable (Fig. 9A), the onset of tumors in birds infected miR-4-0-155 virus was slow (Fig. 6B), possibly due to the differences in the functional context of the two miRNAs or in their processing and expression. This was also evident from the expression levels of miR-M4 and miR-155 in the primary tumor samples collected from birds infected with the two viruses. Compared to the high levels of miR-M4 in tumor samples induced by miR-4-0-R virus (Fig. 9B), the levels of miR-155 in the tumors induced by miR-4-0-155 virus were much lower (Fig. 9B). The lack of expression of miR-M4 and weak expression of miR-155 in the tumors induced by miR-4-0-155 virus was also confirmed by Northern blot (Fig. 9D). Interestingly, there is consistent downregulation of gga-miR-155 in MD tumors and MDV-transformed cell lines [17], [18]. The downregulation of miR-155 in MDV-transformed tumor cell lines is not a permanent defect, as these cells can be induced to express miR-155 by co-infection with retroviruses-expressing v-Rel (unpublished data). Moreover, recent studies have demonstrated differences between miR-M4 and gga-miR-155 in targeting genes such as UL28 despite having identical seed sequences highlighting the significance of the non-seed regions of these miRNAs [22]. Tumors induced by the chimeric miR-4-0-155 virus are the first examples demonstrating the expression of gga-miR-155 in MD tumors. Although our study does not identify the various genes linked to the miR-M4-mediated oncogenesis, demonstration of the significance of miR-M4 in the induction of lymphomas is a major step in understanding the molecular oncogenic mechanisms in MD. From the demonstration of the conserved functions of miR-M4 and miR-155 in the induction of tumors, it is clear that MDV is able to restore the functions of miR-155 through the expression of high levels of the functional homolog miR-M4. Although it is unclear how MDV is able to downregulate miR-155, some feedback regulatory mechanisms are the most likely mechanisms. Interestingly in KSHV-induced tumors also, there is upregulation of the miR-155 functional homolog miR-K12-11, at the expense of significant downregulation of miR-155 [39]. On the other hand, EBV does not encode any functional ortholog but induces miR-155 in transformed B-cells [37], and a recent study has demonstrated its role in the induction of B-cell transformation by EBV in vitro [40]. It is unclear what advantages MDV and KSHV do have in choosing this rather complicated pathway of encoding and expressing high levels of the functional orthologs of a host miRNA, the expression of which is repressed in the transformed cells. One possible advantage of encoding a viral ortholog is the potential to achieve high levels of expression as seen in MDV- and KSHV-induced tumors, overriding the tighter cellular regulatory controls associated with the c-bic/miR-155 expression. Moreover, although miR-155 and the two viral orthologs may regulate common set of target proteins through the conserved seed sequences, other potential differences in their functions, especially due to the differences in the non-seed sequences, may also exist. It is not known whether the changes in the non-seed sequences do affect the functions of these miRNAs. Nevertheless, the differences in the speed in the onset of tumors between MDV expressing miR-M4 and gga-miR-155 would suggest that such differences could be important.
The role of miR-M4 and the rescue of the oncogenic phenotype by miR-155 provide further evidence of the conserved biological functions of miRNA orthologs, a finding of major importance in elucidating the functions of other viral orthologs such as KSHV miR-K12-11. The study also highlights the use of tumor virus disease models as powerful tools to reveal fundamental molecular determinants that trigger the development of cancers. Finally, the demonstration of protection induced by the miR-SS virus against infection by the vv+MDV strain 675A to the same levels as the widely used CVI988 vaccine strain (Fig. 10) demonstrated the prospects of generating molecularly-defined attenuated vaccines by specific deletion of oncogenic sequences such as the miRNAs.
All animal experiments were performed in accordance with the United Kingdom Home Office guidelines under the provisions of the Project License approved by the Institute for Animal Health Ethical Committee.
Primary chicken embryo fibroblast cultures (CEF) were prepared from 10-day old chicken embryos from SPF eggs as previously described [29]. Reconstitution of recombinant viruses was achieved by transfection of 1–2µg BAC DNA into the CEF using Lipofectamine (Invitrogen).
Infectious BAC clone pRB-1B5 [29], [30] was used for the generation of the mutant constructs [41]. List of primers used for the construction are shown separately (Table S1). Selection markers galK-kanamycin (galK-Kn) [42] and thymidylate synthase-spectinomycin (ThyA-spec) [43] cassettes were used in consecutive steps for the deletion of the two copies of the miRNA cluster. For this, the fragment containing the miRNA cluster (GenBank EF523390 - Nucleotides 134362 to 136848) was amplified with 5′ GCCAACTGTACACGCAGGGACGT 3′ and 5′ GTGCAGTGCCTTTGATGTCTG 3′ primers and cloned into pCR8-TOPO vector (Invitrogen). From this vector, the 1665-bp PshAI-StuI fragment (134527–136183) encompassing all the six miRNAs in the cluster 1 from miR-M9 to miR-M4 (Fig. 1) was replaced with galK-Kn cassette to generate the -recombination –construct for replacing the first copy of the microRNA cluster 1. Similarly, an XhoI fragment (135221–135959) from this vector was replaced with a ThyA-spec cassette to make another construct for the specific replacement of the second copy of the miRNA cluster. This shorter (783-bp) deletion removed all the miRNAs except miR-M9 and miR-M4. The two copies of the miRNA cluster were sequentially deleted using recombineering techniques [41], [44] to generate the miR-00 construct. For generation of viruses with mutations in all of the miRNAs within the cluster, a 1.45 kb NgoMIV-EcoRV fragment corresponding to the positions 134785–136216 of the pRB-1B5 sequence was synthesised. The synthetic gene was designed through alternative codon usage so as to destroy all the pre-miRNA's hairpin structures, but retaining the R-LORF8 open reading frame (Fig. 2). The first copy of the miRNA cluster 1 was replaced using a RecA-based strategy [28]. Recombinant clones detected by PCR (primers 5′ GTAGTGTATCGGTCTTCGTG 3′ and 5′ CCCGAATACAAGGAATCCTG 3′) were digested with BglII to identify the clones with the replaced synthetic sequence that contained a unique BglII site. The virus reconstituted from this construct with one copy of the miRNA cluster replaced with the synthetic sequence and the other copy deleted by the insertion of the galK-kn was designated miR-S0. The galK-kn selection marker in the miR-S0 construct was then replaced with the pKOV-miR-syn construct to generate the construct with both copies of the miRNA cluster 1 replaced by synthetic sequences, and named miR-SS. We also generated a revertant construct miR-W0, in which we replaced one copy of wild type miRNA cluster sequence from pRB-1B5, while the other copy remains deleted. We mutated one copy of R-LORF8 start codon by inserting FRT-Kn cassette amplified by PCR using R-LORF8-Kn-For and R-LORF8-Kn-Rev primers (Table S1). The FRT-Kn cassette was ‘Flipped-off’ in E coli strain SW105, leaving only a ‘scar’ sequence. In order to prevent the translation of R-LORF8, we introduced stop codon replacing the ATG initiation codon in the construct miR**. For this, we first destroyed the gene by introducing the galK-kn cassette into the locus. This was then replaced with a modified R-LORF8 sequence generated by PCR using R-LORF8-stop-For and R-LORF8-stop-Rev primers (Table S1). Detailed protocols for the manipulation of the BAC constructs are provided separately [41]. The accuracy of the mutations in different constructs was checked by sequencing.
In order to examine the role of miR-M4 in MDV oncogenicity, we also constructed a series of viruses with mutations only in the miR-M4, by standard mutagenesis techniques [45] on the pRB-1B5 clone [29]. First we deleted one copy of the pre-miR-M4 with FRT-Kn cassette using PCR product derived with miR-M4-kn-For and miR-M4-kn-Rev primers. After excision of the Kn cassette, the second copy was deleted with galK cassette using PCR products derived with miR-M4-galK-For and miR-M4-galK-Rev primers. This construct with deletion of both copies of the miR-M4 was designated miR-M4-00 and the galK locus in this construct was used for generating additional constructs. For the construction of a seed mutant of miR-M4, we generated a PCR product with miR-M4-mu2-Top and miR-M4-mu2-Bottom primers, designed to introduce two-nucleotide mutations in the miR-M4 seed region that has been previously shown to be sufficient to abolish the function of miR-M4. The galK selection marker in the miR-M4-00 construct was replaced with the mutated miR-M4 to generate the miR-4-0-mu2 construct. In order to generate another construct in which the miR-M4 was replaced with the miR-155 sequence, we generated the gga-miR-155 pre-miRNA along with the loop sequence using miR-M4-155-Top and miR-M4-155-Bottom primers. The galK selection marker in the miR-4-00 construct was replaced with the annealed above oligonucleotides to generate the miR-4-0-155. Finally, the revertant construct miR-4-0-R was generated in which the wild type miR-M4 sequence was restored. The accuracy of all the modifications was also confirmed by sequencing the modified regions of the constructs. The whole genome of the miR-4-0-mu2 was determined by deep sequencing to confirm the mutations and to rule out any unexpected mutations or recombinations.
All animal experiments were carried out under licence from the UK Home Office in dedicated negative pressure rooms. Groups (n = 10–12) of one day-old line P (MHC type B19/B19) chicks were infected with 1,000 plaque forming units (pfu) of miR-00, miR-S0, miR-SS, miR-W0, miR**, miR-4-00, miR-4-0-mu2, miR-4-0-155 and miR-4-0-R and pRB1B5 viruses as described. Blood samples collected at regular intervals or post-infection (d. p. i.) were used for quantitation of virus load or for the extraction of RNA from the PBL. Methods for quantitative RT-PCR to measure miRNA/transcript levels have been described [31]. Virus isolations were also performed by inoculating 1×106 PBL to a well of a 6-well plate of CEF and incubated at 38.5 C in 5% CO2 until the appearance of virus plaques. The birds were inspected regularly and were sacrificed at clinical end-points and samples taken post-mortem for histology. Birds showing gross or histological lesions of lymphomas in any of the tissues collected at post-mortem were diagnosed as MD-positive, and the data from the incidence of MD from each of the groups were used to calculate the cumulative survival rates.
Groups (n = 10) of one-day old birds were vaccinated via the intra-muscular route with either sham vaccine, or 925 pfu of miR-SS or pCVI988 viruses. One week after infection, all the birds were infected with 1450 pfu of the vv+MDV strain 675A via the intra-peritoneal route. Birds were observed for the onset of clinical disease and the incidence of MD was recorded up to 68 days post infection. All the birds were necropsied at the end of the experiment and the incidence of MD from the gross and microscopic lesions was used to calculate the survival rates from virus infection.
Western blotting to detect the viral proteins Meq, pp38 and pp14, and the chicken CtBP1 was carried out using methods previously described [11]. Immunofluorescence staining and confocal microscopy were carried out on CEF cultures on 13 mm glass coverslips infected with the mutant viruses using methods described [12]. The infected cell were fixed in 4% paraformaldehyde, permeabilised with 0.1% Triton x-100 and stained with anti-Meq antibody (FD7) or anti-pp38 (BD1) antibodies and detected with Alexa Fluor 488/568-conjugated goat anti-mouse reagents (Invitrogen). Images were taken using Leica TCS SP5 confocal laser scanning microscope.
|
10.1371/journal.ppat.1006185 | A neuropeptide modulates sensory perception in the entomopathogenic nematode Steinernema carpocapsae | Entomopathogenic nematodes (EPNs) employ a sophisticated chemosensory apparatus to detect potential hosts. Understanding the molecular basis of relevant host-finding behaviours could facilitate improved EPN biocontrol approaches, and could lend insight to similar behaviours in economically important mammalian parasites. FMRFamide-like peptides are enriched and conserved across the Phylum Nematoda, and have been linked with motor and sensory function, including dispersal and aggregating behaviours in the free living nematode Caenorhabditis elegans. The RNA interference (RNAi) pathway of Steinernema carpocapsae was characterised in silico, and employed to knockdown the expression of the FMRFamide-like peptide 21 (GLGPRPLRFamide) gene (flp-21) in S. carpocapsae infective juveniles; a first instance of RNAi in this genus, and a first in an infective juvenile of any EPN species. Our data show that 5 mg/ml dsRNA and 50 mM serotonin triggers statistically significant flp-21 knockdown (-84%***) over a 48 h timecourse, which inhibits host-finding (chemosensory), dispersal, hyperactive nictation and jumping behaviours. However, whilst 1 mg/ml dsRNA and 50 mM serotonin also triggers statistically significant flp-21 knockdown (-51%**) over a 48 h timecourse, it does not trigger the null sensory phenotypes; statistically significant target knockdown can still lead to false negative results, necessitating appropriate experimental design. SPME GC-MS volatile profiles of two EPN hosts, Galleria mellonella and Tenebrio molitor reveal an array of shared and unique compounds; these differences had no impact on null flp-21 RNAi phenotypes for the behaviours assayed. Localisation of flp-21 / FLP-21 to paired anterior neurons by whole mount in situ hybridisation and immunocytochemistry corroborates the RNAi data, further suggesting a role in sensory modulation. These data can underpin efforts to study these behaviours in other economically important parasites, and could facilitate molecular approaches to EPN strain improvement for biocontrol.
| Entomopathogenic nematodes (EPNs) use a range of behaviours in order to find a suitable host, some of which are shared with important mammalian parasites. The ethical burden of conducting research on parasites which require a mammalian host has driven a move towards appropriate ‘model’ parasites, like EPNs, which have short life cycles, can be cultured in insects or agar plates, and have excellent genomic resources. This study aimed to develop a method for triggering gene knockdown by RNA interference (RNAi), which would allow us to study the function of genes and the molecular basis of behaviour. We have successfully knocked down the expression of a neuropeptide gene, flp-21 in S. carpocapsae infective juveniles. We find that it is involved in the regulation of behaviours which rely on sensory perception and relate to host-finding. This study provides a method for employing RNAi in a promising model parasite, and characterises the molecular basis of host-finding behaviours which could be relevant to economically important mammalian parasites. EPNs are also used as bioinsecticides, and so understanding their behaviour and biology could have broad benefits across industry and academia.
| Entomopathogenic nematodes (EPNs) borrow their name from the entomopathogenic bacteria (Photorhabdus, Serratia and Xenorhabdus spp.) with which they form a commensal relationship. These nematodes provide a stable environment for the bacteria, and act as a vector between insect hosts. Once the nematode has invaded an insect, the nematode exsheaths (or ‘recovers’) and entomopathogenic bacteria are regurgitated into the insect haemolymph; the bacteria then rapidly kill and metabolise the insect, providing nutrition and developmental cues for the nematode. These entomopathogenic bacteria are then transmitted between nematode generations [1]. The entomopathogenic lifestyle has been found to arise independently in nematodes, at least three times, spanning significant phylogenetic diversity. Heterorhabditis and Oscheius spp. [2] reside within clade 9 along with major strongylid parasites of man and animal [3]; Steinernema spp. reside within clade 10 alongside strongyloidid parasites [4].
Nictation is a dispersal and host-finding strategy, enacted by nematodes which stand upright on their tails, waving their anterior in the air [5]. This behaviour is shared amongst many economically important animal parasitic and entomopathogenic nematodes, alongside the model nematode C. elegans, for which nictation is a phoretic dispersal behaviour of dauer larvae, used to increase the likelihood of attachment to passing animals. Nictation is regulated by amphidial IL2 neurons in C. elegans, which occur in lateral triplets either side of the pharyngeal metacorpus [5, 6]. IL2 neurons display significant remodelling from C. elegans L3 to dauer (the only life-stage to enact nictation behaviours) such that connectivity with other chemosensory and cephalic neurons is enhanced [6]. It has been shown that IL2 neurons express the DES-2 acetylcholine receptor subunit, and that cholinergic signalling is requisite for nictation [5, 7–9]. Additionally, the central pair of IL2 neurons express the FMRFamide-like peptide (FLP) receptor, NPR-1 [10]. To date there are two known NPR-1 agonists; FLP-18 and FLP-21 [11]. However, there is also known redundancy of FLP-18 and FLP-21 in signalling through other neuropeptide receptors (NPR-4, -5, -6–10, -11, and NPR-2, -3, -5, -6, 11, respectively) in heterologous systems [12, 13], making functional linkage difficult. Steinernema spp. also display a highly specialised jumping behaviour which is thought to enhance both dispersal and host attachment. Jumping occurs when a nictating infective juvenile (IJ) unilaterally contracts body wall muscles bringing the anterior region towards the posterior region, forming a loop. This generates high pressure within the IJ pseudocoel, and differential stretching and compression forces across the nematode cuticle. Release of this unilateral contraction, in conjunction with the correction of cuticle pressure, triggers enough momentum for an IJ to jump a distance of nine times body length, to a height of seven times body length [14]. Here we aimed to study the function of Sc-flp-21 in coordinating nictation and other behaviours relevant to host-finding.
The recent publication of five Steinernema spp. genomes, along with stage-specific transcriptomes [15] represents a valuable resource, alongside the previously published genomes of Oscheius sp. TEL-2014 [16] and Heterorhabditis bacteriophora [17]. The genome of Steinernema carpocapsae is the most complete, at an estimated 85.6 Mb, with predicted coverage of 98% [15]. S. carpocapsae was selected as a test subject for our study due to the quality of genome sequence. The close phylogenetic relationship between Steinernema spp. coupled with a diverse behavioural repertoire, particularly in terms of host-finding [18, 19], make this genus an extremely attractive model for comparative neurobiology. The aim of this study was to examine RNAi functionality in S. carpocapsae IJs, and to probe the involvement of FLP-21 in coordinating sensory perception (host-finding, nictation, jumping and dispersal phenotypes), as a prelim to probing the neuronal and molecular underpinnings of host-finding behaviour in this genus.
S. carpocapsae (ALL) was maintained in Galleria mellonella at 23°C. IJs were collected by White trap [20] in a solution of Phosphate Buffered Saline (PBS). Freshly emerged IJs were used for each experiment.
BLAST analysis of RNAi pathway components was conducted as in Dalzell et al. [21], using a modified list of core RNAi pathway components from C. elegans, against predicted protein sets and contigs of the S. carpocapsae genome, through the Wormbase Parasite BLAST server [22, 23].
Sc-flp-21 (Gene ID: L596_g19959.t1) dsRNA templates were generated from S. carpocapsae IJ cDNA using gene-specific primers with T7 recognition sites (see Table 1). Neomycin phosphotransferase (neo) and Green Fluorescent Protein (gfp) dsRNA templates were generated from pEGFP-N1 (GenBank: U55762.1). All dsRNA templates were size matched (200–220 bp). Template PCR products were generated as follows: [95°C x 10 min, 40 x (95°C x 30 s, 60°C x 30 s, 72°C x 30 s) 72°C x 10 min]. PCR products were assessed by gel electrophoresis, and cleaned using the Chargeswitch PCR clean-up kit (Life Technologies). dsRNA was synthesised using the T7 RiboMAX Express Large Scale RNA Production System (Promega), and quantified by a Nanodrop 1000 spectrophotometer.
1000 S. carpocapsae were incubated in 50 μl PBS with dsRNA and 50 mM serotonin (to stimulate pharyngeal pumping) across four experimental regimes; (i) 24 h in 5 mg/ml dsRNA / serotonin / PBS; (ii) 24 h in 5 mg/ml dsRNA / serotonin / PBS, followed by washes to remove the initial dsRNA, and 24 h recovery in PBS only; (iii) 48 h in 5 mg/ml dsRNA / serotonin / PBS; and (iv) 48 h in 1 mg/ml dsRNA and serotonin. Each experiment was conducted as five replicates at 23°C.
Total RNA was extracted from 1000 IJs using the Simply RNA extraction kit (Promega, UK) and Maxwell 16 extraction system (Promega, UK). cDNA was synthesised using the High Capacity RNA to cDNA kit (Applied Biosystems, UK). Each individual qRT-PCR reaction comprised 5 μl Faststart SYBR Green mastermix (Roche Applied Science), 1 μl each of the forward and reverse primers (10 μM), 1 μl water, 2 μl cDNA. PCR reactions were conducted in triplicate for each individual cDNA using a Rotorgene Q thermal cycler under the following conditions: [95°C x 10 min, 40 x (95°C x 20 s, 60°C x 20 s, 72°C x 20 s) 72°C x 10 min]. Primer sets were optimised for working concentration, annealing temperature and analysed by dissociation curve for contamination or non-specific amplification by primer—dimer as standard. The PCR efficiency of each specific amplicon was calculated using the Rotorgene Q software. Relative quantification of target transcript relative to two endogenous control genes (Sc-act and Sc-β-tubulin) was calculated by the augmented ΔΔCt method [24], relative to the geometric mean of endogenous references [25]. The most similar non-target gene (L596_g5821.t1) was identified using BLASTn against the S. carpocapsae genomic contigs (supplemental S1 Text), and primers Sc-L596_g5821.t1-f and Sc- L596_g5821.t1-r were used to assess transcript abundance relative to Sc-act across control and experimental conditions for the 48h dsRNA exposure experiments only (Table 1).
Approximately 5 g of fresh waxworm (Galleria mellonella) and mealworm (Tenebrio molitor) larvae were placed into 20 mL glass tubes and sealed. The holder needle was exposed to the headspace of the tube over a 120 min timecourse (extraction time) at room temperature (22°C). After this time, the SPME syringe was directly desorbed in the GC injection port for 5 min. A fused silica fibre coated with a 95 μm layer of carboxen—polydimethylsiloxane (CAR—PDMS; Supelco) was used to extract the volatile compounds from the samples. Fibres were immediately thermally desorbed in the GC injector for 5 min (with this time we desorb the analytes and re-activated the fiber for the next analysis) at 250°C and the compounds were analysed by GC-MS.
A CTC Analytics CombiPal autosampler was coupled to a 7890N Agilent gas chromatograph (Agilent, Palo Alto, California) and connected to a 5975C MSD mass spectrometer. The manual SPME holder (Supelco, Bellefonte, PA, USA) was used to perform the experiments. Chromatographic separation was carried out on 30 m x 0.25 mm I.D. ZB-semivolatiles, Zebron column (Phenomenex, Macclescfield, UK).The oven temperature was set at 40°C for 3 min, temperature increased from 40 to 250°C at 5°C min-1 and set at the maximum temperature for 4 min. Helium was used as carrier gas at 1 ml min-1. Mass spectra were recorded in electron impact (EI) mode at 70 eV. Scan mode was used for the acquisition to get all the volatile compounds sampled. Quadrupole and source temperature were set at 150 and 230°C respectively. Compounds were identified using MS data from the NIST library (>95% confidence).
100 S. carpocapsae IJs were placed in the centre of a 90 mm PBS agar plate (1.5% w/v) in a 5 μl aliquot of PBS. Plates were divided into four zones; a central zone 15 mm in diameter, and three further zones equally spaced over the remainder of the plate. Plates were allowed to air dry for ~5 min. Evaporation of the PBS allowed the IJs to begin movement over the agar surface. Lids were then placed back onto the Petri dishes, and plates were incubated at 23°C in darkness for one hour. IJs were counted across central and peripheral zones and expressed as percentage of total worms. Our subsequent analysis was conducted on total IJs found within the two central zones. Relative to those found in the two peripheral zones. Five replicate assays were conducted for each treatment.
3.5 g of compost (John Innes No.2) was placed in a petri dish (55 mm), and dampened evenly with 150 μl PBS. Approximately ten IJs were pipetted onto the compost in 5 μl of PBS, and left for 5 minutes at room temperature; this enabled IJs to begin nictating. For the waxworm volatile challenge, one healthy waxworm (UK Waxworms Ltd.) was placed inside a 1 mL pipette tip, without filter. For the mealworm volatile challenge, two mealworms (Monkfield Nutrition, UK), weight-matched to the waxworm, were placed inside a 1 mL pipette tip, without filter. Blank exposure data were captured using an empty 1 ml pipette tip, without filter. In each case, the pipette was set to eject a volume of 500 μL, comprising air and the corresponding insect volatiles. A binocular microscope was used to record IJ behavioural responses following up to five volatile exposures each, on gentle ejection from the pipette within a distance of ~1 cm of the S. carpocapsae IJs. A five second period was allowed between each volatile exposure. Recording ended for any individual when jumping was observed or the IJ abandoned a nictating stance (this always corresponded with migration away from the stimulus). A jumping index was calculated for each treatment group by counting the number of IJs which jump in response to any of the five volatile exposures [1]. Additional behavioural observations were recorded, and subsequently reported as percentage IJs displaying the behaviour over the course of up to five volatile exposures, or until the IJ migrated / jumped out of the field of vision. Five replicate assays were conducted for each treatment.
Two circular holes (approx. 6 mm diameter, centred 4 mm from edge of lid) were drilled either side of a 90 mm petri dish lid. Two microcentrifuge tubes (1.5 ml) with a small hole cut out the bottom (approx. 2mm diameter), were also used for each assay. 200 S. carpocapsae IJs were placed in the centre of a 90 mm PBS agar plate (1.5% w/v) in a 5 μl aliquot of PBS. The arena was segmented into positive and a negative zones either side of the plate (25 mm in length from the edge, circling off the plate at a point 60 mm apart; see Fig 1A). Plates were allowed to air dry for ~5 min, allowing the IJs to begin migration. The lid was placed on top of the plate, and sealed with parafilm. The 1.5 ml tubes were secured in the holes with parafilm; one remained empty, which we term the blank tube, and the other held four live Galleria mellonella fourth instar larvae, or four Tenebrio molitor larvae as appropriate. The lid of the tubes were then closed. The plates were incubated at 23°C in darkness for one hour. IJs were counted in the positive (host side) and negative (blank side) zones and then scored using a chemotaxis index [26]. The assay format was adapted from Grewal et al. [1994] [27]. Five replicate assays were conducted for each treatment.
Freshly emerged S. carpocapsae IJs were fixed in 4% paraformaldehyde overnight at 4°C, followed by a brief wash in antibody diluent (AbD; 0.1% bovine serum albumin, 0.1% sodium azide, 0.1% Triton-X-100 and PBS pH 7.4). The fixed specimens were roughly chopped on a glass microscope slide with a flat edged razor, and incubated in primary polyclonal antiserum raised against GLGPRPLRFamide, N-terminally coupled to KLH, and affinity purified (1:800 dilution in AbD) for 72 h at 4°C. Subsequently, chopped IJs were washed in AbD for 24 h at 4°C, and then incubated in secondary antibody conjugated to fluorescein isothiocyanate (1:100 dilution in AbD) for 72 h at 4°C. A further AbD wash for 24 h at 4°C was followed by incubation in Phalloidin—Tetramethylrhodamine B isothiocyanate (1:100 dilution in AbD) for 24 h at 4°C. Finally, chopped IJs were washed in AbD for 24 h at 4°C. Specimens were mounted onto a glass slide with Vectasheild mounting medium and viewed with a Leica TCS SP5 confocal scanning laser microscope. Controls included the omission of primary antiserum, and pre-adsorption of the primary antiserum with ≥250 ng of GLGPRPLRFamide. Pre-adsorption of the primary antiserum in GLGPRPLRFamide resulted in no observable staining.
PCR primers were designed to amplify a 200 bp region of Sc-flp-21 (Gene ID: L596_g19959.t1) from S. carpocapsae IJ cDNA. Template PCR products were generated as follows: [95°C x 10 min, 40 x (95°C x 30 sec, 60°C x 30 sec, 72°C x 30 sec) 72°C x 10 min]. PCR products were assessed by gel electrophoresis, and cleaned using the Chargeswitch PCR clean-up kit (Life Technologies). Amplicons were quantified by a Nanodrop 1000 spectrophotometer. Sense and antisense probes were generated using amplicons in an asymmetric PCR reaction. The components of each reaction were as follows: 2.0μl of Reverse primer (or Forward primer for control probe); 2.5μl 10X PCR buffer with MgCl2 (Roche Diagnostics); 2μl DIG DNA labelling mix (Roche Diagnostics); 0.25μl 10x Taq DNA polymerase (Roche Diagnostics); 20ng DNA template; distilled water to a volume of 25μl. Probes were assessed by gel electrophoresis.
Freshly emerged S. carpocapse IJs were fixed in 2% paraformaldehyde in M9 buffer overnight at 4°C followed by 4h at room temperature. Nematodes were chopped roughly using a sterile razor blade for 2 minutes and washed three times in DEPC M9. Subsequently, the chopped nematodes were incubated in 0.4 mg/ml proteinase K (Roche Diagnostics) for 20 minutes at room temperature. Following three washes in DEPC M9, the nematodes were pelleted (7000g) and frozen for 15 minutes on dry ice. Subsequently the nematode sections were incubated for 1 minute in -20°C methanol and then in -20°C acetone for 1 minute. The nematodes were then rehydrated using DEPC M9 and incubated at room temperature for 20 minutes, after which three wash steps in DEPC M9 were carried out to remove any acetone.
The nematodes were pre-hybridised in 150 μl hybridisation buffer [prepared as detailed by Boer et al., 1998] for 15 minutes. The hybridisation probes were heat denatured at 95°C for 10 minutes, after which they were diluted with 125 μl hybridisation buffer. The probe-hybridisation mixture was then added to the nematode sections which were incubated at 50°C overnight. Post hybridisation washes were carried out as follows: three washes in 4x Saline Sodium Citrate buffer (15 minutes, 50°C); three washes in 0.1x SSC/0.1x Sodium dodecyl sulphate (20 minutes, 50°C) and; 30 minute incubation in 1% blocking reagent (Roche Diagnostics) in maleic acid buffer (50°C). Subsequently the nematodes were incubated at room temperature for 2 h in alkaline phosphatase conjugated anti-digoxigenin antibody (diluted 1:1000 in 1% blocking reagent in maleic acid buffer). Detection was completed with an overnight incubation in 5-Bromo-4-chloro-3-indolyl phosphate/Nitro blue tetrazolium at 4°C. The staining was stopped with two washes in DEPC treated water. The nematode sections were mounted on to glass slides for visualisation.
Data pertaining to both qRT-PCR and behavioural assays were assessed by Brown-Forsythe and Bartlett’s tests to examine homogeneity of variance between groups. One-way or two-way ANOVA was followed by Bonferroni’s multiple comparisons test. All statistical tests were performed using GraphPad Prism 6.
As is the case for other parasitic nematode species, S. carpocapsae was found to encode a less diverse RNAi pathway than that of C. elegans, in terms of gene for gene conservation [21]. However, the apparent reduction in AGO homologue diversity is offset by significant expansions across several putative ago genes, to give a predicted overall increase in the S. carpocapsae AGO complement (38 in total), relative to C. elegans (24, not including pseudogenes) [28]; WAGO-1 (nineteen), ALG-1 (three), ALG-3 (two), WAGO-5 (four), WAGO-10 (two), WAGO-11 (three) are all expanded relative to C. elegans. Notably, no identifiable homologue of RDE-1, the primary AGO for exogenously triggered RNAi events in C. elegans, could be identified (refer to S1 and S2 Tables).
The presence of PRG-1 and components of the piwi interacting (pi)RNA biosynthetic machinery suggests that a functional piRNA (or 21U RNA) pathway may be present. Whilst ERGO-1 is not conserved, two putative ALG-3 orthologues suggest that a functional endo-siRNA (26G RNA) pathway may also exist, which is supported by broad conservation of associated proteins. MicroRNA-associated AGOs, ALG-1 and ALG-2 are conserved, with a small apparent expansion of ALG-1 to three related proteins in S. carpocapsae. Further understanding of how RNAi pathway complements influence functionality will require small RNA sequencing efforts, and functional genomics approaches.
The RNA-dependent RNA Polymerase (RdRp), RRF-3 is conserved, and known to function antagonistically to exogenously primed RNAi, through competing activity for pathway components required for both exogenous RNAi, and the endo-siRNA (26G RNA) pathway within which RRF-3 operates [29–31]. The RdRps, RRF-1 and EGO-1, which are involved in the biosynthesis of secondary siRNAs (22G RNAs) are also conserved. Loss of the argonaute ERGO-1 which functions upstream of secondary siRNA biogenesis in the endo-siRNA (26G RNA) pathway in C. elegans, also leads to an exogenous ERI phenotype (Enhanced RNAi), but is not conserved in S. carpocapsae, suggesting that ALG-3 / -4 may be solely responsible for endo-siRNA functionality [32, 33].
The apparent absence of the intestinal dsRNA transporter, SID-2 is consistent with findings from other parasitic nematodes [21, 34, 35]. SID-1 also appears to be absent, however CHUP-1, a putative cholesterol uptake protein which contains a SID-1 RNA channel is present, and may assist in the intercellular spread of dsRNA. RSD-3, which also effects intercellular spread of dsRNA is conserved (see Fig 2 for pathway overview and S2 Text).
Various treatment regimens were employed in order to assess the responsiveness of S. carpocapsae IJs to exogenous dsRNA. 24 h incubation in 5 mg/ml dsRNA, with 50 mM serotonin was not sufficient to trigger statistically significant Sc-flp-21 knockdown (Fig 3A), however a 24 h dsRNA / serotonin incubation followed by a 24 h recovery in PBS only, did trigger a small decrease in Sc-flp-21 relative to Sc-act when compared to gfp and neo dsRNA controls (0.70 ±0.11, P<0.05) (Fig 3B). Extended incubation of S. carpocapsae IJs in 5 mg/ml dsRNA and 50 mM serotonin for 48 h triggered robust knockdown of Sc-flp-21 (0.16 ±0.07, P<0.0001) (Fig 3C). 48 h incubation in 1 mg/ml dsRNA, with 50 mM serotonin also triggered significant levels of Sc-flp-21 knockdown (0.49 ±0.27, P<0.01), however this was not as effective as the 5 mg/ml dsRNA treatment (Fig 3D). A BLAST analysis identified predicted S. carpocapsae transcript L596_g5821.t as the non-target gene with most similarity to the Sc-flp-21 dsRNA (S1 Text). The relative expression level of L596_g5821.t1 was unaffected by a 48 h incubation in 5 mg/ml Sc-flp-21 dsRNA with 50 mM serotonin, relative to neo and gfp dsRNA (1.013 ±0.04, P>0.05) (Fig 3E).
Comprehensive volatile signatures were characterised, and significant differences noted between G. mellonella and T. molitor larvae. In total, we identified 9 compounds unique to G. mellonella, four compounds unique to T. molitor, and 15 compounds shared between both species (Table 2). These profiles significantly expand on the number of volatiles identified from headspace GC-MS data presented by Hallem et al. [36] for the same insect species. A semi-quantitative analysis of detected volatiles can be found in supplementary S1 Data.
S. carpocapsae IJs were challenged by exposure to volatiles from G. mellonella or T. molitor following RNAi (48 h 5 mg/ml dsRNA, 50 mM serotonin) and control treatments. A decrease in hyperactive nictation following Sc-flp-21 knockdown was observed (10% ±5.774) relative to untreated (40.75% ±6.75; P<0.01) and neo dsRNA treatment (47.5% ±2.5; P<0.01) following G. mellonella volatile challenge (Fig 4A). Likewise, a decrease in hyperactive nictation was observed following T. molitor volatile challenge to Sc-flp-21 RNAi IJs (5.0% ±2.9), relative to untreated (57.25% ±2.8; P<0.0001) and neo dsRNA treatment (35.0% ±6.5; P<0.001) (Fig 4B). A decrease in the jumping index of IJs following Sc.flp-21 dsRNA treatment was observed when challenged by G. mellonella volatiles (0.08 ±0.02) relative to untreated (0.72 ±0.09; P<0.001) and neo dsRNA treated (0.55 ±0.06; P<0.01) (Fig 4C). Similarly, a decrease in jumping index as a response to T. molitor volatiles was observed following Sc-flp-21 RNAi (0.03 ±0.02) relative to untreated (0.46 ±0.05; P<0.001) and neo dsRNA treatment (0.4 ±0.04; P<0.001) (Fig 4D). An agar host-finding assay was used to further assess the impact of flp-21 knockdown. A decrease in G. mellonella finding ability was observed (0.06 ±0.08) relative to untreated (0.53 ±0.03; P<0.001) and neo dsRNA treated (0.42 ±0.08; P<0.01) (Fig 4E). Likewise, a decrease in T. molitor finding ability was observed (0.01 ±0.06) relative to untreated (0.32 ±0.04; P<0.01) and neo dsRNA treated (0.26 ±0.1; P>0.01) (Fig 4F). It was also found that Sc-flp-21 RNAi resulted in significantly decreased lateral dispersal, relative to both untreated and neo dsRNA treatment (P<0.0001) (Fig 4G). In all instances, dsRNA treatment regimens which triggered lower levels of Sc-flp-21 knockdown relative to the 48h 5 mg/ml dsRNA, 50 mM serotonin approach, failed to trigger null phenotypes.
flp-21/FLP-21 was localised exclusively to paired neurons within the central nerve ring region of S. carpocapsae IJs. Without additional neuroanatomical information on S. carpocapsae IJs it is impossible to further define these cells, however, based on the immuocytochemical localisation the cells appear to project posteriorly (Fig 5). These data suggest that FLP-21 must act as a modulator of sensory function, downstream of the primary chemosensory neurons (amphids).
RNA interference is an extremely important tool for the study of gene function in parasitic nematodes [37, 38]. Three independent reports of a functional RNAi pathway in the entomopathogenic nematode Heterorhabditis bacteriophora have been published. Ciche and Sternberg [39] assessed the efficacy of RNAi through soaking egg / L1 stage H. bacteriophora in 5–7.5 mg/ml dsRNA targeting a number of genes which had been selected on the basis of phenotypic impact on the model C. elegans. Demonstrable phenotypes and target transcript knockdown signified an active pathway. Moshayov, Koltai and Glazer [40] employed the methodology of Ciche and Sternberg [39] to study the involvement of genes in the regulation of IJ exsheathment (or ‘recovery’). Subsequently, Ratnappan et al. [41] demonstrated that microinjection was also a suitable method for introducing dsRNA into hermaphrodite gonads, effectively triggering the RNAi pathway in F1 progeny. To date, no such assessment of a functional RNAi pathway has been published for Steinernema spp.
The RNAi pathway of S. carpocapsae has been characterised by BLAST and validated through silencing Sc-flp-21 in IJs. Our data indicate that neuronal cells are sensitive to RNAi in S. carpocapsae IJs, and that knockdown is highly sequence specific. Like other parasitic nematodes S. carpocapsae encodes an expanded set of WAGO-1 (R06C7.1) family AGOs (19 in total) which function primarily with secondary siRNAs (22G RNAs) in C. elegans, along with CSR-1 which is also conserved. Whilst RDE-1 is primarily responsible for triggering the onset of an exogenous RNAi response, acting upstream of secondary siRNAs (22G RNAs), it is not conserved in S. carpocapsae [21, 31]. Our observation of RNAi sensitivity in S. carpocapsae reveals that RDE-1 is not required to trigger an exogenous RNAi response, however the functional significance of AGO homologue expansions relative to C. elegans remains to be determined. The lack of SID-2 seems to correlate with our observation that relatively high amounts of dsRNA are required to trigger the RNAi pathway by oral delivery.
The nearest non-target gene sequence within the S. carpocapsae genome represents an uncharacterised predicted gene (L596_g5821.t1). The Sc-flp-21 dsRNA shared high levels of sequence similarity over a 21 bp stretch of L596_g5821.t1 (20 of 21 bp shared), however qRT-PCR indicates that L596_g5821.t1 had not been silenced, which could suggest: (i) the level of sequence similarity was either insufficient for gene knockdown; (ii) dsRNA was not diced in the correct register to produce this exact 21 bp sequence within a significant population of siRNAs; or (iii) the L596_g5821.t1 gene is not expressed in cells / tissue which is sufficiently susceptible to dsRNA delivered under the conditions tested. In order to trigger significant knockdown of Sc-flp-21, 48h continuous exposure to dsRNA was required in the presence of 50 mM serotonin. Reducing dsRNA exposure time lead to a corresponding reduction in Sc-flp-21 knockdown, as did a reduction of dsRNA amount from 5mg/ml to 1mg/ml over a 48h time-course. Phenotypes which developed following 48h dsRNA exposure were not observed across any of the experimental variations which resulted in decreased gene knockdown (shorter exposure timeframes / lower dsRNA amounts). This has potentially important implications for RNAi experimental design in other parasitic nematodes, and notably in C. elegans, for which the validation of gene knockdown by qRT-PCR is not common across the literature. Undoubtedly false negative determinations of gene function will be a problem in this context. Our data demonstrate that statistically significant gene knockdown levels are not necessarily sufficient to reveal gene function; careful consideration should be given to the design of RNAi experiments as a result.
The neuronal RNAi sensitivity of S. carpocapsae IJs, and the ease of behavioural assays makes these species ideal models for studying the neurobiology of sensory perception and host-finding behaviours. Within the Steinernematid EPNs, a number of species also display a highly specialised jumping behaviour which can be triggered in nictating IJs on exposure to host Insect volatiles [18]. Silencing Sc-flp-21 triggers pleiotropic effects on sensory behaviours of relevance to host-finding, lateral dispersal, hyperactive nictation and jumping phenotypes. The waxworm and mealworm headspace SPME GC-MS profiles are significantly expanded relative to those presented by Hallem et al. [36] These data could provide a valuable tool for comparative analysis of neurobiology and host-finding behaviours across EPN species.
We find that flp-21 / FLP-21 is localised exclusively to paired neurons in the anterior of the IJ using whole mount in situ hybridisation and immunocytochemistry. This represents the most restricted flp-21 / FLP-21 expression pattern observed in a nematode to date. FLP-21 is expressed in several anterior sensory and motor neurons in C. elegans, where it is known to coordinate aspects of sensory perception [11]. The FLP-21 homologues of two plant parasitic nematodes, Globodera pallida, and Meloidogyne incognita (GSLGPRPLRFamide) are expressed in the sensory amphid neurons and across the central nerve ring of infective stage J2s [42]. These data support a broad role for FLP-21 in coordinating sensory perception across different nematode species. The ICC localisation of FLP-21 in S. carpocapsae reveals positive immunostaining within the two cell bodies, and posteriorly along the axons, terminating at paired synapses at points around the terminal bulb of the pharynx. Additional effort must focus on understanding the neuroanatomy of entomopathogenic nematodes in order to understand these data more fully, and exploit this platform for comparative neurobiology.
Collectively, these data provide the first mechanistic insight to EPN sensory behaviour, which may have implications for biocontrol efficacy. Through isolating genes and signalling pathways which coordinate these behaviours, efforts to identify molecular markers of desired behaviours and traits could facilitate the identification of more suitable isolates and strains for biocontrol use, and the enhancement of current strains through selective breeding / mutagenic approaches. The selection or manipulation of behavioural tendencies could lead to strains which are capable of operating within new ecological niches, expanding their utility. More broadly, these data suggest a broad role for FLP-21 in coordinating sensory perception and host-finding behaviours which may be relevant to other economically important parasites of plant and mammal.
|
10.1371/journal.pgen.1000194 | corona Is Required for Higher-Order Assembly of Transverse Filaments into Full-Length Synaptonemal Complex in Drosophila Oocytes | The synaptonemal complex (SC) is an intricate structure that forms between homologous chromosomes early during the meiotic prophase, where it mediates homolog pairing interactions and promotes the formation of genetic exchanges. In Drosophila melanogaster, C(3)G protein forms the transverse filaments (TFs) of the SC. The N termini of C(3)G homodimers localize to the Central Element (CE) of the SC, while the C-termini of C(3)G connect the TFs to the chromosomes via associations with the axial elements/lateral elements (AEs/LEs) of the SC. Here, we show that the Drosophila protein Corona (CONA) co-localizes with C(3)G in a mutually dependent fashion and is required for the polymerization of C(3)G into mature thread-like structures, in the context both of paired homologous chromosomes and of C(3)G polycomplexes that lack AEs/LEs. Although AEs assemble in cona oocytes, they exhibit defects that are characteristic of c(3)G mutant oocytes, including failure of AE alignment and synapsis. These results demonstrate that CONA, which does not contain a coiled coil domain, is required for the stable ‘zippering’ of TFs to form the central region of the Drosophila SC. We speculate that CONA's role in SC formation may be similar to that of the mammalian CE proteins SYCE2 and TEX12. However, the observation that AE alignment and pairing occurs in Tex12 and Syce2 mutant meiocytes but not in cona oocytes suggests that the SC plays a more critical role in the stable association of homologs in Drosophila than it does in mammalian cells.
| Meiosis is a specialized type of cell division that is needed to produce sperm and egg cells, which carry only half the number of chromosomes of other cells in the body. Meiosis is required for reproduction, but abnormalities in chromosome number caused by errors in the process of meiosis are responsible for many birth defects and mental retardation syndromes in humans. The fruit fly, Drosophila melanogaster, is an excellent organism in which to study meiosis because of the powerful genetic and microscopic techniques that can be implemented with it. Early in meiosis, homologous chromosomes are joined together by an elaborate protein structure called the synaptonemal complex (SC) that plays a critical role in both holding homologous chromosomes together and in facilitating a process known as meiotic recombination. In this study, we have found a protein called Corona that is required for the formation of the SC. Our data show that Corona is required for the proper localization of the SC protein C(3)G. In the absence of Corona, C(3)G fails to polymerize and form the central region of the SC. Increasing our understanding of SC assembly and function will lead to a better understanding of the mechanism for proper chromosome segregation during meiosis.
| During meiosis, the diploid genome is segregated to form haploid nuclei through processes that include the close juxtaposition of homologous chromosomes and recombination between them. In most organisms, a proteinaceous structure called the synaptonemal complex (SC) forms between homologous chromosomes during meiotic prophase. The SC is required for synapsis, the intimate association of homologs along their entire length. The SC and its components are thought to play roles in regulating recombination and generally promoting the establishment of crossovers between the chromosomes [1],[2].
Examination of SCs by electron microscopy (EM) has defined distinct structures present in the SCs of most organisms. During early prophase, axial elements (AEs) form along the longitudinal axis of each pair of sister chromatids using a cohesin-based chromosome core as a scaffold for assembly [3]. As prophase progresses, the AEs of homologous chromosomes become physically connected by perpendicular transverse filaments (TFs) that span the SC central region (CR), which occupies the ∼100 nm space between the two homologous AEs [1],[2]. AEs within the mature SC are referred to as lateral elements (LEs). Finally, a central element (CE) is often observed as a structure overlapping the middle of the TFs and positioned parallel to the two LEs.
Although homologous chromosomes undergo presynaptic pairing and alignment in some organisms [4],[5], synapsis requires a fully formed CR that extends the length of the chromosomes. In this paper we will use the term “alignment” to describe the parallel association of homologs (or AEs) at a distance equal to or greater than the width of the SC and the term “pairing” to describe the close association of homologous sequences as determined by FISH.
Components of TFs, such as ZIP1 (budding yeast), SYCP1 (mouse), SYP-1 (worms), and C(3)G (Drosophila), have been identified as proteins containing long coiled coil segments [6]–[11]. Although these TF proteins are similar in predicted secondary structure, they share very little similarity in amino acid sequence. Despite this lack of sequence conservation, the proteins are all thought to form TFs across the CR of the SC by binding of their C-termini to the AEs with head-to-head orientation of their N-termini at the center of the CE [12]–[16]. TFs are important for ensuring synapsis of homologs and normal levels of interhomolog exchange [6], [8], [10], [17]–[20].
In Drosophila melanogaster oocytes, the TFs are formed by the C(3)G protein [8],[12]. Like other TF proteins, C(3)G is comprised of a central coiled coil-rich domain flanked by N- and C-terminal globular domains. As shown by Jeffress et al. [21], C-terminal deletion of C(3)G results in its failure to attach to the AEs of each set of homologs. Instead, this C-terminal deletion protein forms a large cylindrical polycomplex structure. EM analysis of this structure reveals a polycomplex of concentric rings with alternating dark and light bands, presumably corresponding to long arrays of polymerized TFs. However, deletions of N-terminal regions completely abolished both SC and polycomplex formation. To explain these data, Jeffress et al. [21] proposed that in Drosophila, the N- terminal globular domain of C(3)G is critical for the formation of anti-parallel pairs of C(3)G homodimers, and thus for assembly of complete TFs, while the C-terminus is required to affix these homodimers to the AEs.
The question then arises as to how C(3)G molecules can be polymerized to form a linear array of TFs. The idea that such polymerization events are facilitated by the apposition of paired AEs seems unlikely given the finding that C-terminal deletions of C(3)G form polycomplexes [21]. The observation that the rat homolog of C(3)G (SYCP1) can form polycomplex-like structures in COS-7 cells [22] suggests that the process of TF polymerization may be self-promoting and sustaining, and thus requires no other components. However, in mice, significant extension of SYCP1 to form a full-length CR in meiotic cells requires the functions of the SYCE1, SYCE2, and TEX12 proteins, which localize to the CE of the SC [23]–[26]. SYCE1 and SYCE2 physically associate with each other and the N-terminus of the TF protein SYCP1, while TEX12 binds to SYCE2 [24],[25]. Mice lacking the SYCE2 protein display defects in the formation of the TFs (SYCP1 accumulation), and thus in SC formation [23]. They appear to form only short and, at least in the case of Tex12−/− mice, morphologically abnormal SCs [23],[26]. It remains to be determined whether or not functional homologs of SYCE2 and TEX12 might facilitate C(3)G polymerization, and thus CE formation in Drosophila oocytes.
To discover additional components of the SC and genes involved in other critical processes in meiosis, we previously undertook a novel genetic screen for female meiotic mutants in Drosophila [27]. One of the genes identified in the screen, corona (cona), was found to have both a severe defect in meiotic recombination and a profound effect on the localization of C(3)G. Previous analyses of cona mutants demonstrated a failure of the SC protein C(3)G to localize correctly in the absence of CONA, demonstrating defective SC formation. As is the case for c(3)G mutants, the frequency of meiotic exchange in cona females was reduced 50- to 200-fold compared to wild-type [27] without a similar reduction in the number of DSBs [SLP and RSH, unpublished data]. Moreover, double mutants for c(3)G and cona displayed a defect in recombination that was comparable to either single mutant [SLP and RSH, unpublished data], and thus the two proteins likely function in a common pathway with respect to facilitating meiotic exchange. Like C(3)G, CONA protein is only conserved within the genus Drosophila, but CONA contains no predicted coiled coil domains or other characterized functional motifs [27].
In this study, we show that CONA is a new SC protein that co-localizes with C(3)G in a mutually-dependent fashion. We found that CONA accumulation is required for C(3)G localization into wild-type SC structures and formation of polycomplexes, but is not necessary for the formation of either the AEs or the chromosome cores from which they arise. Our results indicate that CONA is crucial for the assembly of the CR of the SC in Drosophila and may have a function similar to that of the vertebrate CE proteins TEX12 and SYCE2. However, the observation that pairing and alignment of AEs occurs in Tex12 and Syce2 mutant meiocytes, but not in cona oocytes, suggests that the SC plays a more critical role in the stable association of homologs in Drosophila than it does in mammalian cells.
We previously showed that the cona gene corresponds to the transcription unit CG7676 on the basis of the presence of a Doc transposon in the 3′ untranslated region of CG7676 in conaA12 that was not present on the un-mutagenized parental chromosome and the isolation of a second allele, conaf04903, which bears a PiggyBac insertion in sequence flanking the 5′ end of CG7676 [27]. Both conaA12 and conaf04903 drastically reduce the levels of meiotic recombination and produce high levels of nondisjunction (∼32–39%) [27, SLP and RSH, unpublished data].
We raised an antibody against the CONA protein and used it to determine the localization of CONA in meiotic prophase cells in the germaria of Drosophila ovaries (see Materials and Methods). Evidence that this antibody is specific to CONA (i.e., that no signal is observed in pro-oocytes homozygous for conaf04903) is presented in Figure S1. In wild-type ovaries, anti-CONA localization was observed within a subset of nuclei in regions 2A and 2B of the germarium and within the oocyte nucleus in region 3 and early egg chambers within the vitellarium. The distribution of CONA within nuclei was distinctly thread-like and strongly co-localized with the SC protein C(3)G (Figure 1A). These results demonstrate that CONA localizes along the SC.
As an alternate strategy to localize the protein, we constructed a transgene, P{UASP-cona::Venus}, which expresses the full-length CONA protein fused to the yellow fluorescent protein Venus under the control of the GAL4/UAS system [28]. The CONA::Venus fusion protein was functional, as expression driven by nanos-GAL4::VP16 in the female germline rescued the nondisjunction phenotype in conaf04903 homozygotes. Control conaf04903 homozygotes lacking the P{UASP-cona::Venus} transgene showed 31.9% X chromosome nondisjunction, whereas conaf04903 homozygotes expressing CONA::Venus showed a nearly tenfold reduction in nondisjunction to just 3.4% (data not shown).
We examined the pattern of nanos-GAL4::VP16-driven CONA::Venus localization during meiotic prophase. In a conaf04903 mutant background, strong Venus yellow fluorescent protein signal localized in a pattern very similar to that observed for CONA immunolocalization. Immunolocalization of C(3)G in these ovaries revealed extensive co-localization of CONA::Venus and C(3)G in thread-like patterns within nuclei (Figure 1B–C). Nuclear CONA::Venus fluorescence was strongest in a cona mutant background in which little or no wild-type CONA protein is present (Figure S1). When expressed in heterozygotes or homozygotes for a wild-type copy of the cona locus, CONA::Venus fluorescence was weaker in nuclei and increased diffuse fluorescence was often observed in the cytoplasm of germline cells in regions 1 and 2 of the germarium (data not shown). This may be the result of competition with wild-type CONA protein. A similar reduction in signal has been observed for localization of GFP-tagged ORD protein along the SC in the presence of wild-type ORD (RSK and SEB, unpublished data). These data confirm the immunolocalization of CONA and implicate CONA as a component of the SC.
When CONA::Venus was expressed under the control of a nanos-GAL4::VP16 driver in a conaf04903 heterozygote in which wild-type CONA protein was also present, C(3)G was detected as puncta and short threads within early prophase nuclei before CONA::Venus signal was detectable (Figure 2A–B). The spotty to thread-like pattern of C(3)G accumulation observed in Figure 2B is also observed in early region 2A in conaf04903/+ heterozygotes that lack the CONA::Venus transgene, and represents an early stage (zygotene) in SC assembly in which the short threads of C(3)G are associated with endogenous CONA (Figure S1C). As the intensity of CONA::Venus staining increased during the progression of meiotic prophase, CONA::Venus assumed a thread-like staining pattern that co-localized with C(3)G (Figure 2E–F, 2I–J).
A different pattern of C(3)G localization was observed when CONA::Venus was expressed in the conaf04903 homozygote, and therefore was the only form of functional CONA protein present. In nuclei that contained very low or undetectable levels of CONA::Venus signal (Figure 2C–D), C(3)G staining exhibited a more diffuse appearance similar to that previously described for cona mutant pro-oocytes [27]. However, as CONA::Venus staining became more visible at slightly later stages, CONA::Venus and C(3)G co-localized in short thread-like segments and the diffuse C(3)G signal was no longer observed (Figure 2G-H). Eventually, CONA::Venus and C(3)G co-localization resembled that observed in the conaf04903 heterozygote in pachytene nuclei with fully assembled SC (Figure 2K–L). These data further demonstrate that the assembly of C(3)G into a thread-like SC requires the accumulation of CONA and involves co-localization with the CONA protein.
The AEs are believed to form from chromosome core structures that contain sister chromatid cohesion proteins [3]. In mammals, AE-specific proteins such as SYCP2 and SYCP3 associate with components of the cohesin complex during the initial steps of SC assembly [13], [29]–[37]. Similarly, cohesin-based chromosomal cores/AEs form in Drosophila pro-oocytes [38]. Formation of the chromosomal core in Drosophila is dependent on the product of the c(2)M gene, which also localizes along this structure [12],[38],[39]. ORD protein also localizes along chromosome cores and is required for the maintenance of chromosome core integrity during meiotic prophase [38],[40]. Mutants in AE/LE proteins often result in recombination defects and the failure of synapsis, which indicates that properly formed AEs/LEs are required for the normal formation of the SC central region [31], [39], [41]–[44].
To better understand the association of CONA with the SC, we examined the localization of the CONA protein in mutants that disrupt different components of the AE. Mutations in the c(2)M gene result in incomplete SC formation, as indicated by only very short segments of nuclear C(3)G localization, in contrast to the long, thread-like localization observed in wild-type [39]. In contrast, in ord mutants, the thread-like C(3)G staining appears to disassemble earlier than in wild-type due to the dissolution of cohesin-based chromosome cores along the chromosome arms [38],[40].
Analysis of CONA localization in c(2)MEP2115 homozygous pro-oocytes demonstrated that CONA was localized in numerous short segments that corresponded to sites of C(3)G localization (Figure 3A). CONA was consistently co-localized with C(3)G and was not seen to localize elsewhere in the germarium except to the dot-like short segments of C(3)G. The observed localization of CONA in c(2)MEP2115 homozygotes could indicate a robust association with C(3)G and/or an inability to localize to abnormally formed AEs except at sites stabilized by C(3)G accumulation. Nonetheless, the dependency on AEs for localization is similar for both CONA and C(3)G.
CONA localization was also analyzed in ord5/ord10 trans-heterozygotes, in which no ord activity is present [40]. In agreement with published data [40], we found that C(3)G formed extensive thread-like patterns of localization in pro-oocyte nuclei in region 2A (Figure 3B), but appeared as shorter segments in older germline cysts beginning in late region 2B (Figure 3C). Oocyte nuclei in region 3 displayed C(3)G signals that were further shortened or dot-like, indicating early SC disassembly. At all stages, CONA was always observed to co-localize with C(3)G within the germarium. The initial co-localization of CONA with C(3)G in region 2A was thread-like, similar to wild-type, but older germline cysts did not reveal differences in the extent of localization of the two proteins, suggesting that removal of CONA occurred contemporaneously with C(3)G removal. These results indicate that CONA and C(3)G behave similarly in both c(2)M and ord mutant backgrounds and suggest that CONA and C(3)G may comprise parts of the same SC sub-structure.
The consistent co-localization of CONA and C(3)G in wild-type and mutant backgrounds and the requirement of CONA for proper C(3)G localization prompted the question of whether C(3)G is required for CONA localization. If CONA is a protein of the AE/LE that is required for C(3)G attachment, it would be expected to localize to chromosomes regardless of whether C(3)G is present or not. When CONA localization was examined in females homozygous for the null mutation c(3)G68, we found no evidence of CONA antibody staining in pro-oocyte nuclei (Figure 3D). This result is unlike that observed for the AE/LE component C(2)M [21],[39] and suggests that CONA does not act as an AE/LE component that localizes independently of C(3)G. Instead, these data are consistent with a role for CONA within the CR of the SC, which would not be expected to form in the absence of C(3)G.
To further investigate the role of CONA in SC formation, we investigated whether the SC protein C(2)M and the cohesion proteins ORD and SMC1 are able to localize normally in the absence of wild-type cona. All three proteins associate with the AEs/LEs of the SC in wild-type [12],[40]. In these experiments, we considered two aspects of ORD, SMC1, and C(2)M localization: first, whether the proteins localized to chromosomes and second, whether the localization appeared equivalent to that observed in wild-type in which normal SC is present. We utilized chromosome spread preparations in which soluble nuclear proteins are removed and only chromosome-associated proteins remain [38]. As shown in Figure 4A, SMC1 and ORD are able to stably associate with the meiotic chromosomes in cona mutant pro-oocytes. Both cohesion proteins accumulate normally at the centromeres as evidenced by the bright foci present in both wild-type and cona mutant nuclei. Although distinct thread-like staining along the chromosome cores is also visible, the threads of staining appear to be thinner and more numerous than those in wild-type. A similar pattern was also observed for C(2)M localization (Figure 4B). These data suggest that ORD, SMC1, and C(2)M localize to chromosomes and form chromosome cores/AEs in the absence of CONA. However, their localization does not appear equivalent to wild-type, most likely because the AEs do not align and pair. A similar localization pattern for AE/LE proteins has been reported for c(3)G mutant oocytes [38],[39].
We also examined C(3)G localization to determine whether C(3)G protein can associate with chromatin in the absence of CONA. Although the C(3)G signal on cona spreads is diminished compared to wild-type, and long continuous thread-like staining is absent, puncta and short fragments of chromosome-associated C(3)G are visible. In many cases, these short stretches coincide with C(2)M, SMC1, and ORD (Figure 4A–B). Together, these results argue that CONA is not required for the localization of ORD, SMC1, or C(2)M to chromosomes or for the formation of the AEs. However, our data suggest that in the absence of CONA activity, association of C(3)G with AEs is insufficient for assembly of a normal SC central region and the pairing/alignment of AEs.
The co-localization with C(3)G, the profound effect on C(3)G localization, and the minor effect on AE protein localization led us to postulate that CONA localizes within the CR of the SC rather than along AEs. Based on this hypothesis, we predicted that CONA would co-localize with C(3)G protein that is prevented from binding to AEs. C(3)G is thought to interact with AEs via its C-terminal globular domain, which is normally oriented toward the AEs [12]. Jeffress and colleagues [21] found that a deletion of amino acids 651–744 at the C-terminal end of C(3)G (known as C(3)GCdel) abolished the ability for C(3)G to form normal SC along chromosomes, but instead the protein accumulated into aggregates called polycomplexes (PCs). The PCs formed by the C(3)GCdel protein often take on a hollow cylindrical shape, and may form in either the presence or absence of wild-type C(3)G protein.
We analyzed CONA localization in C(3)GCdel PCs by immunofluorescence in females expressing the C(3)GCdel protein in the absence of wild-type C(3)G. As expected, the C(3)GCdel protein was detected in sub-cellular bodies of varying size, which correspond to the PCs, and not in a thread-like pattern along chromosomes. Similarly, strong CONA immunofluorescence was detected on the PCs, but not along chromosomes (Figure 5A). This demonstrates that amino acids 651–744 at the C-terminus of C(3)G are dispensable for CONA co-localization and that CONA does not localize along AEs or chromosome cores in the absence of wild-type C(3)G.
Since CONA is necessary for the assembly of wild-type C(3)G into normal SC, and CONA co-localizes with both C(3)G in wild-type and with C(3)GCdel in PCs, we tested whether CONA is required for the formation of the PCs. Using antibodies specific to the coiled coil region of C(3)G to detect both wild-type C(3)G and C(3)GCdel (Figure 5B and Figure S2A), we examined germaria from females expressing C(3)GCdel in a c(3)G68 conaf04903 double mutant background. Expression of the C(3)GCdel protein results in PC formation in a c(3)G68 single mutant background (Figure 5C and Figure S2B). However, when CONA was absent in the c(3)G68 conaf04903 double mutant, no anti-C(3)G immunofluorescence was visible above background levels, even though pro-oocyte nuclei could be detected by anti-SMC1 staining (Figure 5D and Figure S2D). The diffuse C(3)G staining observed in cona mutants was also not visible in this experiment, possibly due to differences in expression or stability of wild-type C(3)G compared to the C(3)GCdel protein. As a positive control to ensure that the transgenes encoding GAL4::VP16 and C(3)GCdel were both present and functioning in the experiment, and that the anti-C(3)G staining was successful, ovaries from sibling c(3)G68 conaf04903 heterozygote females were stained and analyzed at the same time. This control, in which both c(3)G and cona were heterozygous over wild-type alleles, revealed PC formation indicative of C(3)GCdel expression, as well as thread-like C(3)G staining expected for a c(3)G cona double heterozygote (Figure S2C and S2E).
The failure to detect PC formation in cona homozygotes demonstrates that CONA is required for C(3)GCdel PC formation, similar to the requirement of CONA for SC formation. This observation and the localization of CONA to C(3)GCdel PCs support the hypothesis that CONA is involved in CR formation in SCs. In these experiments we observed the disorganization of chromosomal cores/AEs along chromosome arms when the CR is abrogated by mutations in c(3)G and/or cona. Chromosomal cores/AEs detected using anti-SMC1 antibodies in wild-type appeared long and thread-like, closely matching C(3)G localization (Figure 5E). In the absence of wild-type C(3)G or CONA, however, SMC1 was detected in less intensely stained linear segments that were more numerous (Figure 5F–G). As noted above, this suggests that assembly of chromosome cores/AEs occurs along the sister chromatids but disruption of the CR of the SC results in disorganization of these structures compared to wild-type.
The SC is known to play a role in homologous chromosome pairing in Drosophila oocytes [19],[20], and defects in this process could contribute to the disorganization of AEs and the reduction in exchange in cona mutants. To determine whether cona is required for homologous chromosome pairing, we examined the association of homologous euchromatic DNA sequences in pro-oocytes and oocytes from germarium regions 2A, 2B and 3 using fluorescence in situ hybridization (FISH). Using a FISH probe that hybridizes near the middle of the X chromosome euchromatin, we found paired homologs in 97.7% (85/87) of the wild-type cells examined (Figure 6A). In contrast, paired homologs were detected in only 46.0% (40/87) of conaf04903 homozygous pro-oocytes and oocytes (Figure 6B). This demonstrates a dramatic decrease in the ability of homologous chromosomes to associate in the absence of CONA.
Testing for homolog pairing in females homozygous for c(3)G68 demonstrated that homologs were paired in only 36.5% (19/52) of cells examined (Figure 6C), which is consistent with previously published results that show a role for C(3)G, and thus the SC, in homolog pairing [19],[20]. In c(3)G68 conaf04903 double mutant females, homologs were paired in 29.8% (14/47) of the pro-oocytes and oocytes examined (Figure 6D), a figure not significantly different than that for c(3)G68 alone (χ2 = 0.506; p = 0.477). Since CONA is required for normal C(3)G localization, the pairing defect in the cona mutant may be a result of abnormal C(3)G localization. We noticed that there was a slight, but not significant (χ2 = 3.324; p = 0.068), elevation in pairing frequency in conaf04903 homozygotes compared to c(3)G68 conaf04903 double homozygotes, which could possibly be explained by a low level of homolog pairing promoted by the small amount of C(3)G that localizes to chromosomes in the conaf04903 single mutant (Figure 4). These data demonstrate that both c(3)G and cona are necessary for normal levels of homolog pairing, and are consistent with CONA functioning within the CR of the SC to promote synapsis.
We have characterized Corona (CONA), a novel SC-associated protein that is critical for the higher-order assembly of TFs into the CR of the SC. The normal localization of CONA and C(3)G is mutually-dependent – in the absence of CONA, C(3)G is visible as only spots or short threads along meiotic chromosome cores, and in the absence of C(3)G, CONA appears to be absent from the meiotic nucleus. Three lines of evidence suggest that CONA plays a critical role in the stable assembly of C(3)G into the CR of the SC. First, cona mutant oocytes fail to form long stretches of continuous SC, and only short threads or spots of C(3)G are visible within the pro-oocyte nucleus (Figure 4 and Figure S1). Second, the dependence of SC assembly (as assayed by C(3)G polymerization) on CONA::Venus expression in the absence of endogenous CONA, as well as the co-localization of CONA and C(3)G in c(2)M and ord mutants (Figure 2 and Figure 3) suggest that CONA is required to polymerize C(3)G into long stretches required to form the CR. Third, the requirement for CONA to facilitate C(3)G polymerization is also demonstrated by the fact that CONA localizes to the C(3)G PCs created by expressing C(3)G proteins that lack their C-termini and thus cannot bind chromosomes (Figure 5). Moreover, CONA also is required for the formation of these PCs, demonstrating that CONA has a functional role necessary for the connection of C(3)GCdel molecules in PCs.
The phenotypes of cona mutants make it clear that the CONA-mediated assembly of C(3)G into polymerized TFs is required for most, if not all, aspects of C(3)G function. Despite being present in cona mutants, C(3)G protein is unable to promote homolog synapsis or exchange. Defects in meiotic pairing, synapsis, and recombination are similar in cona, c(3)G and c(3)G cona mutant pro-oocytes (Figure 6, SLP and RSH, unpublished data).
In terms of its role in the formation of the CR of the SC, CONA may have a role similar to the mouse CE proteins SYCE1, SYCE2, and TEX12 [23]–[26]. These proteins co-localize extensively with the TF protein SYCP1, though SYCE2 and TEX12 were reported to have a more punctate appearance. Moreover, SYCE1 and SYCE2 also remain co-localized with SYCP1 when AEs/LEs are disrupted in Sycp3−/− spermatocytes and oocytes, and are unable to localize to meiotic chromosomes in the absence of SYCP1 [24],[25]. Mutation of SYCE2 or TEX12 disrupts synapsis and greatly reduces the amount of SYCP1 that localizes to chromosomes, yet AE proteins localize normally. In Syce2−/− and Tex12−/− meiotic cells, synapsis appears to be initiated at multiple sites along the paired homologs, but they fail to extend and form full-length SC [23],[26]. These findings are quite similar to the cona mutant phenotype, in which only a small amount of C(3)G is found on chromosomes, while the C(2)M, SMC1, and ORD proteins are still localized properly.
SYCE1 has been proposed to stabilize head-to-head interactions between SYCP1 dimers, while SYCE2 and TEX12 have been proposed to connect SYCP1-SYCE1 complexes to form higher-order structures [23],[26]. Either of these roles of CE proteins is consistent with the activities of CONA, in that the N-terminus of C(3)G is localized to the CE and required for normal SC formation [12],[21] and the formation of higher order SC or PC structures fails in the absence of CONA. Moreover, the phenotype exhibited by cona mutants parallels that documented for N-terminal deletions of C(3)G [21]; only spots or short stretches of chromosomally-associated C(3)G are visible. These data suggest that either one large or multiple small domains deleted in these N-terminus-deficient C(3)G proteins may define regions of C(3)G that interact with CONA.
Localization of C(2)M, SMC1, and ORD in cona mutant pro-oocytes indicates that chromosome core/AE structures are present, although they are more numerous and appear thinner than in wild-type. This disorganized pattern resembles that observed for C(2)M and cohesin SMC proteins when C(3)G is absent [38],[39] and argues that AEs cannot align in the absence of synapsis in Drosophila oocytes. In addition, FISH analysis demonstrates that pairing of homologous sequences is severely disrupted in cona (this study) and c(3)G oocytes [19],[20].
Disruption of homolog pairing and alignment in cona and c(3)G mutants contrasts sharply with what is observed in mammalian meiocytes lacking the TF protein SYCP1 or CE proteins SYCE2 or TEX12. Although homologous chromosomes fail to synapse in Sycp1−/−, Syce2−/−, and Tex12−/− meiotic cells, AEs lie in close proximity along their entire length [18],[23],[26]. Presumably, the formation of DSBs at multiple sites along the chromosomes establishes axial associations and these are sufficient to hold homologous chromosomes in close proximity even when the SC fails to propagate [18],[23],[26]. Axial associations likely form the basis for the assembly of the short regions of SC observed in Syce2−/− and Tex12−/− meiotic cells, which could further secure the alignment of homologs. While we cannot rule out the possibility that similar short regions of “synapsis” exist in cona oocytes, it seems likely that even a small number of these along the length of the chromosome would result in at least some examples in which AEs lie as “parallel tracks” in chromosome spreads, a phenomenon that we did not observe.
Our analysis of cona mutant oocytes suggests that, unlike mammals, the SC is critical for early events governing the pairing/alignment of homologous chromosomes in Drosophila. We can envision at least three different models that might explain why homolog alignment is dependent on SC in Drosophila. In the first model, homologous chromosomes enter meiotic prophase already paired and aligned as a result of the persistence of pairing established during preceding cell cycles and the rapid formation of SC is required to maintain these associations [4]. Although this model has been favored in the past, two published reports refute the argument that homologous chromosomes enter meiosis already paired and aligned. As noted by Fung and colleagues [45] as well as Csink and Henikoff [46], the pairing of homologous chromosomes in Drosophila somatic cells is disrupted during both replication and mitosis. Therefore, any pairing that exists prior to meiotic S phase would be lost and need to be re-established, most likely during meiotic prophase.
The second model posits that the different effects on homolog pairing and alignment observed in flies and mammals reflect differences in the ability of CE proteins to stabilize short stretches of SC. In contrast to flies, DSBs are required for synapsis in mice [47]–[49]. The short stretches of SC resulting from the formation of DSBs and early recombination intermediates in mouse meiocytes lacking SYCE2 or TEX12 may maintain the alignment of AEs in the absence of full synapsis. If the requirement for CE function is sufficiently more stringent in flies than in mammals, then short regions of synapsis similar to those observed in Tex12−/− and Syce2−/− meiocytes may be unstable or never form in cona mutant flies. In the absence of such stretches of SC or DSB-induced axial associations, the Drosophila homologs would be expected to quickly dissociate.
Our third model is based on the different temporal relationship between DSB formation and SC assembly in flies and mammals. In mammals, DSB formation and the formation of early recombination intermediates occur commensurate with, and are required for SC formation [47],[48]. In contrast, DSB initiation occurs after the completion of SC assembly in Drosophila and is not required for synapsis [49]–[51]. Because SC assembly in flies occurs via a DSB-independent pathway, pairing/alignment of AEs may be abolished in mutant oocytes in which higher-order assembly of TFs is prevented. According to this model, initial pre-synaptic associations of homologs may be maintained either by the formation of early recombination intermediates and axial associations that lead to the initiation of short stretches of SC (the mammalian paradigm), or by the establishment of extensive synapsis (the Drosophila paradigm). In both cases, the initial event (formation of recombination intermediates or SC formation) is eventually followed and perhaps ‘locked-in’ by the other. One could hypothesize that mammalian CE mutants can maintain alignments because of the earlier formation of recombination intermediates and axial associations. In contrast, lack of SC assembly in cona and c(3)G mutants would compromise the essential early step that maintains the alignment of homologous chromosomes in Drosophila oocytes. If the homologs are already apart by the time DSBs occur in cona and c(3)G mutants, DSBs would be too late to stabilize homolog associations and maintain AE alignment.
In summary, our data demonstrate an essential requirement for CONA in the polymerization of C(3)G that is required for SC formation. Understanding the mechanism by which CONA performs that role will require the identification of CONA-interacting proteins, which we expect will include the N-terminal globular domain of C(3)G and perhaps other CE proteins as well. Elucidating the function of these proteins in SC assembly and the consequences of their loss by mutation may also help us understand the role of the SC in establishing or maintaining the pairing and alignment of homologs in early prophase.
Drosophila stocks and crosses were maintained on a standard medium at 25°C. Descriptions of genetic markers and chromosomes can be found at http://www.flybase.org/ [52]. A w1118 stock was used as a wild-type stock for the immunofluorescence and FISH experiments, except for the experiment shown in Figure 1A, in which a Canton-S strain was used. Df(3R)JDP was constructed by FLP-mediated recombination essentially as described by Parks et al. [53] using FRT sequences in PBac{WH}conaf04903 and P{XP}d01968, inserted at coordinates 14,211,754 and 14,222,824, respectively, on the chromosome 3R genome map (Release 5.6). The entire cona protein-coding region is deleted in Df(3R)JDP.
The transgene construct P{UASP-cona::Venus} was constructed using the plasmid pPWV (obtained from the Drosophila Genomics Resource Center, Bloomington, IN) and the Gateway system (Invitrogen, Carlsbad, CA) using methods as recommended by the manufacturer. pPWV is identical to pUASP except that it contains a Gateway cassette followed by the Venus yellow fluorescent protein coding region [54]. The cona open reading frame was amplified from the cona cDNA bs15d10 (obtained from Geneservice, Ltd., Cambridge, UK) using primers tailed with attB1 and attB2 sequences and inserted into the vector pDONR221 in a BP Clonase (Invitrogen) reaction to form pDONR-cona. The cona cDNA insert from pDONR-cona was then transferred into pPWV in an LR Clonase (Invitrogen) reaction to form pP{UASP-cona::Venus}, with an open reading frame encoding a CONA::Venus fusion protein. After confirming the construct by sequence analysis, it was introduced into Drosophila by standard germline transformation methods (Genetic Services, Inc., Cambridge, MA).
To observe GFP-ORD in chromosome spread experiments, P{gc(2)M-myc}II.5 P{GFP::ORD}48I ord10 bw sp If/+; conaf04903 es ca /FRT82B conaA12 females were obtained by crossing y w/y+Y; (P{gc(2)M-myc}II.5 P{GFP::ORD}48I ord10 bw sp If; conaf04903 es ca)/T(2;3)CyO-TM3, P{GAL4-Hsp70.PB}TR1, P{UAS-GFP.Y}TR1: P{GAL4-Hsp70.PB}TR2, P{UAS-GFP.Y}TR2, Ser1 males to yd2 w1118 P{ey-FLP.N}2 P{GMR-lacZ.C(38.1)}TPN1/Y; FRT82B conaA12/TM6B, P{y+}TPN1, Tb1 females. For chromosome spread experiments to observe C(2)M, homozygous conaA12 females were selected from the stock yd2 w1118 P{ey-FLP.N}2 P{GMR-lacZ.C(38.1)}TPN1/Y; FRT82B conaA12/TM6B, P{y+}TPN1, Tb1.
The full-length cona open reading frame was amplified from the cona cDNA bs15d10 and cloned into pET-19b (Novagen, San Diego, CA). After the construct was verified by sequencing, the 6XHis-tagged CONA protein was expressed in E. coli BL21 cells. The bacterial expressed protein was purified using ProBond Nickel-Chelating Resin (Invitrogen). Polyclonal antibody production in guinea pigs using purified 6XHis-CONA as antigen was performed by Cocalico Biologicals (Reamstown, PA). Pre-immune sera from the immunized guinea pigs did not stain Drosophila ovaries (data not shown).
The anti-CONA antibody was specific to the CONA protein, as anti-CONA signals were not detected in ovaries from conaf04903 females (Figure S1B). Similar observations were made using ovaries from conaA12/Df(3R)JDP females [SLP and WDW, unpublished data]. These observations suggested that little or no endogenous CONA protein is produced in the presence of the conaA12 or conaf04903 mutations.
Immunofluorescence on whole ovarioles was performed as described previously and the ovarioles were mounted on coverslips by embedding in polyacrylamide gel in most experiments [8]. Primary antibodies used for staining whole-mount preparations were guinea pig anti-CONA (1∶125), mouse monoclonal anti-C(3)G 1A8-1G2 [12] (1∶500), mouse monoclonal anti-C(3)G 1G5-2F7 and 5G4-1F1 [12],[21] (1∶500 each), mouse monoclonal anti-ORB 6H4 and 4H8 [55] (1∶50 each), and rat anti-SMC1 [56] (1∶500). Secondary antibodies were Alexa 546 anti-mouse IgG (1∶500), Alexa 488 anti-mouse IgG (1∶500), Alexa 488 anti-guinea pig IgG (1∶500), Alexa 488 anti-rat IgG (1∶500) (Invitrogen), and Cy3 anti-mouse IgG (1∶500) (Jackson Immunoresearch, West Grove, PA).
Microscopy was conducted using a DeltaVision RT restoration microscopy system (Applied Precision, Issaquah, WA) equipped with an Olympus IX70 inverted microscope and CoolSnap CCD camera. Image data were corrected and deconvolved using softWoRx v.2.5 software (Applied Precision). For some experiments, confocal images were collected using a Bio-Rad Radiance 2000 laser scanning confocal microscope and Zeiss LaserSharp2000 software. Maximum intensity projections were produced from confocal data using Zeiss LSM Image Browser v.4.2 software.
Chromosome spread experiments were performed as described previously [38]. Primary antibodies used for immunofluorescence on chromosome spreads were affinity-purified guinea pig anti-SMC1 [38] (1∶500), rabbit anti-C(2)M [39] (1∶500), rabbit anti-GFP (Invitrogen) (1∶500), and mouse monoclonal anti-C(3)G 1A8-1G2 [12] (1∶500). Secondary antibodies were Alexa 488 anti-rabbit IgG (1∶400), Alexa 488 anti-mouse IgG (1∶400) (Invitrogen), Cy3 anti-guinea pig IgG (1∶400), Cy5 anti-guinea pig IgG (1∶400), and Cy5 anti-mouse IgG (1∶400) (Jackson Immunoresearch).
For chromosome spreads, images were captured and processed as described previously [38]. Because the signal intensity varies considerably for different nuclei on the same slide, wild-type and mutant images were enhanced to different degrees during processing to render details visible. In general, the C(3)G signal on chromatin in cona nuclei is significantly weaker than in wild-type.
FISH on ovarioles was performed as described elsewhere [57] with simultaneous immunofluorescence detection of ORB protein. The probe for the FISH experiments was composed of three overlapping bacterial artificial chromosome (BAC) clones from the RP98 library [58] obtained from the BACPAC Resource Center, Children's Hospital Oakland Research Institute. The three BACs (and map locations on the X chromosome) were RP98-26N1 (9F4-10A2), RP98-17B23 (9F11-10A4), and RP98-26J12 (10A4-B1). BAC DNA was isolated using the Qiagen Midi Prep Kit. A DNA mixture containing 3.3 µg of DNA from each of the three BACs was labeled with Alexa 488 (Invitrogen) essentially as described by Dernburg [59] and purified using a Qiaquick column (Qiagen). Immunofluorescence with anti-ORB primary antibodies and Cy3 anti-mouse IgG secondary antibodies was performed after hybridization under the same conditions as described above for whole mount ovarioles. The ovarioles were mounted in Prolong Gold antifade mountant (Invitrogen) [60].
Images were collected using a DeltaVision RT restoration microscopy system as described above. After image collection and processing, hybridization foci within pro-oocyte nuclei were scored for chromosome pairing. In nuclei with two foci, the distance between the pixels of highest fluorescence intensity within each focus was measured in three-dimensional image stacks using softWoRx Explorer software (Applied Precision). Nuclei containing a single hybridization focus or foci separated by 0.7 µm or less were defined as paired [19], while those with foci separated by more than 0.7 µm were defined as unpaired.
|
10.1371/journal.pmed.1002092 | Mothers after Gestational Diabetes in Australia (MAGDA): A Randomised Controlled Trial of a Postnatal Diabetes Prevention Program | Gestational diabetes mellitus (GDM) is an increasingly prevalent risk factor for type 2 diabetes. We evaluated the effectiveness of a group-based lifestyle modification program in mothers with prior GDM within their first postnatal year.
In this study, 573 women were randomised to either the intervention (n = 284) or usual care (n = 289). At baseline, 10% had impaired glucose tolerance and 2% impaired fasting glucose. The diabetes prevention intervention comprised one individual session, five group sessions, and two telephone sessions. Primary outcomes were changes in diabetes risk factors (weight, waist circumference, and fasting blood glucose), and secondary outcomes included achievement of lifestyle modification goals and changes in depression score and cardiovascular disease risk factors. The mean changes (intention-to-treat [ITT] analysis) over 12 mo were as follows: −0.23 kg body weight in intervention group (95% CI −0.89, 0.43) compared with +0.72 kg in usual care group (95% CI 0.09, 1.35) (change difference −0.95 kg, 95% CI −1.87, −0.04; group by treatment interaction p = 0.04); −2.24 cm waist measurement in intervention group (95% CI −3.01, −1.42) compared with −1.74 cm in usual care group (95% CI −2.52, −0.96) (change difference −0.50 cm, 95% CI −1.63, 0.63; group by treatment interaction p = 0.389); and +0.18 mmol/l fasting blood glucose in intervention group (95% CI 0.11, 0.24) compared with +0.22 mmol/l in usual care group (95% CI 0.16, 0.29) (change difference −0.05 mmol/l, 95% CI −0.14, 0.05; group by treatment interaction p = 0.331). Only 10% of women attended all sessions, 53% attended one individual and at least one group session, and 34% attended no sessions. Loss to follow-up was 27% and 21% for the intervention and control groups, respectively, primarily due to subsequent pregnancies. Study limitations include low exposure to the full intervention and glucose metabolism profiles being near normal at baseline.
Although a 1-kg weight difference has the potential to be significant for reducing diabetes risk, the level of engagement during the first postnatal year was low. Further research is needed to improve engagement, including participant involvement in study design; it is potentially more effective to implement annual diabetes screening until women develop prediabetes before offering an intervention.
Australian New Zealand Clinical Trials Registry ACTRN12610000338066
| Women who have had gestational diabetes are much more likely to develop type 2 diabetes.
Although many diabetes prevention programs for people over the age of 50 exist, few are tailored to the needs of young mothers who have had gestational diabetes.
On the assumption that offering prevention earlier is beneficial, researchers developed and tested a diabetes prevention program for women who had gestational diabetes; women participated in the program during their first year after giving birth.
The researchers enrolled 573 women in a one-year study: 284 women were assigned to the diabetes prevention program (one individual session and five group sessions over a three-month period, followed by telephone calls at six and nine months), and 289 were assigned to the control group (usual postnatal care).
After one year, the average changes for women in the diabetes prevention program were a 0.23-kg decrease in weight, a 2.24-cm decrease in waist circumference, and a 0.18-mmol/l increase in fasting blood glucose, while the average changes for women in the control group were a 0.72-kg increase in weight, a 1.74-cm decrease in waist circumference, and a 0.22-mmol/l increase in fasting blood glucose. The between-group difference in weight change was 0.95 kg.
The number of women who attended the diabetes prevention program was lower than anticipated—10% attended all sessions, and 53% attended the individual session plus at least one group session; about a quarter of women did not complete the study, mainly due to becoming pregnant again.
These findings suggest that although a diabetes prevention program designed for women who have had gestational diabetes can prevent weight gain over 12 months, getting women to engage with the program was challenging, so it would not be sustainable in routine health services.
The women who participated in the study had low diabetes risk profiles (only one in ten had impaired glucose tolerance), and most diabetes prevention guidelines would not categorise them as being at sufficiently high risk for participation in a diabetes prevention program.
For diabetes prevention programs in women who have had gestational diabetes, further research is required on the process of engagement and lifestyle interventions at other time points, including participant involvement in the design of interventions. Australian clinical guidelines stipulate that women who have had gestational diabetes should be screened annually for diabetes. One option for management would be to wait until they develop prediabetes before offering a diabetes prevention program, which may prove more effective because their children will be older and women may be easier to engage in improving their health.
| Gestational diabetes mellitus (GDM) and type 2 diabetes mellitus (T2DM) rates are rising worldwide [1], posing an increasing burden on the health and economic welfare of nations [2]. Women with GDM are seven times more likely to develop T2DM than women who have normoglycemic pregnancies [3]. Diabetes prevention is possible; two landmark studies in high risk individuals from the general population showed that T2DM incidence could be reduced by 58% with a combination of weight loss and moderate physical activity [4,5]. The lifestyle modification program for the original US Diabetes Prevention Program (US-DPP) was implemented over a 24-wk intensive intervention period with 16 individual coaching sessions, and a maintenance period with individual sessions every 2 mo for 24 mo [6]. Positive prevention outcomes (50% reduction in risk) were found for women with previous GDM within a subgroup analysis of the US-DPP [7].
Given that women with prior GDM are at high risk of developing T2DM and cardiovascular disease earlier in their lifespan than women with normoglycemic pregnancies [8], intervening early with a suitable diabetes prevention program (DPP) has the potential to yield positive health outcomes. Interpregnancy weight gain contributes to an increased risk of adverse pregnancy outcomes for both mother and baby during subsequent pregnancies [9,10], and the children of women with GDM are at increased risk of obesity and T2DM later in life [8]. Reducing postpartum weight retention decreases perinatal complications [10] and T2DM risk [11,12] and can influence the health status of a woman’s children [13,14], but designing interventions for this life stage is challenging [15,16]. It is well recognised that many barriers exist to mothers engaging in behaviour change during the early infancy period, including tiredness, lack of time, competing work and carer duties, and cultural expectations [17–19]. It is unclear the extent to which these obstacles can be overcome with a lifestyle modification program specifically designed to meet the needs of this population, as trials thus far have reported inconsistent results [7,20–26].
The Mothers After Gestational Diabetes in Australia Diabetes Prevention Program (MAGDA-DPP) study was undertaken to test the effectiveness of a group-based lifestyle modification program offered in the first postnatal year to women with previous GDM. At the time of MAGDA-DPP design, evidence on the optimal intervention was gathered [27], and Greater Green Triangle Diabetes Prevention Program (GGT-DPP) [28] content was adapted to meet known barriers and characteristics of the participant life stage. The program aimed to promote changes, relative to usual care, in anthropometric, behavioural, and biomedical outcomes [29,30].
MAGDA-DPP was a multicentre, prospective, open randomised controlled trial (RCT) to assess the effectiveness of a structured DPP for women with previous GDM. The detailed methods and research design of MAGDA-DPP are described elsewhere [29,30]. The trial recruited women from two Australian state capital cities, Melbourne and Adelaide. The study was approved by the relevant ethics committees and registered prospectively as an RCT. The co-primary outcomes for MAGDA-DPP were change in fasting blood glucose, waist circumference, and weight at 12 mo.
Women aged ≥18 y with a diagnosis of GDM in their most recent pregnancy were eligible for inclusion. GDM was defined by Australasian Diabetes in Pregnancy Society (ADIPS) criteria [31] at the time of study commencement: fasting plasma glucose (FPG) of 5.5 mmol/l or higher, or 2-h glucose of 8.0 mmol/l or higher on a 75-g oral glucose tolerance test (OGTT), or a glucose challenge test result of 11.1 mmol/l or higher. Exclusion criteria were the following: preexisting diabetes (type 1 diabetes mellitus or T2DM); cancer (not in remission); severe mental illness; substance abuse (illicit drugs); myocardial infarction in the preceding 3 mo; difficulty with English; involvement in another postnatal intervention trial; and pregnancy at postnatal baseline testing or at any point during the 12 mo of study involvement. Women diagnosed with T2DM or who became pregnant during the study were excluded based on the influence their condition would have on the primary outcome measures of weight, waist circumference, and FPG.
MADGA-DPP used multiple recruitment strategies, prospective and retrospective, which are described in full within our methodology publications [29,30]. Briefly, prospective recruitment involved approaching women in the antenatal clinic soon after the diagnosis of GDM, at around 28 wk of pregnancy, and conducting eligibility screening. Eligible women were provided with a patient information and consent form to return via prepaid envelope within 4 wk. If consent forms were not received within that time frame, follow-up contact was made with the woman. Once consent was received, the woman was contacted at 3 mo postpartum to initiate baseline testing.
The National Diabetes Services Scheme (NDSS) is an Australian Government initiative to subsidise blood glucose monitoring supplies for people with diabetes and is in effect a national diabetes registry. Women who have had GDM are recorded in a subregistry of the NDSS called the National Gestational Diabetes Register (NGDR). Retrospective recruitment occurred using the following approaches: a mail-out via the NDSS (using data from the NGDR) to women living in relevant postcodes in Adelaide (South Australia) and Melbourne (Victoria); referrals from one private obstetrician (South Australia); and hospital records database mining (South Australia). Women who had GDM diagnosed during their most recent pregnancy were approached and screened for eligibility using the predefined inclusion and exclusion criteria. Written informed consent was obtained from all participants, regardless of recruitment method, once screening confirmed their eligibility. The MAGDA-DPP commenced recruitment 17 January 2011, and the last participant completed final testing 28 May 2015.
Following signed consent, a trained research nurse conducted the study assessments in the participant’s home. All participants completed a baseline assessment, and the assessment was repeated after 12 mo. Standardised protocols for all clinical measures (blood pressure, anthropometry, blood sampling) were implemented [32]. Blood samples were collected by the study nurse/phlebotomist and analysed by Melbourne Pathology (Victoria) or Clinpath Laboratories (South Australia). The study nurse conducted the anthropometric (height, weight, waist circumference) and blood pressure measurements. Women provided fasting venous samples for lipid (triglycerides, total cholesterol, low-density lipoprotein cholesterol [LDL-C] and high-density lipoprotein cholesterol [HDL-C]), HbA1c, and glucose (fasting and 2-h OGTT) analysis. In addition to the baseline and 12-mo assessments, the intervention group repeated all blood tests (except OGTT) and weight and waist measures at 3 mo after baseline testing or as soon as possible after their final MAGDA-DPP group session. Survey data were completed at baseline and 12 mo and included the questions about the following: demographics (breastfeeding, parity, education, employment status, cultural background; baseline only); diet (Food Frequency Questionnaire [33]); physical activity (Active Australia Questionnaire [34]); self-regulation and self-efficacy for diet and physical activity (not reported); social support (Multidimensional Scale of Perceived Social Support; not reported); quality of life (Assessment of Quality of Life 8D; not reported); depressive symptoms and suicidal ideation (Patient Health Questionnaire 9 [PHQ-9]); and health status (including smoking status and history of diabetes, myocardial infarction, cancer, and mental disorders). Health status information on history of diabetes, myocardial infarction, cancer, and mental disorders was collected at baseline for checking of exclusion criteria. All women completing clinical testing were provided with a standardised feedback letter on their test results, and a copy was sent to their nominated family physician.
The MAGDA-DPP adopted the lifestyle modification goals of the Finnish Diabetes Prevention Study (FIN-DPS) [35] for the intervention program content and messaging. The MAGDA-DPP had five lifestyle modification goals: ≤30% of energy from fat, ≤10% of energy from saturated fat, ≥15 g dietary fibre per 1,000 kcal, ≥30 min of moderate physical activity daily, and ≥5% body weight reduction. The first three goals were calculated using Food Frequency Questionnaire data [33], the fourth using Active Australia Questionnaire data [34], and the fifth using actual body weight. In the FIN-DPS, the number of lifestyle modification goals achieved was inversely associated with diabetes incidence over a 13-y period [5].
The trial was registered with the Australian New Zealand Clinical Trials Registry on 28 April 2010, and the first participant was randomised on 1 August 2011. Once baseline diabetes screening (OGTT) results were known, eligible women were randomly allocated into either the intervention or control arm using the MAGDA-DPP management database. Permuted block randomisation was stratified by recruitment location and method. A sequence number was displayed, and the assignment code (usual care or intervention) revealed to the user in the randomisation office at Deakin University.
After randomisation, the active intervention consisted of one individual and five group sessions delivered by specially trained healthcare professionals, with two additional follow-up maintenance telephone calls for each participant, as shown in S1 Fig. As previously described [29,30], the lifestyle intervention was informed by a theoretical framework based on the Health Action Process Approach and supported by social cognitive and self-regulation theory [36]. The intervention was based on the GGT-DPP, which was previously shown to be effective in producing change in diabetes risk factors [28]. The MAGDA-DPP was tailored to reflect relevant barriers for mothers of young children (for example, topics covered the impact of sleep deprivation, stress management and mindful eating, healthy eating for families, weaning, culturally appropriate and cost- and time-saving food preparation, and exercise considerations when caring for young children), and mothers were able to bring their children along to group sessions.
Fidelity measures were incorporated throughout the intervention (facilitator manual, detailed training program, and audio recording of all facilitator sessions). The first DPP session (the individual session) was delivered in the woman’s home by the facilitator, and the MAGDA-DPP handbook was provided. The session focus was on the intention formation component of the Health Action Process Approach, personalisation of T2DM risk using a risk algorithm, and individual goal setting. This was followed by five group sessions held at 2-wk intervals and two subsequent individual phone calls at 3 and 6 mo after the final group session. Each group session was approximately 2 h in duration, with up to 15 women per group. Session content details are reported in the protocol paper [30]. Women were encouraged to set and review at least one personal goal relating to diet and one relating to physical activity at each program session (Box 1). Women in the control group received usual care and were offered the intervention program after their 12-mo final assessment.
The penetration, implementation, participation, and effectiveness (PIPE) framework for evaluating real-world program and product design elements important to implementation is a metric to evaluate the net impact of health improvement programs [37]. Four elements make up the PIPE metrics: penetration of the program into the population of interest; implementation of the proposed set of services; participation in the program; and effectiveness in generating expected outcomes. Penetration is defined as the number of individuals reached/invited divided by the number of individuals in the target population. According to Aziz et al. [37], penetration of 33% or lower is considered low, 34%–66% as moderate, 67% or higher as high. Program implementation is rated on three aspects: frequency of contact, duration of the intervention, and fidelity measures. Frequency of contact is defined based on the number, length, and type of contact within the first 12 mo of a program. A group or individual contact counts as one session, an online/telephone contact counts as 0.5 of a session, and a text/email/fax contact counts as 0.25 of a session. The total number of sessions is divided by the number of sessions delivered within the US-DPP (22 sessions) to calculate frequency (≤33% low, 34%–66% moderate, ≥67% high). Interventions lasting less than 6 mo are defined as low duration, 6–12 mo as moderate duration, and more than 12 mo as high duration. Fidelity is rated as follows: no standard curriculum as low, standard curriculum but no quality assurance measures reported as moderate, and a standard curriculum and quality assurance measures reported as high. Participation is the number of individuals enrolled in the intervention divided by the number of individuals reached/invited (≤33% low, 34%–66% moderate, ≥67% high). For DPPs, effectiveness is rated on three criteria: outcome success (number of participants achieving the main outcome divided by total number of participants completing intervention, where ≤25% low, 26%–40% moderate, >40% high success), average weight loss (≤2.3 kg low, 2.4–4.6 kg moderate, >4.6 kg high), and absolute/relative risk reduction (≤15% low, 16%–30% moderate, >30% high).
Analyses of primary and secondary endpoints were performed using SPSS version 22 and independently verified in GenStat release 16.1. Participants’ baseline characteristics are presented as summary measures. A statistical analysis plan was prepared, finalised, and signed off by the project guarantor prior to statistician unblinding. Analyses were carried out for all participants randomised to the study (ITT set, n = 573) and for the per protocol set (PPS, n = 331). The PPS analysis was confined to the subset of ITT participants excluding those with major protocol deviations such as allocation to the intervention but no exposure to any intervention sessions or ineligibility. Protocol deviations were determined independently of, and prior to, the unblinding of the trial statistician. PPS exclusions included post-baseline assessments beyond the specified time window (n = 2), pregnancy (n = 75), randomised to control group but received the intervention (n = 1), participation in another postnatal intervention during the trial (n = 3), should not have been randomised to trial (T2DM at baseline, n = 1), diagnosed with T2DM during trial (n = 11), lost contact or moved away or overseas (n = 48), and withdrew (n = 19). Similar proportions of women in the usual care and intervention arms—14% (40/289) and 12% (35/284), respectively—became pregnant during the trial. Also excluded from the PPS were women who did not receive minimum exposure to the intervention (n = 78). Minimum exposure was defined in the statistical analysis plan as attending the individual session and at least one group session.
Mixed model analyses of continuous scale endpoints used the residual maximum likelihood (REML) method to cope with missing values. The significance of the F-test for the group by time interaction is reported, as well as t-tests for within-group changes over time and between-group differences at each time point. The proportion of participants in each group known to have achieved each of the first four lifestyle modification goals at baseline and 12 mo was calculated, and the method of generalised estimating equations was used to fit models to enable group by time interactions to be tested (Wald chi-squared test). Lifestyle modification goal 5 (≥5% body weight reduction at 12 mo) was assessed using a two-sample binomial test to compare the proportions in each group. The number of goals achieved (0 to 5) by individuals at 12 mo was assessed using a mixed model analysis. Sensitivity analyses, in which missing assessments were deemed to indicate unmet goals, were conducted for each goal and also for the combined score. Unless otherwise stated, all statistical tests were conducted at the 5% significance level, with no adjustments for multiplicity of either endpoints or comparisons.
The required sample size, using a two-sided 5% significance level and 80% power, was 574 (287 in each arm); this was based on the co-primary endpoint with the smallest conjectured effect size, namely, the change in FPG over 12 mo in the GGT-DPP study [28], and thus the study was powered to detect an effect size of ≥0.27 mmol/l (assuming a mean difference between the intervention and control groups of 0.14 mmol/l and a within-group standard deviation of 0.5 mmol/l). The sample size was increased to allow for an attrition rate of up to 25%.
We approached 8,031 women, either face-to-face or via mailed out invitation, and of those, 2,211 (28%) were screened for eligibility. Of these, 828 women (38%) consented to participate in the trial, and 573 (69%) were randomised. It took 41 mo to recruit and randomise the 573 participants. The trial flow for MAGDA-DPP is shown in Fig 1. While 28% and 38% of participants were overweight or obese, respectively, the level of impaired glucose metabolism was low in the cohort (n = 58 [10%] with impaired fasting glucose; n = 10 [2%] with impaired glucose tolerance [IGT]). The intervention and usual care groups were comparable in their baseline characteristics (Table 1). The mean age of participants’ infants at baseline was 8.0 mo (standard deviation 4.8). The number of women excluded from the PPS analysis was different between the groups (n = 164 [58%] intervention participants excluded; n = 78 [27%] control participants excluded), and this difference remained significant when exclusion for not meeting minimum program exposure was removed (n = 139 [49%] intervention participants excluded; n = 78 [27%] control participants excluded). Retention rates for the intervention and usual care groups were 73% and 79%, respectively. When pregnancy was removed from loss to follow-up data, the retention rates were 93% for the usual care group and 85% for the intervention group. A single adverse event was recorded within the study (needle stick trauma), but this occurred during screening and prior to randomisation.
The intervention group’s mean weight change was −0.23 kg (95% CI −0.89, 0.43) compared with +0.72 kg (95% CI 0.09, 1.35) in the usual care group (change difference −0.95 kg, 95% CI −1.87, −0.04; group by treatment interaction p = 0.04) over 12 mo. The intervention group’s mean change in waist circumference was −2.24 cm (95% CI −3.01, −1.42) compared with −1.74 cm (95% CI −2.52, −0.96) in the usual care group (change difference −0.50 cm, 95% CI −1.63, 0.63; group by treatment interaction p = 0.389) over 12 mo. The intervention group’s mean increase in fasting blood glucose was 0.18 mmol/l (95% CI 0.11, 0.24) compared with an increase of 0.22 mmol/l (95% CI 0.16, 0.29) in the usual care group (change difference −0.05 mmol/l, 95% CI −0.14, 0.05; group by treatment interaction p = 0.331) over 12 mo. Tables 2 and 3 show the results for the ITT analysis set for the 12-mo data; no other statistically significant results were identified across the primary and secondary endpoints when using the F-test of the group by time interaction—a result that was consistent for both the ITT and PPS analyses.
Compared with baseline levels, the between-time comparisons at 3 mo show that mean weight change in the intervention group was −0.92 kg (p = 0.001) (Table 4). Other significant results in the intervention arm at 3 mo were a reduction in waist circumference, total cholesterol, HDL-C, and LDL-C (all p < 0.001). FPG was significantly higher at 3 mo than at baseline (p < 0.001) (Table 4). Reductions in waist circumference, total cholesterol, HDL-C, and LDL-C were maintained at 12 mo but not the reduction in weight. The increase in FPG persisted at 12 mo.
Analysis of the proportion of participants meeting the MAGDA-DPP lifestyle modification goals adopted from the FIN-DPS did not reveal any significant time by group interactions. Similarly, there was no significant difference in the total number of goals achieved between the two groups (Table 5). S2 Fig. displays the association between average weight loss and different levels of engagement within the active intervention period (first 3 mo) and at 12 mo.
The MAGDA-DPP intervention was delivered via an RCT and did not have a specific target population for which penetration could be exactly calculated due to different recruitment streams. Some idea of penetration can be estimated from the NDSS mail-out invitation to participate, which was sent to 5,349 women registered with the NGDR and living in the study’s geographical catchment areas. Only 191 women responded to the NGDR invitation, and 149 of those subsequently agreed to participate in the intervention, resulting in a penetration rate of 4%, which is low. The implementation metric for MAGDA-DPP had a low level for frequency of sessions (32% of the number delivered within the US-DPP), a moderate level for duration (6–12 mo), and a high level for theoretical fidelity (standard curriculum and quality assurance measures implemented). The participation metric, based on enrolment of invited individuals, was low (26%) and reflects the challenge of engaging women in the intervention. The measures for PIPE effectiveness were all low: proportion of successful participants, average weight loss, and diabetes risk reduction (indirect, assessed against achievement of the five lifestyle modification goals adopted from the FIN-DPS, which are inversely associated with diabetes incidence [5]).
Among those randomised to the intervention (n = 284), 66% (n = 188) completed at least the individual session; specifically, 53% met the program minimum exposure definition of completing the individual session and one or more group sessions (n = 149), 34% had no exposure to the intervention (n = 96), 13% completed only the individual session (n = 37), and only 10% completed the individual session and all five group sessions (n = 28). Pregnancy rates and subsequent ineligibility for those completing the minimum intervention (11%, 16/149) and non-completers (14%, 16/135) were similar (p = 0.852). Group facilitators spent an average of 18 min per participant arranging intervention sessions and reminding participants. Despite an average of four attempts at contact made by facilitators via phone call or text message, 31% of women failed to attend a single session. To achieve the minimum intervention exposure (1 individual session and ≥1 group session), facilitators made on average ten contacts (mean total duration 20 min). Group facilitators made on average three contacts (mean total duration 8 min) to ensure attendance at a single session.
This study of a postnatal lifestyle intervention in women with gestational diabetes achieved a 1-kg weight difference compared with the control group. This difference is potentially significant for diabetes prevention, but the participation rate was low, reflecting how difficult it was to engage women in this cohort in the first year after the birth of their child. We found that, on average, women randomised to the MAGDA-DPP intervention group showed no postnatal weight gain, in contrast to women in the usual care group, who continued to gain weight over the 12-mo study period. The changes over 12 mo in the other two primary outcomes, the diabetes risk measures fasting blood glucose and waist circumference, were not significantly different for women in the intervention versus the control group. Intervention participants did show initial significant weight loss and improvements in their waist circumference and fasting blood glucose following the intensive component of the intervention, but these benefits were for the most part lost at 12 mo. This phenomenon is common amongst lifestyle modification programs and is a well-noted challenge in diabetes prevention in general [25,38,39] and in this population in specific [7,25,40,41]. Recruiting and delivering an intervention within this population of women with young families proved challenging, and our outcomes are similar to those recently reported by other studies [22,25,41].
Obesity is one of the strongest modifiable risk factors for T2DM development [5,42], and postnatal weight gain is a key risk factor for women [16,43], especially women with previous GDM [12,44]. Australian women typically gain 650 g annually [45], and the women in the MAGDA-DPP usual care group were no different (720 g average). The US Agency for Healthcare Research and Quality recently identified a 0.5-kg between-group weight gain difference as significant [46], and similarly the US Community Preventive Services Task Force found that even low levels of weight loss are effective in reducing T2DM risk [47]. Clinical significance has been attributed to a ~1-kg weight difference over time for cardiovascular disease [48] and T2DM [43], amongst other diseases. Wang and colleagues modelled a similar weight change within the US population and estimated that 2 million diabetes cases could be avoided with this small change [49]. Given that postnatal weight retention increases diabetes risk [11] and that guidelines recommend postnatal weight management [31,50,51], our findings could be interpreted as supporting the potential for a low intensity program to address postnatal weight retention and therefore lower diabetes risk. We would argue that our findings represent an issue of low penetration and participation in this target group, resulting in low effectiveness.
The number of reported DPPs specifically designed for women with prior GDM has risen exponentially, but their effectiveness in reducing diabetes risk has been low to date. A recent meta-analysis [41] found that no significant reduction in fasting blood glucose or any significant impact on weight loss occurred in DPPs of ≤6 mo duration. When the analysis was expanded to interventions of 12 mo duration, a significant difference in weight change of −1.06 kg (95% CI −1.68, −0.44) was seen, but this was driven by a single interim results publication for a study whose full intervention results have not been published [20]. It is clear that for effective weight loss within DPPs, high session frequency and longer program duration and fidelity are needed [37]. This presents a challenge for women with young families, who commonly cite a lack of time as a major barrier to engagement [17]. Nevertheless, a lower frequency of sessions can be effective for diabetes prevention—when delivered over longer periods of time and where penetration and participation rates are higher [37,47]—which is important when looking to sustainability or scaling up a program for health service delivery.
Central to the issue of penetration and participation is the design of randomised trials, which leads to the recruitment of highly selective populations. One of the largest DPPs in women with previous GDM comes from a study by Ratner and colleagues [7]; systematic reviews consistently [41,52,53] identify this study as high-quality evidence for the role of a DPP in this population, but the generalisability of the results from the population recruited is rarely discussed. Participants (n = 350) were mothers with IGT who were on average 43 y old, obese (with a mean BMI of 34.2 kg/m2), and with 12 y since their index GDM pregnancy. Clearly, their diabetes risk was higher, their child care demands lower, and the chance of engagement greater. In contrast, MAGDA-DPP mothers were 10 y younger, with BMI averaging in the overweight range, and only 2% had IGT. It is to be expected that their diabetes risk and their risk perception were likely to be quite different from those of Ratner et al.’s participants [7]. The recently published GEM trial [25] provides us with a more real-world perspective on the comparative effectiveness at the health service level. The population of the GEM trial (n = 2,280) was similar to that of MAGDA-DPP, but the trial’s penetration and participation was high as a result of the intervention being embedded as usual care within 22 randomised medical facilities and using telephone health coaching. The trial’s 12-mo weight loss outcomes showed significantly less postpartum weight retention in the intervention participants and a −0.64 kg weight difference between the intervention and control groups (95% CI −1.13, −0.14), lending support to our findings being more in line with real-world outcomes.
There are some lessons to be learnt from the factors contributing to the low effect size seen. The relatively low intervention engagement in MAGDA-DPP is reflected in an accordingly low level of behavioural change and resulting weight change. Attending and completing weight loss interventions are known correlates to achieving weight loss [54,55]; when people leave a program early, their skills and coping strategies for achieving and sustaining weight loss are likely to be underdeveloped [56,57]. Risk perception is another important influence on engagement with lifestyle behaviour change [36]. At the individual session, a risk algorithm was used to demonstrate the risk of developing diabetes to participants. Risk algorithms are highly age-dependent; most women were normoglycemic, so it is possible their interpretation was that they did not need to worry about their risk of diabetes until they were older.
Strengths of this randomised trial include the length of follow-up after the active intervention, good retention rates, the fidelity measures included in the intervention design, and the rigorous data collection methodology. Limitations of the MAGDA-DPP study include the low level of participation in the intervention group sessions along with overall low levels of penetration and participation, as defined by the PIPE metric [37]. Although relatively extensive consultation work was undertaken prior to MAGDA-DPP implementation (literature review, qualitative interviews with the population of interest [18], piloting of the program materials in postnatal women who had gestational diabetes), it is possible that a broader qualitative exploration of issues relating to penetration, compliance, and program delivery may have yielded stronger engagement and possibly better outcomes. The diabetes risk profiles for MAGDA-DPP participants were surprisingly low considering the body of evidence behind GDM being a strong risk factor for T2DM development [3,11]. At baseline, 10% of MAGDA-DPP participants were identified as having prediabetes, and their average BMI was only 1 kg/m2 higher than the average Australian woman [58]. It is also possible that those who agreed to participate were a lower-risk group, with healthier baseline behaviours. Another possible study limitation was the study protocol specification of three co-primary endpoints without a plan to test each at a stringent significance level, or in a hierarchical manner, in order to maintain a trial-wise type I error rate below, say, the conventional 5%. Our protocol [30] did state that a “statistically significant change in any one of these three endpoints will be regarded as evidence of a change in diabetes risk”, and we found a statistically significant difference between the groups for weight change. The observed magnitude of the difference is similar to the magnitudes reported in other studies of lifestyle interventions [25,41], and we believe it is important to add the result of this study to the accumulating knowledge about the utility of lifestyle modification programs in mothers with prior GDM.
Our trial explored the effect of offering a DPP in the first year postnatally and showed that it was ineffective. Telephone- or web-based interventions that can adapt to the time demands of raising a young family may have more successful participation rates [23,25] and may have the advantage of being less resource intensive and more suited to scale-up, but it is unlikely that they will be as effective as programs offered to women with the high-risk characteristics of those in the study by Ratner et al. [7].
The extent to which the newer GDM diagnostic criteria of the International Association of Diabetes and Pregnancy Study Groups will affect demand for diabetes prevention services in not yet known [59], but our finding that the majority of our cohort were at low risk (using the previous, higher GDM diagnostic cut-offs) suggests that the relative benefit and cost associated with offering an early postnatal period DPP to all women with a previous GDM pregnancy does not make it a sensible use of scarce health resources. A better health service approach might be to improve the currently recommended annual diabetes screening within family medicine practice for women with previous GDM, so more women with prediabetes, who are at high risk, can be identified [50]. This health service approach could be supported by a reminder system within a national GDM registry, the NGDR being the current Australian example, and women with prediabetes could be more selectively targeted for recruitment into an appropriate DPP.
Our results show that a low intensity, group-delivered DPP was superior to usual care in preventing postnatal weight gain in a cohort of women with previous GDM. However, the level of engagement was low, and DPPs may need to be offered at other time points after pregnancy. Further research on engagement is required, including participant input into the design of interventions, and a more effective option may be to follow up women with previous GDM until they show IGT or HbA1c levels in the prediabetes range before offering entry to a DPP.
|
10.1371/journal.pntd.0003410 | Parasitic Nematode-Induced CD4+Foxp3+T Cells Can Ameliorate Allergic Airway Inflammation | The recruitment of CD4+CD25+Foxp3+T (Treg) cells is one of the most important mechanisms by which parasites down-regulate the immune system.
We compared the effects of Treg cells from Trichinella spiralis-infected mice and uninfected mice on experimental allergic airway inflammation in order to understand the functions of parasite-induced Treg cells. After four weeks of T. spiralis infection, we isolated Foxp3-GFP-expressing cells from transgenic mice using a cell sorter. We injected CD4+Foxp3+ cells from T. spiralis-infected [Inf(+)Foxp3+] or uninfected [Inf(-)Foxp3+] mice into the tail veins of C57BL/6 mice before the induction of inflammation or during inflammation. Inflammation was induced by ovalbumin (OVA)-alum sensitization and OVA challenge. The concentrations of the Th2-related cytokines IL-4, IL-5, and IL-13 in the bronchial alveolar lavage fluid and the levels of OVA-specific IgE and IgG1 in the serum were lower in mice that received intravenous application of Inf(+)Foxp3+ cells [IV(inf):+(+) group] than in control mice. Some features of allergic airway inflammation were ameliorated by the intravenous application of Inf(-)Foxp3+ cells [IV(inf):+(-) group], but the effects were less distinct than those observed in the IV(inf):+(+) group. We found that Inf(+)Foxp3+ cells migrated to inflammation sites in the lung and expressed higher levels of Treg-cell homing receptors (CCR5 and CCR9) and activation markers (Klrg1, Capg, GARP, Gzmb, OX40) than did Inf(-)Foxp3+ cells.
T. spiralis infection promotes the proliferation and functional activation of Treg cells. Parasite-induced Treg cells migrate to the inflammation site and suppress immune responses more effectively than non-parasite-induced Treg cells. The adoptive transfer of Inf(+)Foxp3+ cells is an effective method for the treatment and prevention of allergic airway diseases in mice and is a promising therapeutic approach for the treatment of allergic airway diseases.
| Many studies have investigated the down-regulation of the immune system by parasite infection. CD4+CD25+Foxp3+T (Treg) cells are key players in parasite-mediated immune downregulation. Our previous study suggested that Treg cells recruited by Trichinella spiralis infection were the key cells mediating the amelioration of allergic airway inflammation in mice. In the present study, we investigated the functions of parasite-induced Treg cells using mice expressing GFP-tagged Foxp3. T. spiralis infection increased the number of Treg cells. Adoptive transfer of the parasite-induced Treg cells to mice with allergic airway inflammation ameliorated allergic airway inflammation. The transferred cells were recruited to inflammation sites in the lung. Cells from parasite-infected mice expressed higher levels of Treg-cell homing receptors and activation markers than did cells from uninfected mice. This study might help explain why immune disorders (often of unknown cause) are more prevalent among people in developed countries (areas with low parasite infection) than among those in developing countries (areas with parasite epidemics). Our finding might improve current cell therapy techniques and facilitate the development of new techniques that use parasites or parasite-borne materials to treat diverse immune disorders.
| In humans, trichinellosis, caused by oral infection with Trichinella sp., is typified by an intestinal phase and a muscular phase, corresponding to two distinct periods in the parasite's life cycle in the host [1], [2]. The physiopathological symptoms include heavy muscle aches, fever, and eosinophilia [3]. During each of the two phases, the host immune system activates different responses to the infection. Th2-related cytokine levels increase immediately after T. spiralis larvae invade the intestine [4], and the levels of IL-4 and IL-13 peak before the initiation of nurse cell formation [4], [5]. Additionally, the levels of most Th17-related cytokines increase until the muscle phase begins. Th2- and Th17-related cytokine levels decrease after the recruitment of CD4+CD25+ Forkhead box P3 (Foxp3)+T (Treg) cells to the spleen and lymph nodes [4]. Treg cells appear to play a role in the maintenance of chronic infections or in the suppression of the parasite targeting immune response [4], [6].
Treg cells contribute to the maintenance of host immune homeostasis by actively suppressing various pathological and physiological immune responses [7]. To reduce the infectious burden, parasites can influence natural Treg cells by modifying the T-cell immune response at the infection site, thus allowing the parasite to survive in the host for longer periods [8]. Although some controversy remains, two different mechanisms are thought to underlie the suppression of Treg cells during parasite infection. In the first, the interaction of the T effector ligands CD80 and CD86 with cytotoxic-T-lymphocyte-associated protein (CTLA-4) activates the transmission of immunosuppressive signals on T effector cells, thereby reducing the function of effector T-cells. In the second, cytokines such as IL-10 and transforming growth factor (TGF-β) mediate suppression [8], [9]. After some parasite infections, Treg cells activate specific genes, such as those encoding CD103, Foxp3, glucocorticoid-induced TNFR family related gene (GITR), OX40 (CD134), CTLA-4, secretory leukocyte peptidase inhibitor (Slpi), granzyme B (Gzmb), fatty acid-binding protein 5 (Fabp5), nuclear factor, interleukin 3 regulated (Nfil3), suppressor of cytokine signaling 2 (Socs2), G protein-coupled receptor 177 (Gpr177), and killer cell lectin-like receptor subfamily G, member 1 (Klrg1) [10]–[14]. However, the roles and mechanisms of Treg cell-mediated suppression remain controversial and require further investigation [15]. Although many studies have demonstrated that parasites can activate and induce the Treg-cell population, few studies have investigated the immune regulatory mechanisms of parasite-induced Treg cells after their direct transfer into animals with immune disorders. The OVA-alum allergic airway inflammation model has been widely used as an animal model of immune disorders because it enables the study of Th2-mediated allergic responses [16]–[19].
In a previous study, we observed that T. spiralis infection induced the Treg-cell population and increased IL-10 and TGF-β cytokine levels, and infection may also reduce artificially induced allergic airway inflammation [20]. In this study, to examine the functional roles of parasite-induced Treg cells, we evaluated the expression of Treg-cell surface markers (related to homing, suppression ability, and responses to inflammatory cytokines) and the functional effects of induction with T. spiralis. In addition, we intravenously injected parasite-induced CD4+Foxp3+T cells and natural CD4+Foxp3+T cells into normal mice before and during airway inflammation.
The T. spiralis strain (isolate code ISS623) was maintained in our laboratory by serial passage in rats. Carcasses of infected mice were eviscerated and cut into pieces. The parasite-infected muscles were digested in 1% pepsin-hydrochloride solution with constant stirring for 1 h at 37°C. The muscle-stage larvae were collected under a microscope after removal of the pepsin-hydrochloride solution. The larvae were rinsed more than 10 times in sterile PBS.
During the experimental period, Foxp3-eGFP mice (expressing GFP-tagged Foxp3) purchased from Jackson Laboratory were maintained in a specific pathogen-free facility at the Institute for Laboratory Animals of Pusan National University. Foxp3+ cells were isolated from the spleen of T. spiralis-infected [Inf(+)Foxp3+] and uninfected [Inf(-)Foxp3+] Foxp3-eGFP mice. The spleens were minced into small pieces, which were placed into ACK hypotonic lysis solution (Sigma, USA) at room temperature for 2 min to lyse erythrocytes (red blood cells, RBCs). Following lysis, the remaining cells were filtered through 100-µm meshes (Small Parts, Inc., USA) and washed three times. CD4+ T cells were isolated using a CD4+ T cell isolation kit (Miltenyi Biotech, USA) in accordance with the manufacturer's protocol. Foxp3+ (GFP+) cells were obtained using a FACS cell sorter.
Five-week-old female C57BL/6 mice were purchased from Samtako Co. (Korea). Four groups of mice were used. In the first group of mice, allergic airway inflammation was induced via intraperitoneal (IP) injection of ovalbumin (OVA)-alum for sensitization followed by intranasal (IN) challenge with OVA four times, without adoptive cell transfer [OVA+]. The second group of mice was injected (IP) and challenged (IN) with PBS, without adoptive cell transfer [OVA-]. The third group of mice was administered Inf(+)Foxp3+ cells (5 × 105) intravenously, and allergic airway inflammation was induced with OVA/alum injection (IP) and OVA challenges (IN) [OVA+IV(inf):+(+) group]. The fourth group of mice was administered Inf(-)Foxp3+ cells, following the same protocol used for the third group of mice [OVA+IV(inf):+(-) group] (Fig. 1A). To evaluate the efficiency of the adoptive transfer according to the injection time, we transferred cells before the allergic airway inflammation induction period (Stage I, preventive effect) or after the first allergen challenge (Stage II, therapeutic effect). Allergic airway inflammation was induced as previously reported with some modifications [20]. Briefly, mice were sensitized via IP injection of 75 µg OVA (Sigma-Aldrich, USA) and 2 mg of aluminum hydroxide (alum; Sigma-Aldrich) in 200 µL of 0.9% sterile saline on days 1, 2, 8, and 9. The mice were then challenged with IN administration of 50 µg of OVA on days 15, 16, 22, and 23. Airway hyper-responsiveness (AHR) was measured on day 24, and the mice were sacrificed on day 25 (Fig. 1B).
Lymphocytes were isolated from the spleen and lung-draining lymph nodes (LLN) of mice to determine the levels of allergen (OVA)-specific, cytokine-secreting lymphocytes. The methods were identical to those described above for the preparation of CD4+Foxp3+ T cells. The isolated cells were plated onto OVA-coated wells and non-coated wells at 5×106 cells/mL in RPMI 1640 with 10% fetal bovine serum and penicillin/streptomycin. After incubation for 72 h at 37°C in 5% CO2, culture supernatants was harvested and stored at −20°C for ELISA.
After mice were anesthetized, the tracheas were exposed and cut just below the larynx. A flexible polyurethane tube attached to a blunt 24-gauge needle, with a 0.4-mm outer diameter and length of 4 cm (Boin Medical Co., Korea), 800 µL of cold PBS was inserted into the trachea. The BALF samples were collected and centrifuged for 5 min at 1500 rpm at 4°C. After centrifugation, the supernatants were collected and quickly frozen at −70°C. The cell pellets were resuspended in 100 µL of ACK hypotonic lysis solution (Sigma) and incubated for 2 min to lyse the RBCs. Next, 900 µL of PBS was added, and the cell suspension was centrifuged for 5 min at 3000 rpm at 4°C. The supernatants were then decanted, and the cell pellets were resuspended in 100 µL of PBS. After each procedure, the cells were centrifuged onto microscope slides for 5 min at 500 rpm using a Cytospin apparatus (Micro-12TM; Hanil Co., Korea). The microscope slides were air-dried and stained with Diff-Quik (Sysmex Co., Japan). Cells on the stained slides were counted in a blinded manner under a light microscope. At least 500 cells were counted per slide.
At 24 h after the last allergen challenge, airway responsiveness was evaluated by measuring the change in lung resistance in response to aerosolized methacholine (Sigma) [21]. To measure bronchoconstriction, the enhanced pause (PenH) was measured at baseline (PBS aerosol) and after exposure to increasing doses of aerosolized methacholine (0–50 mg/mL) using whole-body plethysmography (Allmedicus, Korea). In the plethysmography procedure, the mice were acclimated for 3 min, exposed to nebulized saline for 10 min, and treated with increasing concentrations (0, 12.5, 25, and 50 mg/mL) of methacholine using an ultrasonic nebulizer (Omron, Japan). After each nebulization, the PenH values measured every three minutes during the experimental period were averaged. Graphs were generated showing the PenH values in response to increasing methacholine concentrations for each dose-matched group of mice.
Histopathological analyses were performed as described previously [22]. In brief, lung tissues were fixed in a 10% formaldehyde solution and embedded in paraffin. The tissue was then cut into sections, and the sections were stained with hematoxylin-eosin (H&E) and periodic acid-Schiff (PAS) stains. The stained sections were evaluated under a microscope [23].
Total RNA was extracted from the lung using 1 mL of BIOZOL (LPS Solution, Korea), and cDNA was synthesized using MMLV reverse transcriptase (Promega, USA) according to the protocols provided by the manufacturer. MUC2, MUC5, and eotaxin RNA levels were quantified using a real-time PCR machine (Bio-Rad Laboratories, Inc., USA) according to the manufacturer's instructions. Total RNA was extracted from sorted Foxp3+ cells. The transcript levels of chemokine (C-X-C motif) receptor 3 (CXCR3), chemokine (C-C motif) receptor (CCR) 4, CCR5, CCR9, CCR10, Klrg1, capping protein gelsolin-like (Capg), Gzmb, glycoprotein A repetitions predominant (GARP), CTLA-4, CD62L, and OX40 in the Treg cells were analyzed using real-time PCR. The primer sequences are listed in S1 Table. The relative expression of each gene was calculated as the ratio of target gene expression to housekeeping gene (GAPDH) expression using the Gene-x program (Bio-Rad laboratories, Inc.).
After mice were sacrificed, serum was collected via cardiac puncture. The serum was diluted 1∶40 (for IgG1 and IgG2a) in blocking buffer. OVA-specific IgG1, IgG2a, and IgE levels in the serum and cytokine [IL-4, IL-5, IL-10, IL-13, interferon (IFN)-γ, TGF-β] levels in the BALF and culture supernatants of LLN were measured using ELISA in accordance with the manufacturer's instructions (eBioscience, USA). The absorbance of the final reactant was measured at a wavelength of 450 nm with an ELISA plate reader.
To evaluate the recruitment of Treg cells after the adoptive transfer of Foxp3+ cells, live cells were isolated from the spleen and LLN of OVA-induced allergic airway inflammation mice that had been infected or not infected with T. spiralis. The cell preparation method was identical to those of section “lymphocyte preparation”. The samples were acquired using a FACS maschine. The following mAbs were used: anti-CD4-PE, anti-CD25-APC, anti-CD39-efluor 660, and anti-CTLA-4-PE.
Paraffin sections of lung tissue were deparaffinized and then treated with an antigen retrieval solution for 20 min (0.1 M citric acid, 0.1 M sodium citrate, pH 6.0). The slides were rinsed with PBS and immersed in methanol (0.3% H2O2) for 15 min to inhibit endogenous peroxidase activity. After pre-incubation with 1% BSA for 1 h at room temperature, the sections were incubated with hamster anti-mouse CTLA-4 (1∶500; Santa Cruz Biotechnology, USA) for 1 h at 4°C. After several washes in PBS, the Alexa Fluor 594 goat anti-hamster IgG secondary antibody (1∶500; Jackson ImmunoResearch Laboratories, USA) was applied for 1 h at room temperature. The slides were washed in PBS and incubated with DAPI for 2 min. Confocal images of stained lung tissue or stained specific Treg cells were examined under an inverted fluorescence microscope.
In 96-well round-bottomed plates, purified splenocytes were cultured in the presence of 1 µg/mL anti-CD3. The number of responder splenocytes per well was kept constant at 3×104 cells, while the number of suppressor cells was varied. Normal splenocytes were mixed at several different ratios with CD4+Foxp3+ cells isolated from parasite-infected mice or uninfected mice in 200 µL of complete medium. Cell viability analysis using the trypan blue dye exclusion assay was performed three days later. After trypsinization, the number of viable cells in triplicate wells at each concentration was estimated using a hemocytometer. The cells in each well were counted three times, and the experiment was repeated three times [24].
Data were analyzed using SPSS for Windows, version 14 (SPSS, USA). Student's t-test or ANOVA was used to compare the group means.
The study was performed with approval from the Pusan National University Animal Care and Use Committee (Approval No. PNU-2013-0293), in compliance with “The Act for the Care and Use of Laboratory Animals” of the Ministry of Food and Drug Safety, Korea. All animal procedures were conducted in a specific pathogen-free facility at the Institute for Laboratory Animals of Pusan National University.
To determine whether T. spiralis-induced Treg cells can prevent allergic airway inflammation in an OVA-alum asthma mouse model, we injected mice with CD4+Foxp3+ T cells isolated from T. spiralis-infected [Inf(+)Foxp3+] or uninfected [Inf(-)Foxp3+] mice and evaluated allergic inflammation responses after OVA sensitization and challenge (Fig. 1B, Stage I). Fig. 2A shows the changes in the immune cell populations in the BALF of asthma-induced mice after the adoptive transfer of Inf(+)Foxp3+ cells. The adoptive transfer of CD4+Foxp3+ cells from T. spiralis-infected mice, but not of those from uninfected mice, reduced the number of eosinophils in the BALF. To evaluate airway function, airway responsiveness was determined after treating the mice with increasing doses of methacholine. Methacholine increased the PenH value in the OVA-induced allergic airway inflammation group in a dose-dependent manner. The introduction of Inf(+)Foxp3+ cells decreased the PenH value in OVA-induced mice, whereas Inf(-)Foxp3+ cell treatment did not (Fig. 2B).
Following the induction of airway inflammation, an influx of inflammatory cells into the peribronchial spaces was observed. The influx of inflammatory cells led to the destruction of the alveolar wall and generated severe hemorrhage. We observed hypertrophy of goblet cells in the peribronchial epithelium and high amounts of mucus production in the OVA-induced allergic airway inflammation group via PAS staining. In the Inf(+)Foxp3+ cell transfer group, inflammatory cell infiltration in the peribronchial areas decreased somewhat, and a small amount of mucus production and decreased goblet cell hyperplasia were observed in the peribronchial epithelia, with little hypertrophy of goblet cells in the tracheal and bronchial epithelia (Fig. 2C). Gene expression related to mucus production (MUC2 and MUC5) and eosinophil chemoattractant (eotaxin) in the lung was lower in the Inf(+)Foxp3+ cell transfer group than in the OVA-challenged group (Fig. 2D).
To characterize the effects of CD4+Foxp3+ T-cell transfer on cytokine secretion in the BALF, an ELISA was performed to monitor the expression of Th1 (IFN-γ), Th2 (IL-4, IL-5, and IL- 13), and regulatory (IL-10 and TGF-β) cytokines. The concentrations of IL-4, IL-5, and IL-13 in the BALF of the Inf(+)Foxp3+ cell transfer group were lower than those in the other OVA-challenged groups (p<0.05; Fig. 3A). Inf(+)Foxp3+ cell transfer did not affect the production of IL-10 and TGF-β (Fig. 3A). In the LLN, IL-4 and IL-13 cytokine production by lymphocytes decreased in the Inf(+)Foxp3+ cell transfer group, but IL-5 production was not affected by the cell transfer. The cytokine profiles of the Inf(-)Foxp3+ cell transfer group were mostly similar to those of the Inf(+)Foxp3+ cell transfer group. Th2 cytokine production was inhibited by Inf(-)Foxp3+ cell transfer, but the effect was less than that seen in the Inf(+)Foxp3+ cell transfer group (Fig. 3B).
OVA challenge increased the levels of OVA-specific IgE and IgG1. The increase was attenuated by Inf(+)Foxp3+ cell transfer, but not by Inf(-)Foxp3+ cell transfer. The level of OVA-specific IgG2a was similar in each group (Fig. 4).
To understand the mechanism of reduced airway inflammation in CD4+Foxp3+T cell-transferred mice, we examined Treg cells in the spleen and LLN. The Treg-cell subset increased in the spleen and LLN after Inf(+)Foxp3+ cell transfer (Fig. 5A and B). In the LLN, the Treg-cell population also increased after Inf(-)Foxp3+ cell transfer (Fig. 5B). In addition, Foxp3 gene expression in the lung increased after OVA challenge. Inf(+)Foxp3+ cell transfer further increased Foxp3 gene expression than only OVA challenge group, whereas Inf(-)Foxp3+T cell transfer did not (S1A Fig.).
To determine the origin of the Treg cells in the lung, we examined transferred Foxp-eGFP cells in the lung using confocal microscopy. In addition, we evaluated the activation of Treg cells by examining CTLA-4 expression on Foxp-eGFP cells in lung tissue. Interestingly, we detected Foxp-eGFP cells in the lungs of Inf(+)Foxp3+ and Inf(-)Foxp3+ cell-transferred mice. In mice without OVA-induced inflammation, a few Foxp-eGFP cells were found in the lung matrix, but not around the airways (Fig. 5C-a & -b and S1B Fig.). However, numerous Foxp-eGFP cells were detected in the allergic airway of inflammation-induced mice, particularly in Inf(+)Foxp3+ cell-transferred mice (Fig. 5D-c and 5D-d). Almost all of the Foxp-eGFP cells in the lung strongly expressed CTLA-4, a surface marker for Treg-cell activation, and these cells were detected around the sites of airway inflammation (Fig. 5D-c, 5D-d, S1A and B Fig.).
To determine the therapeutic effects of CD4+Foxp3+ T cells on allergic airway inflammation, we introduced two types of CD4+Foxp3+ T cells at the initiation of inflammation and investigated the disease index of allergic airway inflammation, as in the Stage I experiment (Fig. 1B, Stage II). The introduction of Inf(+)Foxp3+ cells ameliorated airway inflammation: airway responsiveness values, mucin secretion in the airway, and MUC2, MUC5, and eotaxin gene expression in the lung were reduced (Fig. 6A and 6B and S2A & S2B Fig.). However, although the number of eosinophils was reduced in the BALF, the change was not significant (S2C Fig.). The Th2 cytokine concentration in the BALF of Inf(+)Foxp3+ cell-transferred mice decreased, but the concentration of TGF-β increased (Fig. 6C). IFN-γ and IL-10 concentrations were not affected by Inf(+)Foxp3+ cell transfer (S3A Fig.). Except for IL-4, Th2 and regulatory cytokine production in lymphocytes isolated from the LLN after CD4+Foxp3+ T-cell transfer did not change (S3B Fig.). The Treg-cell population increased in the LLN, but not the spleen, of Inf(+)Foxp3+ cell-transferred mice (Fig. 7A). As with the Stage I experiment, we detected a few Foxp-eGFP cells in the lungs of mice in which inflammation was not induced (Fig. 7B-a and 7B-b). In the Stage II experiment, many Foxp-eGFP cells were detected in mice with allergic airway inflammation, but in the microscopy fields examined, there appeared to be fewer CTLA-4-expressing cells (Ave. 18.45 cells/105 DAPI+ cell) than those (Ave. 28.95 cells/105 DAPI+ cell) in the Stage I experiment (Fig. 5D & 7C).
To evaluate changes in the molecular characteristics of Treg cells following T. spiralis infection, we analyzed surface markers on Treg cells isolated from the spleen. The number of CD4+CD25+Foxp3+T cells and CD4+CD25-Foxp3+T cells increased following T. spiralis infection. In addition, the Treg-cell activation markers CTLA-4 and CD39 significantly increased after parasitic infection (Fig. 8A). In addition, the expression of CCR5 and CCR9, which encode Treg-cell homing receptors, was higher than in Inf(-)Foxp3+ cells, whereas the expression of CXCR3 and CCR4 in Inf(+)Foxp3+ cells was lower (Fig. 8B). We analyzed the expression of genes related to the function and activation of Treg cells, such as Klrg1, Capg, GARP, Gzmb, and OX40. Except for Klrg1, the expression of these genes in Inf(+)Foxp3+ cells was 3- to 10-fold higher than in Inf(-)Foxp3+ cells (Fig. 8B). We compared the functional properties of Inf(+)Foxp3+ cells and Inf(-)Foxp3+ cells using naive T-cell co-culture. Both types of Treg cells efficiently inhibited T-cell proliferation. However, Inf(+)Foxp3+ cells inhibited T-cell proliferation more effectively than did Inf(-)Foxp3+ cells (50.5% vs. 56.1%, respectively) (Fig. 8C).
To maintain their long-term survival in a host organism, helminthic parasites have immune suppressive abilities that can modulate the host immune response [7]. The immune-modulating functions of helminthic parasites are used in the treatment of several immunological diseases, including inflammatory bowel disease, autoimmune liver diseases, and multiple sclerosis [25]. Mechanisms that might underlie the immunosuppressive effects include inhibition of Th1- (IFN-γ) and Th17-related (IL-17) cytokines, promotion of Th2-related cytokines (IL-4 and IL-5), release of Treg cell-related cytokines (IL-10 and TGF-β), and the induction of regulatory cells [25], [26]. In particular, the roles of Treg cells have been investigated for their role in host immune regulation of many parasitic infections [4], [8], [9], [27], [28]. In addition, Aranzamendi et al. have reported that T. spiralis infection increases the number of Treg cells and that adoptive transfer of CD4+ T cells from T. spiralis-infected mice suppresses lung inflammation [29]. However, there is little information regarding the direct adoptive transfer of Treg cells isolated from parasite-infected animals and its effects. In the present study, we assessed the functional and molecular characteristics of parasite-induced Treg cells using a mouse model of OVA-alum allergic airway inflammation. Adoptive transfer of Inf(+)Foxp3+ cells ameliorated airway inflammation by enhancing Treg-cell recruitment around sites of inflammation and thereby inhibiting the Th2 response.
We had three sets of questions regarding parasite-induced Treg cells. First, we asked, do parasite-induced Treg cells have a stronger effect on airway inflammation than natural Treg cells? If so, what are the differences between the two types of Treg cells? To answer these questions, we characterized the Treg-cell population (CD4+Foxp3+ T cells), which included both natural Treg cells (nTreg; CD4+CD25+Foxp3+T cells) and inducible Treg cells (iTreg; CD4+CD25-Foxp3+T cells), because T. spiralis infection promotes the proliferation and activation of both iTreg and nTreg cells [4]. Several previous studies have shown that both naturally occurring and antigen-driven Treg cells regulate allergen-induced Th2 responses in mice and humans [30]–[32]. In a cockroach allergen-alum model, McGee and Agrawal showed that adoptive transfer of either nTreg cells or iTreg cells reversed airway inflammation and AHR to methacholine; the effect lasted for at least four weeks. In our experiments, although cells from infected and uninfected mice had anti-inflammatory effects, the CD4+Foxp3+T cells isolated from T. spiralis-infected mice [Inf(+)Foxp3+] reduced artificially-induced airway inflammation to a greater extent than did CD4+Foxp3+T cells from uninfected mice [Inf(-)Foxp3+] (Fig. 2–5). These results might reflect the activation of CD4+Foxp3+T cells during T. spiralis infection. Inf(+)Foxp3+ cells expressed several surface proteins, including CTLA-4 and CD39 (Fig. 8B), which are related to dendritic cell (DC) regulatory functions [10]. CTLA-4 on the surface of Treg cells downregulates or precludes the upregulation of CD80 and CD86, the major co-stimulatory molecules on antigen-presenting cells [10]. Extracellular ATP, an indicator of tissue destruction, exerts inflammatory effects on DCs [10]. As an anti-inflammatory mechanism, catalytic inactivation of extracellular ATP by CD39 on Treg cells might prevent the harmful effects of ATP on DC function [10]. CTLA-4 also mediates T-cell downregulation during chronic filarial infections [9]. In addition, GARP and OX-40, which are expressed on the surface of activated Treg cells and regulate the bioavailability of TGF-β, were highly expressed in Inf(+)Foxp3 cells. GARP potently suppresses the proliferation and differentiation of naive T cells into T effector cells and suppresses IL-2 and IFN-γ production, leading to the differentiation of naive T cells into induced Treg cells [33]. The process is linked to Smad2/3 phosphorylation and is partially suppressed by the inhibition of TGF-β signaling. OX40 (CD137) is also generally expressed on mouse Treg cells. We observed that Inf(+)Foxp3+ cells repressed the proliferation of splenocytes, including T cells, to a greater extent than did Inf(-)Foxp3+ cells (Fig. 8C). Blocking OX40 on Treg cells with agonist antibodies inhibits the cells' ability to suppress and restores effecter T-cell proliferation [34], [35]. Capg, a cancer suppressor gene, is specifically upregulated in Treg cells during chronic helminth infection [36], [37]. We found that Gzmb gene expression was upregulated in Inf(+)Foxp3+ cells (Fig. 8B). Activated Treg cells also upregulate Gzmb expression, and Treg cells kill responder cells via Gzmb-dependent mechanisms [38]. Gzmb-deficient Treg cells have reduced suppressive activity in vitro [10]. Gzmb is released by in vitro-activated Treg cells, and it functionally drives apoptosis in naive B cells [39]. Gzmb is also upregulated in infected animals [37]. We found that OVA-specific serum IgE was reduced in Inf(+)Foxp3+ adoptive transfer mice. Thus, in answer to our first questions, these results showed that T. spiralis-induced Treg cells are more potent than natural Treg cells.
In our second set of questions, we asked, how do Treg cells regulate airway inflammation? Are they activated in the spleen, at peripherally located lymph nodes, or at inflammatory sites? In this study, although we could not directly evaluate Treg cell population in the lung by FACS analysis because we have technical limitation such for preparation from lung tissue. However, we observed many activated Foxp3-eGFP cells around inflammation sites in the airways, but it was difficult to detect cells in the absence of inflammation (Figs. 5 and 7, S1 and S4 Figs.). We also found that some Treg-cell homing receptors, for example, CCR5 and CCR9, were highly expressed in Inf(+)Foxp3+ mice, enabling the Treg cells to migrate rapidly to inflammation sites (Fig. 8). Many other homing receptors on Treg cells are involved in the inflammatory recruitment of Treg cells in different immunological settings, including CCR1, CCR2, CCR4, CCR5, CCR8, CCR9, CXCR3, CXCR4, CXCR5, CXCR6, and the P- and E-selectin ligands [40]. Treg cells first migrate from blood to the inflamed allograft, where they are essential for the suppression of inflammation [41]. This process is dependent on the chemokine receptors CCR2, CCR4, and CCR5 and the P- and E-selectin ligands [42]. The absence of CCR5 is associated with impaired recruitment of Treg cells and with decreased IL-10 expression, reflecting the receptor's potent anti-inflammatory activity [42]. Interestingly, levels of CCR9, a gut homing receptor, were higher in Inf(+)Foxp3+ cells than in Inf(-)Foxp3+ cells. The results suggest that T. spiralis stimulates Treg-cell recruitment to the intestine during the intestinal phase of infection.
Finally, we asked, is it more effective to transfer Treg cells before or after the induction of inflammation? We addressed this question in two stages. First, we determined the preventive effects of Treg-cell transfer [Stage I]. Second, we determined the therapeutic effects of the cells [Stage II]. The results suggest that both methods are effective in this allergic airway inflammation model. Introduction of Inf(+)Foxp3+ cells before airway inflammation induction elicited Treg-cell recruitment in the spleen and LLN and increased IL-10 production in the LLN (Figs. 3 and 5, S1 Fig.). Introduction of Inf(+)Foxp3+ cells during OVA challenge did not elicit Treg-cell recruitment in the spleen (Fig. 7). These results indicate that transferred Treg cells can long survive in recipient mice and rapidly migrate to inflammation sites using their activated homing receptors. In this study, more Foxp3 eGFP cells were found in the Inf(+)Foxp3+ adoptive transfer mice before allergic inflammation induction (day 0) than in the Inf(+)Foxp3+ adoptive transfer mice during OVA challenge (day 17) (Fig. 5 and 7). This phenomenon could not be explained by our present data, the possibility of proliferation of injected the cell in recipient mice will be evaluated in further study.
In conclusion, T. spiralis infection promotes the proliferation and functional activation of Treg cells, which migrate to inflammation sites and suppress the immune response more effectively than non-parasite-induced Treg cells. The adoptive transfer of Inf(+)Foxp3+ cells is an effective method for the prevention of allergic airway inflammation in mice and is a promising approach for the treatment of allergic airway diseases.
|
10.1371/journal.pcbi.1006597 | Self-organization of conducting pathways explains electrical wave propagation in cardiac tissues with high fraction of non-conducting cells | Cardiac fibrosis occurs in many forms of heart disease and is considered to be one of the main arrhythmogenic factors. Regions with a high density of fibroblasts are likely to cause blocks of wave propagation that give rise to dangerous cardiac arrhythmias. Therefore, studies of the wave propagation through these regions are very important, yet the precise mechanisms leading to arrhythmia formation in fibrotic cardiac tissue remain poorly understood. Particularly, it is not clear how wave propagation is organized at the cellular level, as experiments show that the regions with a high percentage of fibroblasts (65-75%) are still conducting electrical signals, whereas geometric analysis of randomly distributed conducting and non-conducting cells predicts connectivity loss at 40% at the most (percolation threshold). To address this question, we used a joint in vitro-in silico approach, which combined experiments in neonatal rat cardiac monolayers with morphological and electrophysiological computer simulations. We have shown that the main reason for sustainable wave propagation in highly fibrotic samples is the formation of a branching network of cardiomyocytes. We have successfully reproduced the morphology of conductive pathways in computer modelling, assuming that cardiomyocytes align their cytoskeletons to fuse into cardiac syncytium. The electrophysiological properties of the monolayers, such as conduction velocity, conduction blocks and wave fractionation, were reproduced as well. In a virtual cardiac tissue, we have also examined the wave propagation at the subcellular level, detected wavebreaks formation and its relation to the structure of fibrosis and, thus, analysed the processes leading to the onset of arrhythmias.
| Cardiac arrhythmias are one of the major causes of death in the industrialized world. The most dangerous ones are often caused by the blocks of propagation of electrical signals. One of the common factors that contribute to the likelihood of these blocks, is a condition called cardiac fibrosis. In fibrosis, excitable cardiac tissue is partially replaced with the inexcitable and non-conducting connective tissue. The precise mechanisms leading to arrhythmia formation in fibrotic cardiac tissue remain poorly understood. Therefore, it is important to study wave propagation in fibrosis from cellular to tissue level. In this paper, we study tissues with high densities of non-conducting cells in experiments and computer simulations. We have observed a paradoxical ability of the tissue with extremely high portion of non-conducting cells (up to 75%) to conduct electrical signals and contract synchronously, whereas geometric analysis of randomly distributed cells predicted connectivity loss at 40% at the most. To explain this phenomenon, we have studied the patterns that cardiac cells form in the tissue and reproduced their self-organisation in a computer model. Our virtual model also took into account the polygonal shapes of the spreading cells and explained high arrhythmogenicity of fibrotic tissue.
| The contraction of the heart is controlled by propagating waves of excitation. Abnormal regimes of the wave propagation may cause cardiac arrhythmia, asynchronous contractions of the heart and even lead to cardiac arrest and sudden cardiac death. Cardiac arrhythmias often originate from blocks of propagation [1]. In that case, the wave goes around the block, reenters the same inceptive region and, thus, forms a persistent rotational activity called cardiac reentry, which is one of the main mechanisms of lethal cardiac arrhythmias.
A normal heart has a complex structure, which is composed of the bundles of elongated cardiac cells. Apart from premier excitable cardiac cells, there are also inexcitable and non-conducting cells of connective tissue: cardiac fibroblasts. Their role is to maintain the structural integrity of the heart [2] and repair injuries [3]. Fibroblasts outnumber cardiomyocytes in a healthy human heart although occupying a much smaller total volume [2, 4]. However, many pathological conditions are associated with an excessive growth of the fibrous tissue, called cardiac fibrosis, which is, therefore, considered to be one of the major arrhythmogenic factors [5, 6].
The mechanisms of arrhythmia onset in fibrotic tissue remain poorly understood but generally believed to be associated with the increased probability of waveblock formation. It is a well-established fact that the presence of the non-conducting cells slows down wave propagation [7] and can completely block it if the non-conducting cells’ density is high. The critical density of non-conducting cells above which the conduction terminates is called percolation threshold. This concept originates from the percolation theory, and, by definition, it specifies the point of long-range connectivity loss/formation in random systems. Connectivity here refers to electrical synchronisation in the tissue, or, in other words, the ability to transmit electrical waves of excitation. Percolation threshold, i.e. the critical density of non-conducting cells which breaks long-range connectivity, plays an important role in arrhythmogenicity. It was shown that cardiac tissue is most susceptible to arrhythmias if the density of non-conducting cells is only slightly (∼2-3% [8]) below the percolation threshold. There are two main factors that may facilitate reentry formation. First, a large amount of non-conducting cells (acting as heterogeneities) increases the probability for waveblock formation [9]. Second, a high fraction of non-conducting cells creates a ‘maze’ that effectively lengthens the travel distance for the waves as they follow a longer zig-zag path [10] and, thus, provides sufficient room for the emerging reentrant loops. As a result, high density of non-conducting cells both facilitates the initiation and creates conditions for the existence of reentrant cycles, resulting in a highly arrhythmogenic substrate.
Up to now wave propagation failure was studied only in generic mathematical representations of heterogeneous cardiac tissue with each cell randomly chosen to be either conductive or non-conductive (see e.g. [9, 11]). In this kind of 2D computer models, the propagation of excitation failed at 40% of non-conducting cells [12], which is also within the range of values predicted by classic mathematical models (e.g. 37-44% of the area uncovered by conducting elongated ellipses with the shape similar to cardiomyocytes [13]). However, experimental measurements [14] indicate that wave propagation and synchronous contraction in 2D cardiac monolayers is observed for up to 75% percentage of non-conducting cells.
In this paper, we study the phenomenon of wave propagation in cardiac tissue with a high density of non-conducting cells using a joint in vitro-in silico approach. We performed experiments in 25 monolayers with various percentages of non-conducting cells and detected wave propagation to determine the percolation threshold. We have found that, indeed, the experimentally measured threshold (75% of the area covered by non-conducting cells) is substantially higher than what was predicted in conventional computer modelling (40% [12]) or classic mathematical models. Further morphological examination revealed that the key mechanism of conduction in highly heterogeneous tissue is likely to be tissue patterning. The cardiomyocytes were not located randomly but organised in a branching network that wired the whole sample.
Next, in order to explain cardiac network formation, we applied a virtual cardiac monolayer framework developed in [15], based on the Cellular Potts Models [16–18]. We proposed a hypothesis that such self-organisation occurs due to cytoskeletons’ alignment. Based on this hypothesis, we were able to obtain branching patterns, as well as reproduce the decrease in conduction velocity and wave percolation observed in experiments. We have further studied in silico the process of formation of the wavebreaks leading to reentry formation and analysed the tissue structures that caused them.
This paper is organized as follows. First, we describe the experiments conducted with the neonatal cell cultures. Second, we analyse the patterns and formulate the hypothesis on the mechanism of their formation. Third, we implement the hypothesis and reproduce pattern formation in silico in a Cellular Potts Model (CPM) of cardiac tissue [15]. Next, we compare electrical signal propagation in computer modelling and in experiments. Finally, we show a spontaneous formation of uni-directional blocks in the model resulting in reentry formation and locate and analyse the structures causing it.
We cultured neonatal rat ventricular cardiomyocytes mixed with cardiac fibroblasts in variable proportion (20-88% fibroblasts) and studied electrical activity in these monolayers using optical mapping.
In Fig 1a and in S1 Video the wave propagation in a sample with high fraction of non-conducting cells is shown (66%). In spite of high percentage of non-conducting cells, this sample is still conducting electrical waves, however the wave propagation pattern is complex. The wave originates from the stimulation point marked by a yellow spike at the bottom of the tissue and initially spreads in all directions. However, due to a large number of non-conducting cells, the wave is blocked at multiple sites. After a short delay (in areas outlined with dashed ellipses), the wave propagates further into the left and the right parts of the sample, and again spreads in various directions. This process repeats multiple times, resulting in a complex, fractionated wave pattern containing narrow pathways and some bulk excited regions. The conduction blocks in Fig 1a are shown in red, which are the places where the wave propagation was blocked and the wave had to go around. The main propagation paths are shown with white and black arrows. In this sample, the electrical wave propagation was still possible and long-range connectivity was still present, however the amount of non-conducting cells was close to the percolation threshold.
We have found that the percolation threshold for the neonatal rat cardiac monolayers was 75 ± 2% of non-conducting cells. We have also measured conduction velocities in the samples below the percolation threshold (Fig 1b). The measurements show that the velocity decreased when approaching the percolation threshold. In the samples with low level of non-conducting cells, the velocity was approximately 10-14 cm/s, and it decreased twofold in the samples with 70% non-conducting cells. The number of conduction blocks was higher in samples with high percentage of non-conducting cells. As a result, the mean conduction velocity decreased with the increase in a percentage of non-conducting cells.
After optical mapping of the wave propagation, we have fixated the samples and studied their morphology using immunohistochemical labeling. We have found that the cardiomyocytes in the samples have formed connected networks capable of electrical wave propagation. In Fig 2 cardiomyocytes are shown in pink, and the cluster that they have formed is outlined with a white contour. The cardiomyocytes were organised in a branching structure that wired the whole sample. We have followed the pathway using a confocal microscope and it was possible to find long-range connectivity in the tissue. Therefore, we observed that the cardiomyocytes were organised in conduction pathways and assumed that there must be a mechanism responsible for their self-organisation.
Pattern formation in cell populations during development was extensively studied using Cellular Potts Models [16, 17, 19], including many studies [20–22] of the branching structures similar to the one in Fig 2. Therefore, we have considered several existing hypotheses to explain the formation of the cardiac pathways.
First, we suggested that differential adhesion together with cell elongation may be enough to explain the observed patterning. A similar system with autonomously elongating cells was successfully used to describe vasculogenesis [20]. However, we did not impose obligatory elongation on the cardiac cells, because they are not necessarily elongated according to our experimental observations (see Fig 3 below or S2 Video). Cardiomyocytes obtain their typical brick-like shape with the guidance of the extracellular matrix over the course of development. However, there is no evidence for any internal autonomous mechanism for elongation. In experiments on the glass, cells were not only bipolar but also tripolar and multipolar (see S2 Video). In our tissue growth model [15], which was accurate at replicating realistic cell shapes, this feature was reproduced with explicit introduction of the actin bundles. Cooperation between aligned actin bundles pulling in one direction shaped the cell more efficiently, which led to clustering of these virtual actin bundles and resulted in multipolar cell shapes. Similar approach was also previously used to describe the shapes of dendritic cells [23]. Therefore, since the elongation was not imposed, there was no mechanism forcing cardiomyocytes out of clusters to search for new connections.
Next, we considered various mechanisms that were previously used to explain similar branching patterns in angiogenesis. The main sources of instability in those models were chemotaxis and contact-inhibition [21]. The percolation in the networks formed due to sharp gradients of chemoattractants was also studied previously in a continuous model [22]. However, there was no evidence of any directed migration of cardiac cells and for type-specific contact-inhibition, similar to those that select a tip cell in a growing blood vessel. Therefore, we discarded this hypothesis either.
After trying several approaches used before, we have not found an existing model that could have been 1) applicable to the cardiac tissue and at the same time 2) could reproduce the experimentally observed branching structures. Finally, after careful analysis of our experimental preparation, we found that an essential feature of our structure was the alignment of the cytoskeletons in the neighbouring cells. This alignment of the actin fibres is clearly seen in our experimental preparations. In Fig 3 the bundle of neonatal rat cardiomyocytes is shown. The red arrows indicate the intercalated disks between the cells. The white arrows point at the actin bundles on the opposite sides of the intercalated disk, that smoothly continue one another.
The precise biological mechanism of such alignment is unknown. However, in our view, it can be derived from a well-known property of actin cytoskeleton reinforcement in response to the external force [24]. Actin filaments, as well as the adaptors that link them, remodel under applied tension. In case of contact of two cells, the actin filaments are connected through adherens junctions, which transmit the tensile forces between the filaments of the neighbouring cells. It was shown that higher tension stabilises the whole complex [24]. The tension is maximal, if the actin filaments are aligned with each other, which gives a preference for intercellular alignment of the cytoskeletons.
Therefore, we have incorporated this mechanism into our Cellular Potts Model (CPM) of cardiac cells [15]. This model was already adjusted to reproduce characteristic shapes of the cardiac cells in virtual tissue model, and here we extended it with a new energy term, that corresponded to the alignment of actin bundles in the neighbouring cells.
In our CPM model, cell-substrate adhesion sites, to which actin bundles are anchored, were represented as separate entities: specially labeled subcells of the lattice. If two adhesion sites of two cells came into contact, we established a new connection between them. The connection means, that an additional bond energy was applied to them. This energy depended on the angle between the linked actin bundles (see Fig 4). The minimum of the energy corresponded to smooth coupling between the bundles, or zero angle. In this case, the bond was the most stable, but couplings with the non-zero angle between the bundles had a tendency to break apart.
With this new energy term that favours cytoskeletons alignment, the cardiomyocytes in simulations created branching patterns. In Fig 5a, a resulting simulated structure of the sample with 70% non-conducting cells is shown. We see that in this sample only 30% of cardiomyocytes were able to build a network, and even with such a high density of non-conducting cells, this network was fully interconnected. Our further studies showed that such interconnection was established for every sample with 30% of cardiomyocytes (n = 10) of 1 cm × 1 cm size, which is close to the size of the Petri dish used in our experiments. For samples with 28% of cardiomyocytes, the network was interconnected with 20% probability. We also see that the patterns in computer simulation (Fig 5a) and in experiment (Fig 5b) have similar features, such as long single-cell-wide connections that bridged the gaps between cell clusters, or isolated non-conducting clusters trapped within the main cardiac pathway. Therefore, using the hypothesis on cytoskeletons’ alignment allowed us to reproduce not only the percolation threshold, but also the main features of the pattern.
S2 Video shows the growth of the branching pattern, highlighting the large connected clusters of the cardiomyocytes by different colours. One can see from the video, that subtle movements of the cells (left) result in dramatic changes in connectivity (right). This emphasises the important role of the protrusions of the cells, which are taken into consideration in our model, in electrical signal propagation.
Using this model with cytoskeleton alignment, we have studied further wave excitation patterns and the percolation threshold for electrical waves in the system.
We have reproduced the experiments from Fig 1a in silico using Majumder et al. [25] model for neonatal rat ventricular cardiomyocytes.
Fig 6a and S3 Video show wave propagation in a virtual sample with 70% non-conducting cells. The wave propagation pattern is complex and similar to one observed in experiment (see Fig 1a). One can see, that the number of propagation blocks per unit area is similar to that of the experimental activation pattern, and the trajectories of the waves share the same features. Note, that the spatial scale of the simulated activation map is slightly smaller than those of the experimental one.
The percolation threshold in virtual cardiac monolayers was equal to 71.5 ± 1.5% of non-conducting fibroblasts, meaning that the samples with such level of fibroblasts had 50% probability of percolation. In our simulations, 100% of the samples (n = 10) with 70% fibroblasts were interconnected, whereas samples with more than 73% fibroblasts (n = 10) were never conducting. For 72% fibroblasts 20% of the samples were functional and for 71% fibroblasts, those were 80%. The conduction velocity dependence on the density of fibroblasts is shown in Fig 6b. For each simulated sample we have indicated the mean value of the velocity, and the standard deviation of the velocity distribution was shown with error bars. One can see, that the dependency of the velocity on the fibroblasts’ density is similar to one measured in experiments (see Fig 1b). Closer to the percolation threshold the fluctuations in conduction velocities amplified due to the stochastic nature of the percolation block. Moreover, variations in conduction velocities within one sample were very high in that range, because of a large number of conduction blocks.
Thus, the percolation threshold in our simulations with Ebond = 5.0 was only slightly lower (≈ 72%) than in experiment (75%). However, both values are much higher than the predictions of the models with random cells distribution (40%).
Fig 6b also shows the percolation threshold in a computer model without cytoskeleton alignment (Ebond = 0, blue), which was equal to 64%. It is still higher than the threshold for randomly distributed cells (40%) since some clustering of the cells takes place. The reason for this clustering is that the adhesion between FB is slightly stronger than between cardiomyocytes or between cells of different types (JFB−FB = 500 vs. JCM−CM = JCM−FB = 700). Such a difference is required to reproduce the difference in cell shapes correctly: all surface energies of the fibroblasts have to be lower than those of cardiomyocytes to reproduce dynamic fluctuations of their boundaries. The model in this study inherits some parameters from our previous paper [15], however it was reparameterized to ensure long-term stability of the model (on the scale of 20’000-50’000 MCS). We have first tried to enhance differential adhesion, but discovered that the differential adhesion alone was not capable of reproducing the branching pattern observed in experiments (as was discussed above). Thus, differential adhesion was tuned down as much as possible (JCM−CM set to be equal to JCM−FB), preserving the desired cell shapes and volumes. Nonetheless, the resulting parameters cause some clustering even for Ebond = 0, which, in turn, raises the percolation threshold from 40% up to 64%. These data show to which extent the accurate representation of cell shapes and clustering of the cells of different types affect the percolation threshold (+24% compared to a simple square lattice model), and how much it is affected by cytoskeletons’ alignment (extra +8%).
We have studied various values of Ebond and have found that there is an optimal range of values that provides the highest percolation threshold. One can see in Fig 7 that both low (Ebond = 0.0) and high (Ebond > 7.5) cytoskeletons’ coupling energies result in the low percolation thresholds. When the coupling is weak, the cardiomyocytes do not form strong connections between each other and form clusters rather than networks (see top-right pattern). When the coupling is strong, the cardiomyocytes form very strong connections instead, which effectively “freeze” these cells in place and do not allow any motility, which is essential for network formation (see bottom-right pattern). Therefore, intermediate values of the Ebond are optimal for network formation, as they favour cytoskeleton alignment sufficiently, yet without imposing too much restriction on cells’ motility.
We have chosen Ebond = 5.0 for this study, since it belongs to the optimal range of values and, at the same time, reproduces the most qualitative features of the experimental samples. One can see in the middle-right image in Fig 7, that there are some fibroblasts (white cells) trapped in the clusters of cardiomyocytes (red cells). It is an important characteristic feature of the pattern, which is also present in our experimental samples (see right-hand image in Fig 5).
In Fig 7, we have shown two definitions of the percolation threshold corresponding to 10% and 50% probability of successful network formation. The 50% is typically used in theoretical studies [8], however here we compare the probability corresponding to 10% with the experiment. The reason for this is that it is not possible to carefully assess the probability of network formation in experiments. There are many other factors that may cause the propagation failure in cell culture and large statistics is needed to find the exact value corresponding to 50% probability of conducting network formation. Therefore, we have recorded all successfully conducting samples including those that belong to a vicinity of the percolation threshold and were still, by chance, conducting. Therefore, we suggest that 10% probability threshold provides a practical estimation of the highest percentage of non-conducting cells which we are likely to witness electrically connected, and thus this value is better for comparison with the experiment.
We have shown that in virtual tissues a spontaneous onset of reentry can occur. It was observed in samples with a high level of non-conducting cells. Fig 8 shows such process for a sample with 70% non-conducting cells. We see that after the first stimulus applied to the left border of the sample (shown in yellow), the wave propagation was blocked at the lower part of the sample, but it pursued through the upper part. After reaching the right border, the wave turned around and entered the bottom region. However, this first wave was blocked in the middle of the sample, as the tissue in the upper part of the sample have not been recovered yet. The wave from the second stimulus, applied at the same site, has followed the same path and formed a sustained circulation along the path shown with the red dashed line in Fig 8.
Detailed analysis of the structure revealed that formation of the reentry here is solely due to specific structure which is shown in the middle (inside the yellow square), which acts as an area of uni-directional block (or “diode”): the waves can propagate from right to left, but not in the opposite direction. The diode is formed by two cell clusters that barely touch each other. These cell clusters have slightly different areas adjacent to this connection. Therefore, if the wave propagates from left to right (from a smaller cluster to a larger one), then the small cell cluster does not produce enough current and can not depolarise a bigger one. The transmembrane potential in the largest cluster raises (which is shown in pink colour), but not enough for the sodium channels to open. This effect is called source-sink mismatch [1]. When the source (a smaller cluster) is insufficient compared to the sink (a larger inactive cell cluster), the wave propagation is blocked. This effect was observed in chemical systems [26] and later in cardiac monolayers [27].
In a sample in Fig 8, the “diode” could initiate a reentry if the sample is stimulated from the left or from the top with a high frequency (4 Hz or more).
We have performed studies in 6 large samples with 70% of non-conducting cells. All of these samples had 2-5 areas of the uni-directional block and many bi-directional blocks, but only one of these samples was arrhythmogenic. Thus, in addition to “diodes”, some extra geometrical conditions are required. The precise conditions are to be studied in the future, however, they are related to the presence of the long conducting circuits in the tissue. In fact, in Fig 8 we see that apart from the diode, there is also a loop, which is shown with a red dashed curve in the bottom-right image. The diode is a part of this loop, however, the loop is large enough to account for recovery of the bottom part of the tissue after one rotation. The samples with a higher density of non-conducting cells were even less likely to have long circuits, thus we did not observe sustained reentry there. We concluded, that reentry formation requires not only uni-directional blocks but also long circuits, and the densities of non-conducting cells slightly below the percolation threshold are the most arrhythmogenic ones. This result was previously shown for random cell distributions, that reentry is most likely to occur 5-10% below the percolation threshold [8]. The principle holds true in our model, but quantitatively the densities of non-conducting cells are different.
Our model shows, that the areas of the uni-directional block can be naturally formed during tissue growth. Every time one spreading cell comes in contact with the other cell cluster, there is a chance that this connection will be asymmetric. It may explain the fact that reentries are frequently observed in the experimental setups with neonatal cardiac monolayers.
We observed paradoxical electrical wave propagation in samples with up to 73–75% of non-conducting cells instead of mathematically predicted 40% for randomly distributed cells. We have shown both in vitro and in silico, that electrical wave propagation was possible due to the formation of the conduction pathways that rewired the whole monolayer. We have proved the existence of this branching network with immunohistochemical images. We have measured the conduction velocity, which decreased with the increase of the portion of non-conducting cells in the monolayer in a similar way in experimental and computational studies.
To explain the formation of the pathways, we have considered several existing hypotheses based on differential adhesion, cells elongation and directed migration. However, all of them were drawn inapplicable to the cardiac tissue. As a result, we developed and proposed a new hypothesis, based on cytoskeleton alignment. Assuming that cardiomyocytes align their cytoskeletons to fuse into cardiac syncytium, the morphology of conductive pathways was successfully reproduced in computer modelling. The virtually generated monolayers were then used for the studies of the electrical wave propagation, and experimentally observed velocity decay and high percolation threshold were reproduced as well.
The proposed hypothesis of cytoskeletons alignment has never been considered before. This mechanism, however, has some similarities to diffusion limited aggregation (DLA): a process in which random-walking particles form fractal aggregates. In our model, a cell ‘sticks’ to the growing pattern, but, unlike DLA, our cells also align cytoskeleton’s orientation with the pattern. Such alignment actually takes place as a part of syncytium formation, when cardiomyocytes fuse and align their cytoskeletons. Once a cell aligns cytoskeleton with its neighbours, this structure maturates and fixes the cell in place. In some sense, this process causes contact inhibition only between cardiomyocytes and not between cardiomyocytes and non-cardiomyocyte cells. Such type-specific contact inhibition results in the branching structure like one shown in angiogenesis [21].
Formation of conduction pathways and complex texture of the tissue affects arrhythmogenicity, and may also be important for arrhythmia treatment. For example, it was shown that texture of cardiac tissue at subcellular level can substantially affect the propagation of external current during defibrillation [28, 29]. It would also be interesting to quantify the excitation patterns in terms of the number of re-entrant sources and wavefront complexity [30]. Therefore, it will be interesting to perform a similar study for the textures generated with our model.
There are several limitations of the methods used in this study. First of all, we have conducted experiments with cell cultures, which are different from the real cardiac tissue. It would be interesting, yet more complicated, to study patterning of the real 3D cardiac tissue. However, the mechanisms of patterning that we discovered here are likely to be universal and might be applicable to the real cardiac development as well. Second, fibrosis is a more complex condition than just excessive growth of fibroblasts. It also involves collagen deposition that insulates the cardiac fibers from one another and is also considered to increase arrhythmogenicity. In this study, we did not take the extracellular matrix into account, but it would be also interesting to measure its effect on the percolation threshold in the future studies. Third, recent studies suggest that there are other conducting cell types in the heart, such as myofibroblasts [31] and macrophages [32, 33]. The fraction of these other cell types in the heart does not exceed 5-7% [34, 35], and, moreover, their amount in ventricular cell population, used in this study, is negligible [32]. However, presence of such cells in other cell populations may affect the percolation threshold. Finally, percolation depends on the size of the sample and scales with this size. For random systems, the scaling laws are known. In our case, we did not consider scaling and used similar size of the samples in silico as in the experiment. However, it would be interesting to study scaling of the percolation threshold and see how the size can affect the probability of wavebreak formation.
We conclude, that the cardiomyocytes in heterogeneous tissues with high portion of non-conducting cells can form a connected network and allow electrical signal propagation in monolayers containing up to 75% of non-conducting cells.
All studies conformed to the Guide for the Care and Use of Laboratory Animals, published by the United States National Institutes of Health (Publication No. 85-23, revised 1996) and approved by the Moscow Institute of Physics and Technology Life Science Center Provisional Animal Care and Research Procedures Committee, Protocol #A2-2012-09-02.
In this study, we have used a mathematical model that was previously developed for cardiac monolayers formation [15]. It is based on the Cellular Potts Model (CPM) formalism, which describes cells as the domains of the regular lattice and assigns energy to the system of cells to describe their growth and motility. In our previous work [15], we have selected the main features of the cardiac cells (such as area, number of protrusions, etc.) and parametrised our model to reproduce the cell shapes. In this study, we extended the model with a new energy term responsible for syncytium formation.
The evolution of the cardiac monolayer is described by the Hamiltonian:
H = H adhesive + H elastic + H protr + H nuclei + H junctions , (1)
where Hadhesive + Helastic is the basic CPM model and Hprotr is the term describing the protrusion dynamics of the cardiac cells, which produces a characteristic polygonal shapes of these cells. Hnuclei corresponds to higher rigidity of the nuclei compared to the cell body. Finally, Hjunctions is a new term, that describes the stability of adherens junctions and alignment of the cytoskeletons of the neighbouring cells.
The core feature of the model of cardiac cells is the explicit representation of cell attachments as the labelled subcells of the lattice. They are first assigned when the cell expands, and can be destroyed with a certain penalty, if, for example, the cell is stretched due to its movement. The number of attachments per cell is limited by the amount of actin present in the cytoplasm, which is also reflected in the model. If the maximal number of attachments is reached, new attachments do not form.
The spreading of virtual cells is shown in S5 Video. One of the important features of our approach, is that elongation of the cells is not imposed, but rather emerges due to interactions with the environment. Cardiac cells in a real heart are indeed elongated, as they are guided by the extracellular matrix, whereas in a Petri dish cells can spread in any direction with no preference. However, this does not result in a circular shape due to the limited number of attachment sites where spreading occurs. Instead of explicit declaration of elongation, we suggested that virtual cells have only a small (≈10) number of well-developed mature attachment sites. Moreover, these sites in a model tend to cluster spontaneously (see S5 Video), as the actin strands pull the cell more efficiently acting together, rather than separately. Therefore, these clustered attachment sites turn cell shapes into bipolar, tripolar or multipolar states. The bipolar state is the most stable one for isolated cells, however tripolar cells are also relatively common. This correctly represents the population of cells observed in experiments.
Hprotr is an energy term, that decreases with the distance from the centre of mass of the cell, and which is applied only to these labeled attachment sites. As a result, the attachment sites spread out and reproduce characteristic polygonal shape of the cardiac cells. Here, in this model, we omit the details of the spreading process, which involves polymerization and depolymerisation of the actin, attachment/detachment, etc., but we mimic the overall dynamics of the protrusions. Also, we assume, that for every attachment site a corresponding actin bundle exists, which stretches from the attachment site towards the proximity of the nuclei.
If two attachment sites in the neighbouring cells come into contact, the bond between these attachment sites can be formed. In the algorithm, this bond appears if one attachment site attempts to move over the other. In such case, instead of copying of the subcell, the connection establishes. A new energy term Hjunctions applies to the subcells, which are involved in a newly established junction. This term determines the stability of the cell-to-cell junction and depends on the angle between the actin bundles, associated with the attachment sites involved (see Fig 4). It is determined as follows:
Hjunctions=∑i→,j→labeledjunctionEbond(1−cos(αi,j)),
where αi,j (shown in Fig 4) is the angle between two cytoskeleton bundles of two neighbouring cells which ends at points i and j labelled as the parts of the junction. The wider is the angle, the less stable is the junction. Therefore, the junctions with continuous actin bundles on both sides persist, but the kinked bundles tend to lose the connection. As a result, the actin bundles of the neighbouring cells tend to align and stay in the aligned state.
The parameters of the model used in simulations (see Table 1) in this paper were adjusted to compensate the additional energy term Hjunctions. The most of them are close to the parameters used in our original paper [15]. The value of the new energy term was set to Ebond = 5.0, which provides enough stability for the junctions to maintain the branching structure but at the same time not too much stability to allow cells to search for possible new connections. Addition of Hjunctions effectively increased the adhesion between cardiomyocytes. Therefore, the differential adhesion was toned down in a model to allow cardiomyocytes to migrate randomly before they maturate and stick to the pattern. In this study, we have changed type-specific adhesion coefficients Jcell−cell. Our choice of the parameters was guided by the desired balance between branching and clustering: 1) lower adhesion energy (J) to the cells of the same type increased the size of unstructured, irregularly shaped clusters; 2) on the other hand, high adhesion energy forced cells to migrate out of clusters. We have chosen neutral values of JCM−CM = JCM−FB, which meant that cardiomyocytes had no preference in neighbours. Their energy was equal for being surrounded with either other cardiomyocytes, or with the non-conducting cells. The non-conducting cells had a slight tendency to cluster (JFB − FB < JCM − FB). This choice of parameter allowed us to qualitatively reproduce the patterns observed in experiments (see Figs 2 and 5).
The code used in this study is available on Github: https://github.com/NinelK/VCT. Examples 3 and 4 in the repository correspond to formation of cardiac pathways with or without cytoskeletons alignment.
The structure of the virtual samples was first generated with the Cellular Potts Model. After the cells were grown and statistical characteristics of the cells and clusters stopped changing (50’000 MCS), the resulting lattice was converted into a matrix of coupling coefficients for electrophysiological studies. The electrical coupling between non-conducting cells and any other cells was set to zero, coupling of the lattice points within the cells was set to Din, and, finally, coupling between cardiomyocytes was set to either DL or DT depending on the orientation of the cells’ virtual cytoskeletons. The methodology was already described in detail in our previous paper [15] and the ratio between coefficients was adjusted to match the anisotropy of the wave propagation in aligned cells (Din = 100 × DL, DT ≪ DL and negligible). The magnitude of the coupling coefficients was slightly adjusted (Din = 0.4 cm2/s, DL = 0.4 × 10−2 cm2/s, DT = 0 cm2/s) to reproduce experimentally measured maximal conduction velocity in the control samples with low portion of non-conducting cells. The same coefficients were then used for all simulations.
|
10.1371/journal.ppat.1004424 | Exosomes from Hepatitis C Infected Patients Transmit HCV Infection and Contain Replication Competent Viral RNA in Complex with Ago2-miR122-HSP90 | Antibodies targeting receptor-mediated entry of HCV into hepatocytes confer limited therapeutic benefits. Evidence suggests that exosomes can transfer genetic materials between cells; however, their role in HCV infection remains obscure. Here, we show that exosomes isolated from sera of chronic HCV infected patients or supernatants of J6/JFH1-HCV-infected Huh7.5 cells contained HCV RNA. These exosomes could mediate viral receptor-independent transmission of HCV to hepatocytes. Negative sense HCV RNA, indicative of replication competent viral RNA, was present in exosomes of all HCV infected treatment non-responders and some treatment-naïve individuals. Remarkably, HCV RNA was associated with Ago2, HSP90 and miR-122 in exosomes isolated from HCV-infected individuals or HCV-infected Huh7.5 cell supernatants. Exosome-loading with a miR-122 inhibitor, or inhibition of HSP90, vacuolar H+-ATPases, and proton pumps, significantly suppressed exosome-mediated HCV transmission to naïve cells. Our findings provide mechanistic evidence for HCV transmission by blood-derived exosomes and highlight potential therapeutic strategies.
| Since its first isolation and identification in 1989, Hepatitis C virus (HCV), has caused significant disease burden to humans worldwide. So far, there is no vaccine against HCV, and neutralizing antibody therapies to block receptor–mediated transmission of HCV to liver cells have so far achieved limited therapeutic benefits. This indicates that HCV can transmit infection via receptor-independent mechanisms. Evidence suggests that small host extracellular vesicles (exosomes) can mediate receptor-independent transfer of genetic material between cells, though their role in HCV transmission remains uncertain. Here, we found that the HCV virus can utilize host exosomes to transmit infection to naïve liver cells, even in the presence of potent blocking anti-HCV receptor antibody treatments. Additionally, we identified alternative treatment strategies that can block host exosomes from transmitting HCV infection. Our study provides novel insights to an alternative mechanism of HCV transmission that can compromise anti-HCV immune therapies and proposes potential therapeutic approaches to block exosome-mediated transmission of HCV infection.
| Hepatitis C virus (HCV) infection is one of the leading causes of liver disease with over 170 million individuals chronically infected worldwide [1], [2]. Severe complications including fibrosis, cirrhosis, and hepatocellular carcinoma are among the long-term effects of HCV infection, making liver transplantation the ultimate choice of treatment for advanced liver disease [3]. Even with successful liver transplantation, patients face eminent HCV re-infection of the newly transplanted liver. Recent therapies with anti-HCV E1-E2 or other neutralizing antibodies that attempted to block HCV transmission achieved only limited success [4]–[7].
HCV is a positive-sense single-stranded RNA enveloped virus of the Flaviviridae family. The 9.6 kb HCV genomic RNA encodes a single polypeptide that is proteolytically cleaved to structural (core, E1, and E2) and non-structural (p7, NS2, NS3, NS4A, NS4B, NS5A and NS5B) HCV viral proteins [8]. The HCV viral envelope E1 and E2 proteins engage numerous host cell proteins for viral entry including CD81 [9]–[11]. CD81 interaction with HCV E1/E2 is critical in HCV entry and anti-CD81 or anti-E1/E2 antibodies have been shown to block HCV virus entry [7], [12]. Given the importance of these viral envelope proteins in regulating HCV infection, numerous immune therapies have been developed to specifically target and/or neutralize HCV envelope proteins [7], [13]–[15]. Targeted antibody therapies have offered limited success in preventing liver allograft infection by HCV. Recently, a potent human-derived monoclonal antibody was demonstrated to effectively prevent and treat HCV1 infection in chimpanzees [7]. However, the same antibody was not completely effective in humans [7], raising the possibility of other mechanisms of virus entry into hepatocytes. Previous reports have suggested receptor independent transmission of HCV [6], [16], though the precise mechanisms or possible therapeutic strategies remain to be explored.
Exosomes are a subpopulation of extracellular vesicles that originate from multivesicular bodies (MVBs), ranging from 40–150 nm in size and are produced by most cell types. These vesicles can be detected in blood, urine, and other body fluids [17]. Exosomes can modulate signal transduction, antigen presentation to T-cells, and transmission of genetic material between cells [18]. Over the past decade, a great body of evidence shows that exosomes can be secreted into the extracellular space and can mediate indirect cell-to-cell communication by transferring bio macromolecules, functional proteins, and RNAs between cells [19], [20].
HCV infection occurs via cell free virus and direct cell-to-cell transmission [6]. Indirect cell-to-cell transmission is another pathway to consider. Previously, HCV viral RNA has been identified in supernatant of HCV-SGR cells [21] and exosome-like structures have been detected in the supernatant of HCV infected cells [22] and in the plasma of HCV-infected patients [23]. Recently, Dreux et al (2013) showed that HCV-RNA-containing exosomes can mediate transfer of RNA to co-cultured plasmacytoid dendritic cells (pDCs) and trigger the production of type I interferon (IFN) in vitro [24].
Here we tested the hypothesis that exosomes derived from HCV infected hepatocytes or from the sera of HCV infected patients carry viral RNA and exploit the cellular exosomal delivery system to mediate receptor-independent HCV transmission to hepatocytes. We found that exosomes derived from HCV infected Huh7.5 cells and HCV infected patients contained HCV RNA that induced active infection in primary human hepatocytes (PHH). These exosomes were rich in replication-competent HCV-RNA in complex with miR-122, Ago2, and HSP90; and mediated HCV transmission independent of CD81, SB-RI and APOE. Mechanistically, functional inhibition of miR-122, HSP90 or modification of cellular micro-environmental pH using a vacuolar-type H+-ATPase (V-ATPase) and proton pump inhibitor significantly suppressed the capacity of exosomes to mediate HCV transmission.
Given our interest to investigate the capacity of exosomes derived from HCV J6/JFH-1-infected Huh7.5 cells and serum of HCV infected patients to mediate active transmission, we had to efficiently separate exosomes from the free HCV virus. Because HCV virions and exosomes have very similar sizes and densities, the traditional ultracentrifugation and sucrose gradient isolation method is insufficient for isolating pure exosomes free of virus contamination. To overcome this limitation, we optimized a CD63 immuno-magnetic isolation method to purify exosomes from cell culture supernatants of HCV J6/JFH-1 infected Huh7.5 cells and sera of HCV infected patients. Briefly, after serial filtration (1 µm, 0.44 µm and 0.22 µm) of supernatants, exosomes were initially isolated using Exoquick. To further purify exosomes and exclude other microparticles or free HCV contamination, Exoquick-isolated exosomes were subjected to immuno-magnetic selection with CD63, a selection marker of exosomes. This protocol was verified by analysis for other exosomal markers by western blotting (CD9 and CD81), electron microscopy, and Nanoparticle tracking analysis (NTA) (Figures S1A, S1B & S1C). The Exoquick-CD63 immuno-magnetic selection procedure recovered more exosomes compared to, ultracentrifugation-CD63 immuno-selection of exosomes (Fig. 1A & 1B); based on this observation we used the Exoquick followed by CD63 immuno-magnetic selection for subsequent experiments. We observed that in a fixed total volume there were significantly more free HCV viral particles compared to the number of exosome particles in HCV J6/JFH-1 infected Huh7.5 cell supernatants [approximately 7∶1 ratio] and in HCV infected patient serum samples [approximately 4∶1 ratio] (Fig. 1C). We also found higher HCV viral copy numbers in the free virus fraction compared to exosomes in J6/JFH-1 infected Huh7.5 cell supernatants (Fig. 1D) and HCV patients' serum (Fig. 1E). Purified exosomes were analyzed by transmission electron microscopy, demonstrating their vesicular shape and size range between 50 and 100 nm (Figure S1A). Further analysis with NanoSight demonstrated comparable histogram size plots of exosomes from culture supernantants of HCV J6/JFH-1 infected Huh7.5 cells and exosomes from serum of HCV infected patients (Figure S1B & S1C).
To rule out exosome contamination with free HCV virus, we carried out a simulation experiment mixing cell free HCV virus with uninfected exosomes from Huh7.5 cell culture supernatants for 24 h and re-isolated exosomes with Exoquick followed by CD63 immuno-selection or ultracentrifugation followed by CD63 immuno-selection. The uninfected exosomes exposed to free HCV virus showed no detectable HCV viral RNA while HCV RNA was present in the flow through following immuno-magnetic CD63 selection of exosomes (Fig. 2A & 2B). Further characterization of exosomes and free virus showed that isolated exosomes contained Apolipoprotein B (APOB) which was not present in cell free HCV viruses (Fig. 2C). Apolipoprotein E (APOE) was found to be associated to a large extent with HCV virus fraction and significantly lower in exosomes compared to cell free virus fractions (Fig. 2D). These observations suggest that our purified exosomes were to a large extent devoid of lipo-viral contamination. RNase H treatment to destroy free RNA in the cell free virus concentrate and isolated exosomes from HCV infected Huh 7.5 cells failed to prevent transfer of HCV infection to naïve cells, thus ruling out the possibility of envelope free viral RNA mediating HCV infection. These data indicate that both HCV derived exosomes and HCV virus are resistant to RNase treatment similar to the previous report [25] and still cause productive infection even after RNase treatment (Figure S2).
First, we established efficient methods for exosome purification using Exoquick followed by CD63-based isolation as described above. These exosomes were devoid of free HCV virus contamination as detailed. Exosomes isolated from sera of some HCV-infected patients or supernatants of HCV J6/JFH1 infected Huh7.5 cells contained comparable HCV RNA content for the same number of free HCV viral particles compared to the same number of HCV exosome particles (Fig. 3A). These observations allowed us to use the same number of infectious HCV viral particles and HCV exosomes for subsequent experiments. Treatment-naïve and non-responder (interferon plus ribavirin) patients with active HCV infection had detectable HCV RNA in serum-derived exosomes (Fig. 3B). In contrast, treatment responders, who cleared HCV infection, showed no detectable HCV in exosomes (Fig. 3B). Additionally, interferon alpha treatment of Huh7.5 cells had no effect on the number of exosomes released from hepatocytes (Figure S3) compared to untreated cells.
Recent studies have consistently demonstrated that miR-122, Ago2, and HSP90 enhance HCV replication [26]–[30]. We found that HCV J6/JFH-1 infected Huh7.5 cells produced exosomes that are enriched in Ago2 and contain barely detectable HSP90 protein compared to control exosomes from Huh7.5 cells (Fig. 3C). Interestingly, exosomes from HCV infected treatment-naïve and treatment non-responder individuals, but not treatment responders, were rich in Ago2 and HSP90 (Fig. 3D) compared to control healthy uninfected individuals. GW182 a RISC complex protein which we recently identified as an enhancer of HCV replication associated with alcohol use [31], was not detected in exosomes in our experimental conditions.
Micro RNA-122, a host factor utilized by HCV for replication, was present in exosomes isolated from both HCV J6/JFH-1 infected Huh7.5 cells and HCV-infected individuals (Fig. 3E & 3F). We observed that exosomes from HCV J6/JFH-1 infected Huh 7.5 cells showed higher levels of miR-122 compared to exosomes from non-infected cells (Fig. 3E), while exosomes from HCV-infected patients contained lower miR-122 levels compared to those from healthy controls (Fig. 3F).
Exosomes were recently shown to mediate retroviral infection independent of envelope protein-receptor interaction [32]. More recently, exosomes from Huh7.5 infected cells were found to induce type I interferon production in dendritic cells [24].
Observation of the presence of HCV RNA in exosomes prompted us to evaluate if exosomes from J6/JFH-1-infected hepatocytes or from HCV infected individuals could transmit infection to uninfected cells. We found that exosomes derived from supernatants of HCV J6/JFH-1 infected Huh7.5 cells mediated HCV infection after co-culture with uninfected Huh 7.5 cells (Fig. 4A and Figure S4) which could be inhibited by Telaprevir (VX-950) an NS3.4A serine protease inhibitor (Fig. 4B). Further, exposure of primary human hepatocytes (PHH) to exosomes isolated from treatment-naïve or treatment non-responder HCV infected patients resulted in effective virus infection and replication as indicated by detectable HCV RNA in the culture supernatants (Fig. 4C). Active virus replication after infection of PHH with HCV exosomes was indicated by a 2–3 log increase in HCV copy numbers in PHH at 48 hours after infection compared to the initial HCV copy numbers introduced by the HCV exosomes used for induction of infection (Fig. 4D). Additionally, the use of Telaprevir (VX-950), an NS3.4A serine protease inhibitor, could inhibit HCV replication caused by free virus and HCV exosomes in infected PHH (Fig. 4E).
CD81, SB-RI, APOE and HCV E1/E2 proteins are important host and viral molecules for HCV infection [33]. We and others have shown that anti-CD81 and anti-HCV E1/E2 antibodies can block HCV infection [7], [12], [34]; however, in some instances antibody therapy in patients could not fully prevent HCV infection [16]. Based on these observations, we tested whether the presence of anti-CD81 antibodies would block exosome and cell free virus transmission of HCV. We found that anti-CD81 pre-treatment effectively blocked free HCV virus infection of target Huh7.5 cells, indicated by significantly low HCV RNA expression (Fig. 5A) and by lack of expression of HCV NS3 protein (Fig. 5B). However, exosomes containing HCV RNA could still transmit HCV infection despite anti-CD81 antibody pre-treatment (1∶50 dilution) (Fig. 5A & 5B). These findings were validated in primary human hepatocytes where anti-CD81 pre-treatment significantly inhibited free HCV virus infection but failed to prevent patient exosome-mediated HCV transmission (Fig. 5C & 5D). Additionally, SB-RI (Fig. 5E) or APOE (Fig. 5F) antibody pre-treatment could block HCV J6/JFH-1 free virus transmission but not HCV-exosome transmission of HCV to naïve Huh7.5 cells (Fig. 5E & 5F).
We next tested CD81-deficient Huh7.25-CD81 cells [35] and found that HCV exosomes could still mediate HCV transmission but infection rate with the free virus entry was significantly diminished (Fig. 6A). In the parental Huh7.0 cells, both exosomes and free HCV virus resulted in comparable extent of HCV infection (Fig. 6B). HCV E1 and E2 envelope glycoproteins which can modulate HCV infection [36] have been shown to associate with exosomes [23], thus we tested if anti-HCV E2 antibody treatment could block HCV transmission by exosomes. We found that anti-HCV E2 antibody treatment of HCV J6/JFH-1 virus could significantly block HCV transmission by free HCV particles but not by exosomes (Fig. 6C).
Recent reports have demonstrated the role of Ago2 and miR-122 in enhancing HCV replication when bound to the 5′-UTR of HCV dsRNA [26]. We observed that the same MOI of free HCV viruses or HCV-exosomes resulted in a trend (but not statistically significant) of greater HCV transmission by exosomes compared to the cell free virus (Fig. 5). Based on this observation, we surmised that exosomes might contain replication-competent RNA in association with RISC complex proteins that could enhance HCV RNA stability and enhance viral replication [26], [37], [38]. Using RNA-chromatin immunoprecipitation (RNA-ChIP) analysis of exosomes isolated from HCV J6/JFH-1 infected Huh 7.5 cells or HCV infected patients after Ago2 pull-down, we found that Ago2 was associated with miR-122 (Fig. 7A), positive sense HCV RNA (Fig. 7B upper panel) and, in some cases, negative sense HCV RNA (Fig. 7B lower panel). Using free HCV virus RNA and RNA from HCV infected cells we confirmed primer specificity for detection of positive and negative sense HCV RNA (Figures S5A & S5B). Additionally, using co-immuno precipitation, we confirmed that HSP90 and Ago2 formed complexes within the HCV containing exosomes likely providing further stabilization of the HCV RNA-replication complex (Fig. 7C) [39].
These striking observations indicate that serum exosomes from some HCV infected treatment-naïve patients contain positive sense RNA of HCV virus and are able to transmit active HCV infection. We found that, even in the few patients where we could not detect viral RNA in the exosomes due to the limitation of the sensitivity of the Real Time PCR method (Table 1 & 2), HCV infection of PHH was still evident (Fig. 4C). Furthermore, replication competent, negative sense HCV RNA was also present in some treatment-naïve and in all non-responder patients (Table 1 & 2).
Given that exosomes from HCV-infected treatment-naïve and treatment non-responders contained Ago2 in complex with miR-122, and HSP90, we tested the effect of miR-122 or HSP90 inhibitors which have been suggested for HCV treatment [29], [37], [40]. Delivery of a miR-122 inhibitor resulted in about 50% reduction in miR-122 levels in Huh7.5 cells that is significant considering the high abundance of miR-122 in hepatocytes (Fig. 8A). However, inhibition of miR-122 in Huh7.5 cells prior to infection with HCV exosome failed to significantly suppress HCV transmission (Fig. 8A). Given that exosomes harbored HCV in complex with miR-122/HSP90, we hypothesized that miR-122 in the exosomes provides advantages for HCV transmission. To test this hypothesis, we transfected HCV-exosomes with a miR-122 inhibitor or control, washed and re-purified the miR-122 inhibitor- or control inhibitor-loaded HCV-exosomes and used them for infection of naïve Huh7.5 cells. The miR-122 inhibitor-loaded HCV-exosomes resulted in a significant reduction in intracellular miR-122 levels in Huh7.5 cells (Fig. 8B). Importantly, we found reduced virus transmission by HCV-exosomes loaded with the miR-122 inhibitor as indicated by decreased HCV NS3 protein compared to the controls (Fig. 8C).
We also assessed the potential of the HSP90 activity inhibitor, 17-DMAG, or HSP90 siRNA treatment to modulate HCV infection transmitted by exosomes (Fig. 8D). We found that DMAG treatment but not HSP90 siRNA treatment could significantly block exosome-mediated HCV transmission (Fig. 8D).
Previous data showed that viral HCV entry and subsequent infection can be prevented by administering vacuolar-type H+-ATPase inhibitor [41]. Moreover, Meertens et al [42] reported that entry of HCV pseudoparticles (HCVpp) was efficiently blocked by bafilomycin A1, a specific vacuolar-type H+-ATPase inhibitor, which neutralizes the pH in early endosomes and injures progression of endocytosis beyond this level. Exosome entry through endocytosis is reported to be pH dependent in the traffic of tumor exosomes in regulating both their release and uptake by tumor cells [43]. Based on these reports, we set up an in vitro model utilizing a vacuolar-type H+-ATPase inhibitor (bafilomycin A1) or a proton pump inhibitor (Lansoprazole) to study the role of low pH in favoring HCV infected exosome uptake in Huh 7.5 cells. We found that low pH plays a role in the entry of infected exosome into the Huh 7.5 cells and infection by HCV infected exosomes can be blocked using vacuolar-type H+-ATPase or a proton pump inhibitor. Our data show that both Lansoprazole (Fig. 9A & 9B) and bafilomycin A1 (Fig. 9C & 9D) could significantly inhibit HCV transmission by exosomes and cell free HCV viruses to Huh7.5 hepatoma cells in a dose-dependent manner without causing significant cellular cytotoxicity (Figures S6A & S6B).
Exosomes are found in different biofluids and represent a small (40–150 nm) subpopulation of extracellular vesicles of endocytic origin released by almost all cell types. They act as natural carriers of genetic materials, namely miRNA, mRNA and proteins [44], [45]. Notably, exosomes have been shown to mediate disease transmission caused by bacteria, infectious prion protein, and viruses [46], [47]. In the context of HCV, recent studies showed that hepatocyte-derived exosomes containing viral RNA induced production of IFN-α in plasmacytoid dendritic cells (pDCs) in vitro [24]. In this study, we demonstrate that circulating exosomes derived from sera of treatment-naïve HCV infected individuals or HCV treatment non-responder individuals contain HCV virus that can transmit active HCV infection to primary human hepatocytes, confirmed with the observation of a 2–3 log increase in HCV copy numbers in PHH compared to the initial HCV copy numbers in exosomes used for infection indicated virus replication. A recent study also reported exosome-mediated transmission of HCV in Huh7.5 cells [48]. Our observations confirmed and extended a recent report that also found that exosomes derived from HCV J6/JFH-1 infected Huh7.5 cells can shuttle virus to normal Huh7.5 cells and establish a productive infection. Using a stringent isolation methodology of serial filtration followed by density separation and immune magnetic CD63-positive exosome isolation, we optimized a method of HCV exosome isolation without carryover of free virus thereby further underscoring the capacity of exosomes to transmit HCV infection.
Our findings showed for the first time that exosomes from sera of HCV infected patients or culture supernatants of HCV J6/JFH-1 infected Huh7.5 cells can mediate effective CD81, SB-RI, HCV E2 and APOE -independent HCV transmission to hepatocytes. A recent report by Ramakrishnaiah et al [48] indicated that exosomes can mediate partial CD81-independent HCV transmission in Huh7.5 cells, however in that study cell free HCV transmission could not be fully excluded. Our results indicate that exosomes that are devoid of free virus contamination are capable of HCV transmission even in the presence of a potent anti-CD81, anti-SB-RI, anti-HCV E2 and anti-APOE antibody treatment and in CD81-deficient cells. These observations could explain in part why neutralizing antibodies or therapies that target host/viral protein interactions at the level of cell entry can be compromised and likely occur via cell-to-cell transmission by exosomes.
Given that infections with HCV-exosomes compared to the same MOI of free HCV virus particles, showed a tendency for higher levels of HCV transmission to hepatocytes, it was unclear if these exosomes contained replication competent HCV RNA, factors that enhanced virus replication or facilitated mechanisms of exosomes entry to target cells. Recently, reports have consistently demonstrated that RISC-like complexes involving Ago2 and miR-122 can protect the HCV 5′ internal ribosome entry site (5′ IRES) and enhance HCV replication [26], [38]. We found higher miR-122 expression in HCV J6/JFH-1 infected Huh7.5 cells derived exosomes compared to HCV infected patient exosomes and their respective controls, possibly as a result of suppressed interferon production in Huh7.5 cells since Huh7.5 cells harbor a mutation in the dsRNA sensor retinoic acid-inducible gene-I (RIG-I) [49]–[51]. However, using RNA ChIP analyses we found that exosomes from HCV J6/JFH-1 infected Huh 7.5 cells and exosomes from the two patient groups that have active infection, treatment-naïve and treatment non-responders, showed increased proportion of miRNA-122 in complex with Ago2. Additionally, it was remarkable that Ago2 and miR-122 bound to the HCV 5′-UTR was also in association with HSP90 which has been shown to stabilize RISC complexes [52] and potentially increase HCV replication. Our observations support a hypothesis whereby exosomes mediate higher HCV transmission because they contain replication-competent viral RNA, as well as, known HCV replication enhancers- Ago2 [26], miR-122 [26], [29], [37], and HSP90 [37], [53], [54]. Additionally, HCV RNA in exosomes might mediate higher levels of infection possible due to the higher stability of HCV RNA when associated with Ago2 and miR-122 as suggested by Shimakami et al [39].
The presence of host proteins within HCV-exosomes is a clever strategy by the virus to ensure effective replication once in the endoplasmic reticulum (ER) given that the ER does not contain these exosomal proteins [55]. Our novel findings may translate and offer possible clinical implications to HCV treatment resistance with interferon/ribavirin given that exosomes from HCV-infected treatment resistant patients contained HCV negative sense RNA, which is mostly associated with replication-competent HCV RNA. Strikingly, only some of the treatment-naïve patients with HCV positive-sense RNA detected in their exosomes contained negative-sense HCV RNA (Table 2). Importantly, none of HCV treatment responder patients harbored detectable HCV RNA in their serum-derived exosomes consistent with their status of HCV viral clearance. The implication of our findings needs additional clinical follow-up to determine whether treatment and/or disease outcome using anti-HCV immune therapies would be influenced by the composition of serum-derived exosomes in HCV infected patients.
Based on our novel findings that exosomes can mediate virus transmission via CD81, SB-RI and APOE -independent mechanisms potentially compromising the efficacy of HCV immunotherapies, we next aimed to test therapeutic alternatives. We analyzed the potential use of miR-122 inhibition and DMAG treatment both of which have been successfully explored for HCV treatment but not yet assessed in the context of exosome-mediated HCV transmission. We found that using an exosome targeted miR-122 inhibitor system or the HSP90 inhibitor, DMAG, which could inhibit the effective function of these host factors which modulate HCV infection/replication, could significantly suppress HCV transmission by exosomes. Strikingly, attenuation of HSP90 or miR-122 levels by siRNA knockdown and miR-122 inhibitor in target cells was not sufficient to inhibit HCV transmission via exosomes. This could be due to the fact that HCV exosomes contain all the necessary viral and host protein factors that are otherwise not present in the endoplasmic reticulum [18], and can thus mediate effective replication once cellular entry is accomplished by exosome uptake.
Since exosomes originate from lumen of multivesicular bodies (MVBs), their release and uptake are associated with the endocytic pathway [56]. Acidification of intracellular organelles is reported to be fundamental to the function of the endocytic pathways and exosomes uptake [57]. The vacuolar H+-ATPases (V-ATPases) and proton pumps are responsible for generating and maintaining intra-cellular pH gradients across cell membranes. Disruption to their functions were reported to be accompanied by lysosomal dysfunction and impaired endocytosis [58], [59]. From another perspective, several reports show a crucial role of low pH and endosome acidification for triggering virus entry, not addressing the distinction between exosomes and cell free viruses [41], [60]. Here we show that the use of bafilomycin A1, a specific vacuolar H+-ATPase proton pump inhibitor, and Lansoprazole, a proton pump inhibitor, prevented the capacity of exosomes and cell free virus to transmit infection, suggesting their use in the treatment regimens for HCV infection. This usage seems to be more influential as it is reported that the intracellular pH was not noticeably changed by dosages less than 100 nM of bafilomycin A1 and for a short period of time (4 h), which is reported to be the critical time point for effect of bafilomycin A1 to prevent viral entry [61].
In summary, our novel findings, illustrated in Figure 10 provide mechanistic insights into how exosomes can mediate indirect cell-to-cell viral receptor independent transmission of HCV. Furthermore, we provide evidence that circulating exosomes of HCV infected patients can infect primary human hepatocytes. Additionally, our findings further support the rationale for using miR-122 inhibitors, HSP90 inhibitor, and potentially proton-pump and Vacuolar-type H+-ATPase inhibitors to prevent exosome-mediated HCV transmission.
Huh7.5, Huh7.0 (a gift from Dr. Charlie Rice, Rockefeller University, New York) and CD81-deficient Huh7.25 [a gift from Dr. Takaji Wakita (National Institute of Infectious Disease, Tokyo, Japan) and Dr. T. Jake Liang (NIDDK, National Institutes of Health, USA)] cells were cultured as previously described [35], [49] with slight modification, using exosome depleted FBS (System Bioscience cat. #EXO-FBS-50A-1). Primary human hepatocytes were obtained from the National Institutes of Health (NIH) liver tissue cell distribution system (LTCDS; Minneapolis, MN, USA; Pittsburgh, PA; Richmond, VA, USA), which was funded by NIH contract #N01-DK-7-004/HHSN2670070004C and from BD Bioscience. Highly infectious and replication competent HCV J6/JFH-1 virus (genotype 2a) were generated as previously described [62]. The pFL-J6/JFH-1 plasmid used for virus generation was provided by Dr. Charlie Rice and Dr. Takaji Wakita (National Institute of Infectious Disease, Tokyo, Japan). HCV J6/JFH-1 virus concentration in culture supernatants was determined using NanoSight LM10 (MOI of infectious viral particles or infectious exosomes) and by quantitative real-time PCR as previously described [37].
Subjects were recruited from the Hepatology clinic at the University of Massachusetts Medical School. This research protocol was reviewed by the Committee for the Protection of Human Subjects in Research at the University of Massachusetts Medical School (IRB #2284). All subjects who donated samples for this project provided signed written informed consent. Subjects were assessed for baseline demographics, Hepatitis C viral serology and liver function parameters (Table 1). Healthy control subjects had no evidence of systemic disease, HCV infection, or other liver diseases. Informed consent was obtained from all subjects. Blood samples were drawn and serum samples were analyzed for HCV RNA using RT-PCR and processed as subsequently indicated.
Huh7.5 cells and HCV J6/JFH-1 infected Huh7.5 cells were maintained in DMEM low glucose medium supplemented with 10% exosome depleted FBS and 1% penicillin/streptomycin (Gibco, cat. #15140-163). Cell culture supernatants following cell infection or not, or patient serum samples were collected, centrifuged at 2500 rpm for 10 mins at 4°C to remove cell debris, then filtered through a 0.2 µm filter. The 40 mL of filtered culture supernatant for exosome isolation was concentrated to a final 1 mL volume using the Amicon Ultra-15 Centrifugal Filter Unit with Ultracel-100 membrane (Millipore, cat. #UFC910024). Concentrated culture supernatants or filtered patient serum (500 uL) were mixed with the appropriate volume of Exoquick-TC reagent (System Biosciences cat. #EXOTC10A-1) or Exoquick (System Bioscience cat. #EXOQ5A-1) respectively, for exosome isolation according to the manufacturers' specification. Samples were gently mixed and incubated for 1 h at 4°C. Following incubation, exosomes were precipitated by centrifugation at 1400 rpm for 10 mins at 4°C. The recovered exosomes were re-suspended in 1× phosphate buffered saline (PBS). Positive selection of exosomes was done using anti-CD63 immuno-magnetic capturing with primary anti-CD63 antibody (Abcam cat. #ab8219 and Santa Cruz cat. #15363) followed by corresponding secondary antibody coupled to magnetic beads (Miltenyi Biotec cat. #130-048-602). The Miltenyi Biotec MidiMACS separator was used with LD columns (cat. # 130-042-901) for exosome isolation.
Exosomes isolated by positive anti-CD63 immuno-magnetic bead selection were re-suspended in PBS and transferred to a formvar-coated copper grid then allowed to settle/attach for 30 minutes. The grid was washed by sequentially positioning droplets of PBS on top and using absorbing paper in between. The samples were then fixed by drop-wise addition of 2% paraformaldehyde onto parafilm and placing the grid on top of the paraformaldehyde drop for 10 min. Fixation was followed by five washes with deionized water and samples contrasted by adding 2% uranyl acetate for 15 minutes. Afterward, the samples were embedded by adding a drop of 0.13% methyl cellulose and 0.4% uranyl acetate for 10 minutes. The grid was visualized using a Philips CM10 transmission electron microscope and images were captured using a Gatan CCD digital camera.
Quantification of immuno-magnetic CD63 bead captured infectious HCV J6/JFH-1-exosomes and HCV J6/JFH-1 virus preparations was determined using NanoSight LM10 system (NanoSight, Amesbury, UK) equipped with a fast video capture and Nanoparticle Tracking Analysis (NTA) system, according to the manufacturer's instructions. Quantification of HCV RNA copy numbers was done as previously described [31].
The following siRNA and miRNA inhibitors were used: human HSP90 siRNA (Santa Cruz cat. #sc-35608); control siRNA (Santa Cruz cat. #sc-44236), hsa-miR-122 anti-miR miRNA Inhibitor (Ambion, Austin, Tx cat. #AM11012) and anti-miR Negative Control (Ambion, Austin, Tx cat. #AM17010). miRNA or control inhibitors were complexed with the liver specific in vivo Altogen delivery reagent (Altogen Biosystems cat. #5060) which was loaded into control exosomes or HCV exosomes then co-cultured with target cells as indicated. Specific SiRNA or control siRNA was complexed with FugeneHD (Roche cat. # 04709705001) and transfected into target cells according to the manufacturer's specifications as indicated.
Exosomes isolated from cell culture supernatants or patient serum samples were fixed at room temperature with 4% formaldehyde buffered saline. Afterward, exosomes were lysed in SDS ChIP lysis buffer (Millipore cat. # 20-163) supplemented with protease inhibitor and RNase inhibitor. Total exosome proteins were pre-cleared with protein G beads. 50 µg of total protein was incubated with Ago2 antibody. Immunoprecipitation was performed for 90 minutes at 4°C using 10 µg/ml primary Ago2 antibody and normal rabbit IgG (Santa Cruz cat # sc-2027) non-specific antibody used as IP control. A mixture of Protein A/G PLUS-Agarose beads (Santa Cruz cat. #sc-2003) was added, and the incubation was continued for an additional 60 minutes. The samples were washed with SDS ChIP lysis buffer supplemented with protease inhibitor and RNase inhibitor. The immunoprecipitated protein-RNA complex was either used for Western blot analysis or RNA purification after Ago2 pull down using the Zymo research Direct-zol RNA MiniPrep kit (cat. #R2050), according to the manufacturer's specification. TaqMan MicroRNA assay was used for quantification of miRNA, using a CFX Connect Real-Time PCR Detection System (Philadelphia, USA). The exosome miRNA data was normalized to Cel39 and fold change was calculated using delta-delta ct method as previously described [63].
Western blots were performed using the following established protocols. Briefly, proteins were resolved on 10% SDS-PAGE gels. After electrophoresis resolved proteins were transferred onto nitrocellulose membranes. Following protein transfer, membranes were blocked for 1 hour in PBS containing 5% non-fat dry milk and 0.1% Tween-20. Blots were then incubated overnight with primary antibody at 4°C. The following primary antibodies were used: anti-HCV NS3 (Abcam cat. #ab13830); anti-HSP90 (Cell Signaling cat. #4874); anti-CD63 (Abcam cat. #ab8219 used for western blotting and Santa Cruz Biotechnology cat. #sc-15363 used for exosomes purification); anti-Ago2 (Sigma cat. #SAB4200274); anti-CD81 (Santa Cruz Biotechnology cat. #sc-23962), normal rabbit IgG-AC antibody (Santa Cruz Biotechnology cat. # sc-2345); anti-beta actin [Ac-15] (Abcam, cat. #ab6276). The membranes were then incubated for 1 hour with horseradish peroxidase-conjugated secondary antibodies (dilution 1∶10,000) that included: goat anti-mouse IgG-HRP (Santa Cruz Biotechnology cat. #sc-2005); goat anti-rabbit IgG-HRP (Santa Cruz Biotechnology cat. #sc-2004). Finally, the proteins were visualized with the Clarity Western ECL substrate (BioRad, cat. #170-5061) chemiluminescence system according to the manufacturer's protocol using the Fujifilm LAS-4000 luminescent image analyzer.
Prior to total RNA isolation, equal volume of plasma (500 µL) or 500 uL of 10 mL concentrated culture supernatant samples were thawed on ice, mixed with QIAzole (Qiagen), vortexed and incubated at RT for 5 mins. Synthetic C. elegans (cel)-miR-39 was spiked and after this step total RNA was extracted using Zymo research Direct-zol RNA MiniPrepKit as per instructions. TaqMan miRNA Assay (Applied Biosystems) was used to analyze the miRNA from serum or plasma samples. Cel-miR-39 was used to normalize the technical variation between the exosomes samples and when comparing miRNA or HCV RNA content in cell lines compared to exosomes. Quantification of miR-122 was performed using Taqman microRNA assays (Applied Biosystems). RNU48 was used as an endogenous control for miR-122 expression in cells and Cel-miR-39 was used as an exogenous control to normalize for technical variation in RNA isolation for determining miR-122 levels in exosomes.
After RNA isolation as indicated, reverse transcription was performed by two different methods both of which were designed to amplify the 5′-UTR of HCV as previously described [37]. Briefly, positive sense RNA was amplified involving a first cDNA synthesis reaction using 500 ng of total RNA using the Bio-Rad cDNA synthesis kit according to the manufacturer's specification. The positive sense HCV 5′ UTR was then amplified using the following primer sequence: HCV Forward Primer: 5′-TCTGCGGAACCGGTGAGTAC-3′; HCV Reverse primer: 5′-TCAGGCAGTACCACAAGGCC-3′. HCV negative sense RNA was detected using primers and PCR conditions as previously described [64].
Bafilomycin A1 was purchased from Sigma Aldrich and the proton pump inhibitor; Lansoprazole (Prevacid 24 hr OTC, Novartis), was purchased over the counter. Lansoprazole was dissolved in DMSO and applied to Huh7.5 cells at concentrations of 2.5 µg/ml, 5 µg/ml, and 10 µg/ml. Telaprevir (VX-950) was purchased from Selleckchem and used as previously described [65]. One hour later, HCV virus suspension and HCV infected exosomes (captured with CD63) were added to the cells. Twenty-four hour later, the cells were washed 3 times and assessed for viral structural protein, NS3. Bafilomycin A1 was dissolved in DMSO and applied to the Huh 7.5 cells at concentrations of 12.5 nM, 25 nM, 50 nM, and 100 nM, while the concentration of DMSO in the final treatments was 0.01%. One hour later, HCV virus suspension and HCV infected exosomes (captured with anti-CD63 antibody) were added to the cells. After 24 h, the cells were washed 3 times and assessed for viral RNA entry.
The LDH toxicity assay kit (Abcam Cat. # ab65393) was used according to the manufacturer's specification. Briefly, released LDH in culture supernatants of Huh7.5 cells after 24 h co-culture with different concentration of Bafilomycin A1 and Lansoprazole was measured as the indicator of lysed cells. The percentage of cytotoxicity was measured by subtracting LDH content in remaining viable cells from total LDH in untreated controls. Staurosporine (20 nM) (Abcam, Cambridge, MA) treatment of Huh7.5 cells for 12 h was used as positive control. The final absorbance was measured at 490 nm. All experiments were performed in triplicate.
Cells, as indicated, were treated with blocking antibodies to target HCV host receptors for one hour prior to infection with either HCV exosomes, HCV J6/JFH-1 virus or not as indicated. Blocking antibodies used included: anti-CD81 antibody (Santa Cruz Biotechnology cat. # sc-23962), anti-Scavenging Receptor (SR-BI) antibody (Abcam, cat. # ab52629), anti-HCV E2 antibody (GeneTex cat. # GTX103353) and anti-ApoE antibody (Millipore Cat. #: AB947).
Culture supernatants of Huh7.5 cells infected or non-infected with HCV (J6/JFH-1) were centrifuged at 1,000× rpm for 10 minutes to remove cells followed by another spin at 2,000× rpm for 15 minutes to remove cellular debris. Exosomes were positively selected with CD63 immunomagnetic beads as described above and the flow through collected which included cell free virus and viral particles. Levels of APOE and APOB proteins in the exosomes were identified by using Apolipoprotein E (APOE) Human ELISA Kit (Abcam cat # ab108813) and Human Apolipoprotein B (APOB) Quantikine ELISA Kit (R&D Systems cat # DAPB00) according to the manufacturers' protocols. The same number of control exosomes (obtained from non-infected Huh 7.5 cells), exosomes derived from HCV infected Huh 7.5 cells and viral particles were used for the experiment and quantified by Nanosight measurements. The optical density of the color reactions for both plates was read on plate reader at 450 nm. Standard curves were generated and concentrations of APOE and APOB were calculated as stipulated in the manufacturer's protocol. Liver cell protein lysate was used as positive control.
Data are representative of at least 3 independently repeated experiments presented as mean + standard error of the mean (SEM). A non-parametric Mann-Whitney U test and multiple comparisons for repeated-measures were done using ANOVA performed with GraphPad Prism Version 5.0 (GraphPad Software). A p value of <0.05 was considered significant.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.